Skip to content

Examples

OmniGibson ships with many demo scripts highlighting its modularity and diverse feature set intended as a set of building blocks enabling your research. Let's try them out!

A quick word about macros

Why macros?

Macros enforce global behavior that is consistent within an individual python process but can differ between processes. This is useful because globally enabling all of OmniGibson's features can cause unnecessary slowdowns, and so configuring the macros for your specific use case can optimize performance.

For example, Omniverse provides a so-called flatcache feature which provides significant performance boosts, but cannot be used when fluids or soft bodies are present. So, we ideally should always have gm.USE_FLATCACHE=True unless we have fluids or soft bodies in our environment.

macros define a globally available set of magic numbers or flags set throughout OmniGibson. These can either be directly set in omnigibson.macros.py, or can be programmatically modified at runtime via:

from omnigibson.macros import gm, macros

gm.<GLOBAL_MACRO> = <VALUE> # (1)!
macros.<OG_DIRECTORY>.<OG_MODULE>.<MODULE_MACRO> = <VALUE> # (2)!
  1. gm refers to the "global" macros -- i.e.: settings that generally impact the entire OmniGibson stack. These are usually the only settings you may need to modify.
  2. macros captures all remaining macros defined throughout OmniGibson's codebase -- these are often hardcoded default settings or magic numbers defined in a specific module. These can also be overridden, but we recommend inspecting the module first to understand how it is used.

Many of our examples set various macros settings at the beginning of the script, and is a good way to understand use cases for modifying them!

Environments

These examples showcase the full OmniGibson stack in use, and the types of environments immediately supported.

BEHAVIOR Task Demo

This demo is useful for...

  • Understanding how to instantiate a BEHAVIOR task
  • Understanding how a pre-defined configuration file is used
python -m omnigibson.examples.environments.behavior_env_demo

This demo instantiates one of our BEHAVIOR tasks (and optionally sampling object locations online) in a fully-populated scene and loads a Fetch robot. The robot executes random actions and the environment is reset periodically.

behavior_env_demo.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
import os

import yaml

import omnigibson as og
from omnigibson.macros import gm
from omnigibson.utils.ui_utils import choose_from_options

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Generates a BEHAVIOR Task environment in an online fashion.

    It steps the environment 100 times with random actions sampled from the action space,
    using the Gym interface, resetting it 10 times.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Ask the user whether they want online object sampling or not
    sampling_options = {
        False: "Use a pre-sampled cached BEHAVIOR activity scene",
        True: "Sample the BEHAVIOR activity in an online fashion",
    }
    should_sample = choose_from_options(
        options=sampling_options, name="online object sampling", random_selection=random_selection
    )

    # Load the pre-selected configuration and set the online_sampling flag
    config_filename = os.path.join(og.example_config_path, "fetch_behavior.yaml")
    cfg = yaml.load(open(config_filename, "r"), Loader=yaml.FullLoader)
    cfg["task"]["online_object_sampling"] = should_sample

    # Load the environment
    env = og.Environment(configs=cfg)

    # Allow user to move camera more easily
    og.sim.enable_viewer_camera_teleoperation()

    # Run a simple loop and reset periodically
    max_iterations = 10 if not short_exec else 1
    for j in range(max_iterations):
        og.log.info("Resetting environment")
        env.reset()
        for i in range(100):
            action = env.action_space.sample()
            state, reward, terminated, truncated, info = env.step(action)
            if terminated or truncated:
                og.log.info("Episode finished after {} timesteps".format(i + 1))
                break

    # Always close the environment at the end
    og.clear()


if __name__ == "__main__":
    main()

This demo is useful for...

  • Understanding how to instantiate a navigation task
  • Understanding how a pre-defined configuration file is used
python -m omnigibson.examples.environments.navigation_env_demo

This demo instantiates one of our navigation tasks in a fully-populated scene and loads a Turtlebot robot. The robot executes random actions and the environment is reset periodically.

navigation_env_demo.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
import os

import yaml

import omnigibson as og
from omnigibson.utils.ui_utils import choose_from_options


def main(random_selection=False, headless=False, short_exec=False):
    """
    Prompts the user to select a type of scene and loads a turtlebot into it, generating a Point-Goal navigation
    task within the environment.

    It steps the environment 100 times with random actions sampled from the action space,
    using the Gym interface, resetting it 10 times.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Load the config
    config_filename = os.path.join(og.example_config_path, "turtlebot_nav.yaml")
    config = yaml.load(open(config_filename, "r"), Loader=yaml.FullLoader)

    # check if we want to quick load or full load the scene
    load_options = {
        "Quick": "Only load the building assets (i.e.: the floors, walls, ceilings)",
        "Full": "Load all interactive objects in the scene",
    }
    load_mode = choose_from_options(options=load_options, name="load mode", random_selection=random_selection)
    if load_mode == "Quick":
        config["scene"]["load_object_categories"] = ["floors", "walls", "ceilings"]

    # Load the environment
    env = og.Environment(configs=config)

    # Allow user to move camera more easily
    og.sim.enable_viewer_camera_teleoperation()

    # Run a simple loop and reset periodically
    max_iterations = 10 if not short_exec else 1
    for j in range(max_iterations):
        og.log.info("Resetting environment")
        env.reset()
        for i in range(100):
            action = env.action_space.sample()
            state, reward, terminated, truncated, info = env.step(action)
            if terminated or truncated:
                og.log.info("Episode finished after {} timesteps".format(i + 1))
                break

    # Always close the environment at the end
    og.clear()


if __name__ == "__main__":
    main()

Learning

These examples showcase how OmniGibson can be used to train embodied AI agents.

Reinforcement Learning Demo

This demo is useful for...

  • Understanding how to hook up OmniGibson to an external algorithm
  • Understanding how to train and evaluate a policy
python -m omnigibson.examples.learning.navigation_policy_demo

This demo loads a BEHAVIOR task with a Fetch robot, and trains / evaluates the agent using Stable Baseline3's PPO algorithm.

navigation_policy_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
"""
Example training code using stable-baselines3 PPO for one BEHAVIOR activity.
Note that due to the sparsity of the reward, this training code will not converge and achieve task success.
This only serves as a starting point that users can further build upon.
"""

import argparse
import os
import time

import yaml

import omnigibson as og
from omnigibson import example_config_path
from omnigibson.macros import gm
from omnigibson.utils.python_utils import meets_minimum_version

try:
    import gymnasium as gym
    import tensorboard  # noqa: F401
    import torch as th
    import torch.nn as nn
    from stable_baselines3 import PPO
    from stable_baselines3.common.callbacks import CallbackList, CheckpointCallback, EvalCallback
    from stable_baselines3.common.evaluation import evaluate_policy
    from stable_baselines3.common.torch_layers import BaseFeaturesExtractor
    from stable_baselines3.common.utils import set_random_seed

except ModuleNotFoundError:
    og.log.error(
        "torch, stable-baselines3, or tensorboard is not installed. "
        "See which packages are missing, and then run the following for any missing packages:\n"
        "pip install stable-baselines3[extra]\n"
        "pip install tensorboard\n"
        "pip install shimmy>=0.2.1\n"
        "Also, please update gym to >=0.26.1 after installing sb3: pip install gym>=0.26.1"
    )
    exit(1)

assert meets_minimum_version(gym.__version__, "0.28.1"), "Please install/update gymnasium to version >= 0.28.1"

# We don't need object states nor transitions rules, so we disable them now, and also enable flatcache for maximum speed
gm.ENABLE_OBJECT_STATES = False
gm.ENABLE_TRANSITION_RULES = False
gm.ENABLE_FLATCACHE = True


class CustomCombinedExtractor(BaseFeaturesExtractor):
    def __init__(self, observation_space: gym.spaces.Dict):
        # We do not know features-dim here before going over all the items,
        # so put something dummy for now. PyTorch requires calling
        super().__init__(observation_space, features_dim=1)
        extractors = {}
        self.step_index = 0
        self.img_save_dir = "img_save_dir"
        os.makedirs(self.img_save_dir, exist_ok=True)
        total_concat_size = 0
        feature_size = 128
        for key, subspace in observation_space.spaces.items():
            # For now, only keep RGB observations
            if "rgb" in key:
                og.log.info(f"obs {key} shape: {subspace.shape}")
                n_input_channels = subspace.shape[0]  # channel first
                cnn = nn.Sequential(
                    nn.Conv2d(n_input_channels, 4, kernel_size=8, stride=4, padding=0),
                    nn.ReLU(),
                    nn.MaxPool2d(2),
                    nn.Conv2d(4, 8, kernel_size=4, stride=2, padding=0),
                    nn.ReLU(),
                    nn.MaxPool2d(2),
                    nn.Conv2d(8, 4, kernel_size=3, stride=1, padding=0),
                    nn.ReLU(),
                    nn.Flatten(),
                )
                test_tensor = th.zeros(subspace.shape)
                with th.no_grad():
                    n_flatten = cnn(test_tensor[None]).shape[1]
                fc = nn.Sequential(nn.Linear(n_flatten, feature_size), nn.ReLU())
                extractors[key] = nn.Sequential(cnn, fc)
                total_concat_size += feature_size
        self.extractors = nn.ModuleDict(extractors)

        # Update the features dim manually
        self._features_dim = total_concat_size

    def forward(self, observations) -> th.Tensor:
        encoded_tensor_list = []
        self.step_index += 1

        # self.extractors contain nn.Modules that do all the processing.
        for key, extractor in self.extractors.items():
            encoded_tensor_list.append(extractor(observations[key]))

        feature = th.cat(encoded_tensor_list, dim=1)
        return feature


def main():
    # Parse args
    parser = argparse.ArgumentParser(description="Train or evaluate a PPO agent in BEHAVIOR")

    parser.add_argument(
        "--checkpoint",
        type=str,
        default=None,
        help="Absolute path to desired PPO checkpoint to load for evaluation",
    )

    parser.add_argument(
        "--eval",
        action="store_true",
        help="If set, will evaluate the PPO agent found from --checkpoint",
    )

    args = parser.parse_args()
    tensorboard_log_dir = os.path.join("log_dir", time.strftime("%Y%m%d-%H%M%S"))
    os.makedirs(tensorboard_log_dir, exist_ok=True)
    prefix = ""
    seed = 0

    # Load config
    with open(f"{example_config_path}/turtlebot_nav.yaml", "r") as f:
        cfg = yaml.load(f, Loader=yaml.FullLoader)

    # Make sure flattened obs and action space is used
    cfg["env"]["flatten_action_space"] = True
    cfg["env"]["flatten_obs_space"] = True

    # Only use RGB obs
    cfg["robots"][0]["obs_modalities"] = ["rgb"]

    # If we're not eval, turn off the start / goal markers so the agent doesn't see them
    if not args.eval:
        cfg["task"]["visualize_goal"] = False

    env = og.Environment(configs=cfg)

    # If we're evaluating, hide the ceilings and enable camera teleoperation so the user can easily
    # visualize the rollouts dynamically
    if args.eval:
        ceiling = env.scene.object_registry("name", "ceilings")
        ceiling.visible = False
        og.sim.enable_viewer_camera_teleoperation()

    # Set the set
    set_random_seed(seed)
    env.reset()

    policy_kwargs = dict(
        features_extractor_class=CustomCombinedExtractor,
    )

    os.makedirs(tensorboard_log_dir, exist_ok=True)

    if args.eval:
        assert args.checkpoint is not None, "If evaluating a PPO policy, @checkpoint argument must be specified!"
        model = PPO.load(args.checkpoint)
        og.log.info("Starting evaluation...")
        mean_reward, std_reward = evaluate_policy(model, env, n_eval_episodes=50)
        og.log.info("Finished evaluation!")
        og.log.info(f"Mean reward: {mean_reward} +/- {std_reward:.2f}")

    else:
        model = PPO(
            "MultiInputPolicy",
            env,
            verbose=1,
            tensorboard_log=tensorboard_log_dir,
            policy_kwargs=policy_kwargs,
            n_steps=20 * 10,
            batch_size=8,
            device="cuda",
        )
        checkpoint_callback = CheckpointCallback(save_freq=1000, save_path=tensorboard_log_dir, name_prefix=prefix)
        eval_callback = EvalCallback(eval_env=env, eval_freq=1000, n_eval_episodes=20)
        callback = CallbackList([checkpoint_callback, eval_callback])

        og.log.debug(model.policy)
        og.log.info(f"model: {model}")

        og.log.info("Starting training...")
        model.learn(
            total_timesteps=10000000,
            callback=callback,
        )
        og.log.info("Finished training!")


if __name__ == "__main__":
    main()

Scenes

These examples showcase how to leverage OmniGibson's large-scale, diverse scenes shipped with the BEHAVIOR dataset.

Scene Selector Demo

This demo is useful for...

  • Understanding how to load a scene into OmniGibson
  • Accessing all BEHAVIOR dataset scenes
python -m omnigibson.examples.scenes.scene_selector

This demo lets you choose a scene from the BEHAVIOR dataset, loads it along with a Turtlebot robot, and cycles the resulting environment periodically.

scene_selector.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
import omnigibson as og
from omnigibson.macros import gm
from omnigibson.utils.asset_utils import get_available_g_scenes, get_available_og_scenes
from omnigibson.utils.ui_utils import choose_from_options

# Configure macros for maximum performance
gm.USE_GPU_DYNAMICS = True
gm.ENABLE_FLATCACHE = True
gm.ENABLE_OBJECT_STATES = False
gm.ENABLE_TRANSITION_RULES = False


def main(random_selection=False, headless=False, short_exec=False):
    """
    Prompts the user to select any available interactive scene and loads a turtlebot into it.
    It steps the environment 100 times with random actions sampled from the action space,
    using the Gym interface, resetting it 10 times.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Choose the scene type to load
    scene_options = {
        "InteractiveTraversableScene": "Procedurally generated scene with fully interactive objects",
        # "StaticTraversableScene": "Monolithic scene mesh with no interactive objects",
    }
    scene_type = choose_from_options(options=scene_options, name="scene type", random_selection=random_selection)

    # Choose the scene model to load
    scenes = get_available_og_scenes() if scene_type == "InteractiveTraversableScene" else get_available_g_scenes()
    scene_model = choose_from_options(options=scenes, name="scene model", random_selection=random_selection)

    cfg = {
        "scene": {
            "type": scene_type,
            "scene_model": scene_model,
        },
    }

    # If the scene type is interactive, also check if we want to quick load or full load the scene
    if scene_type == "InteractiveTraversableScene":
        load_options = {
            "Quick": "Only load the building assets (i.e.: the floors, walls, ceilings)",
            "Full": "Load all interactive objects in the scene",
        }
        load_mode = choose_from_options(options=load_options, name="load mode", random_selection=random_selection)
        if load_mode == "Quick":
            cfg["scene"]["load_object_categories"] = [
                "floors",
                "walls",
                "ceilings",
                "lawn",
                "driveway",
                "roof",
                "rail_fence",
            ]

    # Load the environment
    env = og.Environment(configs=cfg)

    # Allow user to move camera more easily
    if not gm.HEADLESS:
        og.sim.enable_viewer_camera_teleoperation()

    # Run a simple loop and reset periodically
    steps = 0
    while not short_exec or steps < 1000:
        state, reward, terminated, truncated, info = env.step(None)
        if terminated or truncated:
            og.log.info("Episode finished after {} timesteps".format(steps + 1))
            break
        steps += 1

    # Always close the environment at the end
    og.clear()


if __name__ == "__main__":
    main()

Scene Tour Demo

This demo is useful for...

  • Understanding how to load a scene into OmniGibson
  • Understanding how to generate a trajectory from a set of waypoints
python -m omnigibson.examples.scenes.scene_tour_demo

This demo lets you choose a scene from the BEHAVIOR dataset. It allows you to move the camera using the keyboard, select waypoints, and then programmatically generates a video trajectory from the selected waypoints

scene_tour_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
import torch as th

import omnigibson as og
import omnigibson.lazy as lazy
from omnigibson.macros import gm
from omnigibson.utils.asset_utils import get_available_g_scenes, get_available_og_scenes
from omnigibson.utils.ui_utils import KeyboardEventHandler, choose_from_options


def main(random_selection=False, headless=False, short_exec=False):
    """
    Prompts the user to select any available interactive scene and loads it.

    It sets the camera to various poses and records images, and then generates a trajectory from a set of waypoints
    and records the resulting video.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Make sure the example is not being run headless. If so, terminate early
    if gm.HEADLESS:
        print("This demo should only be run not headless! Exiting early.")
        og.shutdown()

    # Choose the scene type to load
    scene_options = {
        "InteractiveTraversableScene": "Procedurally generated scene with fully interactive objects",
        # "StaticTraversableScene": "Monolithic scene mesh with no interactive objects",
    }
    scene_type = choose_from_options(options=scene_options, name="scene type", random_selection=random_selection)

    # Choose the scene model to load
    scenes = get_available_og_scenes() if scene_type == "InteractiveTraversableScene" else get_available_g_scenes()
    scene_model = choose_from_options(options=scenes, name="scene model", random_selection=random_selection)
    print(f"scene model: {scene_model}")

    cfg = {
        "scene": {
            "type": scene_type,
            "scene_model": scene_model,
        },
    }

    # If the scene type is interactive, also check if we want to quick load or full load the scene
    if scene_type == "InteractiveTraversableScene":
        load_options = {
            "Quick": "Only load the building assets (i.e.: the floors, walls, ceilings)",
            "Full": "Load all interactive objects in the scene",
        }
        load_mode = choose_from_options(options=load_options, name="load mode", random_selection=random_selection)
        if load_mode == "Quick":
            cfg["scene"]["load_object_categories"] = ["floors", "walls", "ceilings"]

    # Load the environment
    env = og.Environment(configs=cfg)

    # Allow user to teleoperate the camera
    cam_mover = og.sim.enable_viewer_camera_teleoperation()

    # Create a keyboard event handler for generating waypoints
    waypoints = []

    def add_waypoint():
        nonlocal waypoints
        pos = cam_mover.cam.get_position_orientation()[0]
        print(f"Added waypoint at {pos}")
        waypoints.append(pos)

    def clear_waypoints():
        nonlocal waypoints
        print("Cleared all waypoints!")
        waypoints = []

    KeyboardEventHandler.initialize()
    KeyboardEventHandler.add_keyboard_callback(
        key=lazy.carb.input.KeyboardInput.X,
        callback_fn=add_waypoint,
    )
    KeyboardEventHandler.add_keyboard_callback(
        key=lazy.carb.input.KeyboardInput.C,
        callback_fn=clear_waypoints,
    )
    KeyboardEventHandler.add_keyboard_callback(
        key=lazy.carb.input.KeyboardInput.J,
        callback_fn=lambda: cam_mover.record_trajectory_from_waypoints(
            waypoints=th.tensor(waypoints),
            per_step_distance=0.02,
            fps=30,
            steps_per_frame=1,
            fpath=None,  # This corresponds to the default path inferred from cam_mover.save_dir
        ),
    )
    KeyboardEventHandler.add_keyboard_callback(
        key=lazy.carb.input.KeyboardInput.ESCAPE,
        callback_fn=lambda: og.clear(),
    )

    # Print out additional keyboard commands
    print("\t X: Save the current camera pose as a waypoint")
    print("\t C: Clear all waypoints")
    print("\t J: Record the camera trajectory from the current set of waypoints")
    print("\t ESC: Terminate the demo")

    # Loop indefinitely
    steps = 0
    max_steps = -1 if not short_exec else 100
    while steps != max_steps:
        env.step([])
        steps += 1


if __name__ == "__main__":
    main()

Traversability Map Demo

This demo is useful for...

  • Understanding how to leverage traversability map information from BEHAVIOR dataset scenes
python -m omnigibson.examples.scenes.traversability_map_example

This demo lets you choose a scene from the BEHAVIOR dataset, and generates its corresponding traversability map.

traversability_map_example.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
import os

import cv2
import matplotlib.pyplot as plt
import torch as th

import omnigibson as og
from omnigibson.utils.asset_utils import get_available_og_scenes, get_og_scene_path
from omnigibson.utils.ui_utils import choose_from_options


def main(random_selection=False, headless=False, short_exec=False):
    """
    Traversable map demo
    Loads the floor plan and obstacles for the requested scene, and overlays them in a visual figure such that the
    highlighted area reflects the traversable (free-space) area
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    scenes = get_available_og_scenes()
    scene_model = choose_from_options(options=scenes, name="scene model", random_selection=random_selection)
    print(f"Generating traversability map for scene {scene_model}")

    trav_map_size = 200
    trav_map_erosion = 2

    trav_map = cv2.imread(os.path.join(get_og_scene_path(scene_model), "layout", "floor_trav_0.png"))
    trav_map = cv2.resize(trav_map, (trav_map_size, trav_map_size))
    trav_map = cv2.erode(trav_map, th.ones((trav_map_erosion, trav_map_erosion)).cpu().numpy())

    if not headless:
        plt.figure(figsize=(12, 12))
        plt.imshow(trav_map)
        plt.title(f"Traversable area of {scene_model} scene")
        plt.show()


if __name__ == "__main__":
    main()

Objects

These examples showcase how to leverage objects in OmniGibson.

Load Object Demo

This demo is useful for...

  • Understanding how to load an object into OmniGibson
  • Accessing all BEHAVIOR dataset asset categories and models
python -m omnigibson.examples.objects.load_object_selector

This demo lets you choose a specific object from the BEHAVIOR dataset, and loads the requested object into an environment.

load_object_selector.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
import torch as th

import omnigibson as og
from omnigibson.utils.asset_utils import get_all_object_categories, get_all_object_category_models
from omnigibson.utils.ui_utils import choose_from_options


def main(random_selection=False, headless=False, short_exec=False):
    """
    This demo shows how to load any scaled objects from the OG object model dataset
    The user selects an object model to load
    The objects can be loaded into an empty scene or an interactive scene (OG)
    The example also shows how to use the Environment API or directly the Simulator API, loading objects and robots
    and executing actions
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)
    scene_options = ["Scene", "InteractiveTraversableScene"]
    scene_type = choose_from_options(options=scene_options, name="scene type", random_selection=random_selection)

    # -- Choose the object to load --

    # Select a category to load
    available_obj_categories = get_all_object_categories()
    obj_category = choose_from_options(
        options=available_obj_categories, name="object category", random_selection=random_selection
    )

    # Select a model to load
    available_obj_models = get_all_object_category_models(obj_category)
    obj_model = choose_from_options(
        options=available_obj_models, name="object model", random_selection=random_selection
    )

    # Create and load this object into the simulator
    obj_cfg = dict(
        type="DatasetObject",
        name="obj",
        category=obj_category,
        model=obj_model,
        position=[0, 0, 50.0],
    )

    cfg = {
        "scene": {
            "type": scene_type,
        },
        "objects": [obj_cfg],
    }
    if scene_type == "InteractiveTraversableScene":
        cfg["scene"]["scene_model"] = "Rs_int"

    # Create the environment
    env = og.Environment(configs=cfg)

    # Place the object so it rests on the floor
    obj = env.scene.object_registry("name", "obj")
    center_offset = obj.get_position_orientation()[0] - obj.aabb_center + th.tensor([0, 0, obj.aabb_extent[2] / 2.0])
    obj.set_position_orientation(position=center_offset)

    # Step through the environment
    max_steps = 100 if short_exec else 10000
    for i in range(max_steps):
        env.step(th.empty(0))

    # Always close the environment at the end
    og.clear()


if __name__ == "__main__":
    main()

Object Visualizer Demo

This demo is useful for...

  • Viewing objects' textures as rendered in OmniGibson
  • Viewing articulated objects' range of motion
  • Understanding how to reference object instances from the environment
  • Understanding how to set object poses and joint states
python -m omnigibson.examples.objects.visualize_object

This demo lets you choose a specific object from the BEHAVIOR dataset, and rotates the object in-place. If the object is articulated, it additionally moves its joints through its full range of motion.

visualize_object.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
import argparse
import math

import torch as th

import omnigibson as og
import omnigibson.utils.transform_utils as T
from omnigibson.utils.asset_utils import get_all_object_categories, get_all_object_category_models
from omnigibson.utils.ui_utils import choose_from_options


def main(random_selection=False, headless=False, short_exec=False):
    """
    Visualizes object as specified by its USD path, @usd_path. If None if specified, will instead
    result in an object selection from OmniGibson's object dataset
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Assuming that if random_selection=True, headless=True, short_exec=True, we are calling it from tests and we
    # do not want to parse args (it would fail because the calling function is pytest "testfile.py")
    usd_path = None
    if not (random_selection and headless and short_exec):
        parser = argparse.ArgumentParser()
        parser.add_argument(
            "--usd_path",
            default=None,
            help="USD Model to load",
        )
        args = parser.parse_args()
        usd_path = args.usd_path

    # Define objects to load
    light0_cfg = dict(
        type="LightObject",
        light_type="Sphere",
        name="sphere_light0",
        radius=0.01,
        intensity=1e5,
        position=[-2.0, -2.0, 2.0],
    )

    light1_cfg = dict(
        type="LightObject",
        light_type="Sphere",
        name="sphere_light1",
        radius=0.01,
        intensity=1e5,
        position=[-2.0, 2.0, 2.0],
    )

    # Make sure we have a valid usd path
    if usd_path is None:
        # Select a category to load
        available_obj_categories = get_all_object_categories()
        obj_category = choose_from_options(
            options=available_obj_categories, name="object category", random_selection=random_selection
        )

        # Select a model to load
        available_obj_models = get_all_object_category_models(obj_category)
        obj_model = choose_from_options(
            options=available_obj_models, name="object model", random_selection=random_selection
        )

        kwargs = {
            "type": "DatasetObject",
            "category": obj_category,
            "model": obj_model,
        }
    else:
        kwargs = {
            "type": "USDObject",
            "usd_path": usd_path,
        }

    # Import the desired object
    obj_cfg = dict(
        **kwargs,
        name="obj",
        usd_path=usd_path,
        visual_only=True,
        position=[0, 0, 10.0],
    )

    # Create the scene config to load -- empty scene
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [light0_cfg, light1_cfg, obj_cfg],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Set camera to appropriate viewing pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([-0.00913503, -1.95750906, 1.36407314]),
        orientation=th.tensor([0.6350064, 0.0, 0.0, 0.77250687]),
    )

    # Grab the object references
    obj = env.scene.object_registry("name", "obj")

    # Standardize the scale of the object so it fits in a [1,1,1] box -- note that we have to stop the simulator
    # in order to set the scale
    extents = obj.aabb_extent
    og.sim.stop()
    obj.scale = (th.ones(3) / extents).min()
    og.sim.play()
    env.step(th.empty(0))

    # Move the object so that its center is at [0, 0, 1]
    center_offset = obj.get_position_orientation()[0] - obj.aabb_center + th.tensor([0, 0, 1.0])
    obj.set_position_orientation(position=center_offset)

    # Allow the user to easily move the camera around
    og.sim.enable_viewer_camera_teleoperation()

    # Rotate the object in place
    steps_per_rotate = 360
    steps_per_joint = steps_per_rotate / 10
    max_steps = 100 if short_exec else 10000
    for i in range(max_steps):
        z_angle = 2 * math.pi * (i % steps_per_rotate) / steps_per_rotate
        quat = T.euler2quat(th.tensor([0, 0, z_angle]))
        pos = T.quat2mat(quat) @ center_offset
        if obj.n_dof > 0:
            frac = (i % steps_per_joint) / steps_per_joint
            j_frac = -1.0 + 2.0 * frac if (i // steps_per_joint) % 2 == 0 else 1.0 - 2.0 * frac
            obj.set_joint_positions(positions=j_frac * th.ones(obj.n_dof), normalized=True, drive=False)
            obj.keep_still()
        obj.set_position_orientation(position=pos, orientation=quat)
        env.step(th.empty(0))

    # Shut down at the end
    og.clear()


if __name__ == "__main__":
    main()

Highlight Object

This demo is useful for...

  • Understanding how to highlight individual objects within a cluttered scene
  • Understanding how to access groups of objects from the environment
python -m omnigibson.examples.objects.highlight_objects

This demo loads the Rs_int scene and highlights windows on/off repeatedly.

highlight_objects.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
import torch as th

import omnigibson as og


def main(random_selection=False, headless=False, short_exec=False):
    """
    Highlights visually all object instances of windows and then removes the highlighting
    It also demonstrates how to apply an action on all instances of objects of a given category
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene
    cfg = {
        "scene": {
            "type": "InteractiveTraversableScene",
            "scene_model": "Rs_int",
        }
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Grab all window objects
    windows = env.scene.object_registry("category", "openable_window")

    # Step environment while toggling window highlighting
    i = 0
    highlighted = False
    max_steps = -1 if not short_exec else 1000
    while i != max_steps:
        env.step(th.empty(0))
        if i % 50 == 0:
            highlighted = not highlighted
            og.log.info(f"Toggling window highlight to: {highlighted}")
            for window in windows:
                # Note that this property is R/W!
                window.highlighted = highlighted
        i += 1

    # Always close the environment at the end
    og.clear()


if __name__ == "__main__":
    main()

Draw Object Bounding Box Demo

This demo is useful for...

  • Understanding how to access observations from a GymObservable object
  • Understanding how to access objects' bounding box information
  • Understanding how to dynamically modify vision modalities

*[GymObservable]: Environment, all sensors extending from BaseSensor, and all objects extending from BaseObject (which includes all robots extending from BaseRobot!) are GymObservable objects!

python -m omnigibson.examples.objects.draw_bounding_box

This demo loads a door object and banana object, and partially obscures the banana with the door. It generates both "loose" and "tight" bounding boxes (where the latter respects occlusions) for both objects, and dumps them to an image on disk.

draw_bounding_box.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
import matplotlib.pyplot as plt
import torch as th

import omnigibson as og


def main(random_selection=False, headless=False, short_exec=False):
    """
    Shows how to obtain the bounding box of an articulated object.
    Draws the bounding box around the loaded object, a cabinet, and writes the visualized image to disk at the
    current directory named 'bbox_2d_[loose / tight]_img.png'.

    NOTE: In the GUI, bounding boxes can be natively viewed by clicking on the sensor ((*)) icon at the top,
    and then selecting the appropriate bounding box modalities, and clicking "Show". See:

    https://docs.omniverse.nvidia.com/prod_extensions/prod_extensions/ext_replicator/visualization.html#the-visualizer
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Specify objects to load
    banana_cfg = dict(
        type="DatasetObject",
        name="banana",
        category="banana",
        model="vvyyyv",
        bounding_box=[0.643, 0.224, 0.269],
        position=[-0.906661, -0.545106, 0.136824],
        orientation=[0, 0, 0.76040583, -0.6494482],
    )

    door_cfg = dict(
        type="DatasetObject",
        name="door",
        category="door",
        model="ohagsq",
        bounding_box=[1.528, 0.064, 1.299],
        position=[-2.0, 0, 0.70000001],
        orientation=[0, 0, -0.38268343, 0.92387953],
    )

    # Create the scene config to load -- empty scene with a few objects
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [banana_cfg, door_cfg],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Set camera to appropriate viewing pose
    cam = og.sim.viewer_camera
    cam.set_position_orientation(
        position=th.tensor([-4.62785, -0.418575, 0.933943]),
        orientation=th.tensor([0.52196595, -0.4231939, -0.46640436, 0.5752612]),
    )

    # Add bounding boxes to camera sensor
    bbox_modalities = ["bbox_3d", "bbox_2d_loose", "bbox_2d_tight"]
    for bbox_modality in bbox_modalities:
        cam.add_modality(bbox_modality)

    # Take a few steps to let objects settle
    for i in range(100):
        env.step(th.empty(0))

    # Grab observations from viewer camera and write them to disk
    obs, _ = cam.get_obs()

    for bbox_modality in bbox_modalities:
        # Print out each of the modalities
        og.log.info(f"Observation modality {bbox_modality}:\n{obs[bbox_modality]}")

        # Also write the 2d loose bounding box to disk
        if "3d" not in bbox_modality:
            from omnigibson.utils.deprecated_utils import colorize_bboxes

            colorized_img = colorize_bboxes(
                bboxes_2d_data=obs[bbox_modality], bboxes_2d_rgb=obs["rgb"].cpu().numpy(), num_channels=4
            )
            fpath = f"{bbox_modality}_img.png"
            plt.imsave(fpath, colorized_img)
            og.log.info(f"Saving modality [{bbox_modality}] image to: {fpath}")

    # Always close environment down at end
    og.clear()


if __name__ == "__main__":
    main()

Object States

These examples showcase OmniGibson's powerful object states functionality, which captures both individual and relational kinematic and non-kinematic states.

Slicing Demo

This demo is useful for...

  • Understanding how slicing works in OmniGibson
  • Understanding how to access individual objects once the environment is created
python -m omnigibson.examples.object_states.slicing_demo

This demo spawns an apple on a table with a knife above it, and lets the knife fall to "cut" the apple in half.

slicing_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
import math

import torch as th

import omnigibson as og
import omnigibson.utils.transform_utils as T
from omnigibson.macros import gm

# Make sure object states and transition rules are enabled
gm.ENABLE_OBJECT_STATES = True
gm.ENABLE_TRANSITION_RULES = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of slicing an apple into two apple slices
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene with table, knife, and apple
    table_cfg = dict(
        type="DatasetObject",
        name="table",
        category="breakfast_table",
        model="rjgmmy",
        bounding_box=[1.36, 1.081, 0.84],
        position=[0, 0, 0.58],
    )

    apple_cfg = dict(
        type="DatasetObject",
        name="apple",
        category="apple",
        model="agveuv",
        bounding_box=[0.098, 0.098, 0.115],
        position=[0.085, 0, 0.92],
    )

    knife_cfg = dict(
        type="DatasetObject",
        name="knife",
        category="table_knife",
        model="lrdmpf",
        bounding_box=[0.401, 0.044, 0.009],
        position=[0, 0, 20.0],
    )

    light0_cfg = dict(
        type="LightObject",
        name="light0",
        light_type="Sphere",
        radius=0.01,
        intensity=4000.0,
        position=[1.217, -0.848, 1.388],
    )

    light1_cfg = dict(
        type="LightObject",
        name="light1",
        light_type="Sphere",
        radius=0.01,
        intensity=4000.0,
        position=[-1.217, 0.848, 1.388],
    )

    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [table_cfg, apple_cfg, knife_cfg, light0_cfg, light1_cfg],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Grab reference to apple and knife
    apple = env.scene.object_registry("name", "apple")
    knife = env.scene.object_registry("name", "knife")

    # Update the simulator's viewer camera's pose so it points towards the table
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.544888, -0.412084, 1.11569]),
        orientation=th.tensor([0.54757518, 0.27792802, 0.35721896, 0.70378409]),
    )

    # Let apple settle
    for _ in range(50):
        env.step(th.empty(0))

    knife.keep_still()
    knife.set_position_orientation(
        position=apple.get_position_orientation()[0] + th.tensor([-0.15, 0.0, 0.2], dtype=th.float32),
        orientation=T.euler2quat(th.tensor([-math.pi / 2, 0, 0], dtype=th.float32)),
    )

    if not short_exec:
        input("The knife will fall on the apple and slice it. Press [ENTER] to continue.")

    # Step simulation for a bit so that apple is sliced
    for i in range(1000):
        env.step(th.empty(0))

    if not short_exec:
        input("Apple has been sliced! Press [ENTER] to terminate the demo.")

    # Always close environment at the end
    og.clear()


if __name__ == "__main__":
    main()

Dicing Demo

This demo is useful for...

  • Understanding how to leverage the Dicing state
  • Understanding how to enable objects to be diceable
python -m omnigibson.examples.object_states.dicing_demo

This demo loads an apple and a knife, and showcases how apple can be diced into smaller chunks with the knife.

dicing_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
import math

import torch as th

import omnigibson as og
import omnigibson.utils.transform_utils as T
from omnigibson.macros import gm

# Make sure object states, GPU dynamics, and transition rules are enabled
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True
gm.ENABLE_TRANSITION_RULES = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of dicing an apple into apple dices
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene with table, knife, and apple
    table_cfg = dict(
        type="DatasetObject",
        name="table",
        category="breakfast_table",
        model="rjgmmy",
        bounding_box=[1.36, 1.08, 0.84],
        position=[0, 0, 0.58],
    )

    apple_cfg = dict(
        type="DatasetObject",
        name="apple",
        category="apple",
        model="agveuv",
        bounding_box=[0.098, 0.098, 0.115],
        position=[0.085, 0, 0.92],
        abilities={"diceable": {}},
    )

    knife_cfg = dict(
        type="DatasetObject",
        name="knife",
        category="table_knife",
        model="lrdmpf",
        bounding_box=[0.401, 0.044, 0.009],
        position=[0, 0, 20.0],
    )

    light0_cfg = dict(
        type="LightObject",
        name="light0",
        light_type="Sphere",
        radius=0.01,
        intensity=1e7,
        position=[1.217, -0.848, 1.388],
    )

    light1_cfg = dict(
        type="LightObject",
        name="light1",
        light_type="Sphere",
        radius=0.01,
        intensity=1e7,
        position=[-1.217, 0.848, 1.388],
    )

    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [table_cfg, apple_cfg, knife_cfg, light0_cfg, light1_cfg],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Grab reference to apple and knife
    apple = env.scene.object_registry("name", "apple")
    knife = env.scene.object_registry("name", "knife")

    # Update the simulator's viewer camera's pose so it points towards the table
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.544888, -0.412084, 1.11569]),
        orientation=th.tensor([0.54757518, 0.27792802, 0.35721896, 0.70378409]),
    )

    # Let apple settle
    for _ in range(50):
        env.step(th.empty(0))

    knife.keep_still()
    knife.set_position_orientation(
        position=apple.get_position_orientation()[0] + th.tensor([-0.15, 0.0, 0.2]),
        orientation=T.euler2quat(th.tensor([-math.pi / 2, 0, 0])),
    )

    if not short_exec:
        input("The knife will fall on the apple and dice it. Press [ENTER] to continue.")

    # Step simulation for a bit so that apple is diced
    for _ in range(1000):
        env.step(th.empty(0))

    if not short_exec:
        input("Apple has been diced! Press [ENTER] to terminate the demo.")

    # Always close simulator at the end
    og.clear()


if __name__ == "__main__":
    main()

Folded and Unfolded Demo

This demo is useful for...

  • Understanding how to load a softbody (cloth) version of a BEHAVIOR dataset object
  • Understanding how to enable cloth objects to be foldable
  • Understanding the current heuristics used for gauging a cloth's "foldness"
python -m omnigibson.examples.object_states.folded_unfolded_state_demo

This demo loads in three different cloth objects, and allows you to manipulate them while printing out their Folded state status in real-time. Try manipulating the object by holding down Shift and then Left-click + Drag!

folded_unfolded_state_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
import torch as th

import omnigibson as og
from omnigibson.macros import gm
from omnigibson.object_states import Folded, Unfolded
from omnigibson.utils.constants import PrimType
from omnigibson.utils.python_utils import multi_dim_linspace

# Make sure object states and GPU dynamics are enabled (GPU dynamics needed for cloth)
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of cloth objects that can potentially be folded.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene + custom cloth object
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [
            {
                "type": "DatasetObject",
                "name": "carpet",
                "category": "carpet",
                "model": "ctclvd",
                "bounding_box": [0.897, 0.568, 0.012],
                "prim_type": PrimType.CLOTH,
                "abilities": {"cloth": {}},
                "position": [0, 0, 0.5],
            },
            {
                "type": "DatasetObject",
                "name": "dishtowel",
                "category": "dishtowel",
                "model": "dtfspn",
                "bounding_box": [0.852, 1.1165, 0.174],
                "prim_type": PrimType.CLOTH,
                "abilities": {"cloth": {}},
                "position": [1, 1, 0.5],
            },
            {
                "type": "DatasetObject",
                "name": "shirt",
                "category": "t_shirt",
                "model": "kvidcx",
                "bounding_box": [0.472, 1.243, 1.158],
                "prim_type": PrimType.CLOTH,
                "abilities": {"cloth": {}},
                "position": [-1, 1, 0.5],
                "orientation": [0.7071, 0.0, 0.7071, 0.0],
            },
        ],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Grab object references
    carpet = env.scene.object_registry("name", "carpet")
    dishtowel = env.scene.object_registry("name", "dishtowel")
    shirt = env.scene.object_registry("name", "shirt")
    objs = [carpet, dishtowel, shirt]

    # Set viewer camera
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.46382895, -2.66703958, 1.22616824]),
        orientation=th.tensor([0.58779174, -0.00231237, -0.00318273, 0.80900271]),
    )

    def print_state():
        folded = carpet.states[Folded].get_value()
        unfolded = carpet.states[Unfolded].get_value()
        info = "carpet: [folded] %d [unfolded] %d" % (folded, unfolded)

        folded = dishtowel.states[Folded].get_value()
        unfolded = dishtowel.states[Unfolded].get_value()
        info += " || dishtowel: [folded] %d [unfolded] %d" % (folded, unfolded)

        folded = shirt.states[Folded].get_value()
        unfolded = shirt.states[Unfolded].get_value()
        info += " || tshirt: [folded] %d [unfolded] %d" % (folded, unfolded)

        print(f"{info}{' ' * (110 - len(info))}", end="\r")

    for _ in range(100):
        og.sim.step()

    print("\nCloth state:\n")

    if not short_exec:
        # Fold all three cloths along the x-axis
        for i in range(3):
            obj = objs[i]
            pos = obj.root_link.compute_particle_positions()
            x_min, x_max = th.min(pos, dim=0).values[0], th.max(pos, dim=0).values[0]
            x_extent = x_max - x_min
            # Get indices for the bottom 10 percent vertices in the x-axis
            indices = th.argsort(pos, dim=0)[:, 0][: (pos.shape[0] // 10)]
            start = th.clone(pos[indices])

            # lift up a bit
            mid = th.clone(start)
            mid[:, 2] += x_extent * 0.2

            # move towards x_max
            end = th.clone(mid)
            end[:, 0] += x_extent * 0.9

            increments = 25
            for ctrl_pts in th.cat(
                [multi_dim_linspace(start, mid, increments), multi_dim_linspace(mid, end, increments)]
            ):
                obj.root_link.set_particle_positions(ctrl_pts, idxs=indices)
                og.sim.step()
                print_state()

        # Fold the t-shirt twice again along the y-axis
        for direction in [-1, 1]:
            obj = shirt
            pos = obj.root_link.compute_particle_positions()
            y_min, y_max = th.min(pos, dim=0).values[1], th.max(pos, dim=0).values[1]
            y_extent = y_max - y_min
            if direction == 1:
                indices = th.argsort(pos, dim=0)[:, 1][: (pos.shape[0] // 20)]
            else:
                indices = th.argsort(pos, dim=0)[:, 1][-(pos.shape[0] // 20) :]
            start = th.clone(pos[indices])

            # lift up a bit
            mid = th.clone(start)
            mid[:, 2] += y_extent * 0.2

            # move towards y_max
            end = th.clone(mid)
            end[:, 1] += direction * y_extent * 0.4

            increments = 25
            for ctrl_pts in th.cat(
                [multi_dim_linspace(start, mid, increments), multi_dim_linspace(mid, end, increments)]
            ):
                obj.root_link.set_particle_positions(ctrl_pts, idxs=indices)
                env.step(th.empty(0))
                print_state()

        while True:
            env.step(th.empty(0))
            print_state()

    # Shut down env at the end
    print()
    og.clear()


if __name__ == "__main__":
    main()

Overlaid Demo

This demo is useful for...

  • Understanding how cloth objects can be overlaid on rigid objects
  • Understanding current heuristics used for gauging a cloth's "overlaid" status
python -m omnigibson.examples.object_states.overlaid_demo

This demo loads in a carpet on top of a table. The demo allows you to manipulate the carpet while printing out their Overlaid state status in real-time. Try manipulating the object by holding down Shift and then Left-click + Drag!

overlaid_demo.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
import torch as th

import omnigibson as og
from omnigibson.macros import gm
from omnigibson.object_states import Overlaid
from omnigibson.utils.constants import PrimType

# Make sure object states and GPU dynamics are enabled (GPU dynamics needed for cloth)
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of cloth objects that can be overlaid on rigid objects.

    Loads a carpet on top of a table. Initially Overlaid will be True because the carpet largely covers the table.
    If you drag the carpet off the table or even just fold it into half, Overlaid will become False.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene + custom cloth object + custom rigid object
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [
            {
                "type": "DatasetObject",
                "name": "carpet",
                "category": "carpet",
                "model": "ctclvd",
                "bounding_box": [1.346, 0.852, 0.017],
                "prim_type": PrimType.CLOTH,
                "abilities": {"cloth": {}},
                "position": [0, 0, 1.0],
            },
            {
                "type": "DatasetObject",
                "name": "breakfast_table",
                "category": "breakfast_table",
                "model": "rjgmmy",
                "bounding_box": [1.36, 1.081, 0.84],
                "prim_type": PrimType.RIGID,
                "position": [0, 0, 0.58],
            },
        ],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Grab object references
    carpet = env.scene.object_registry("name", "carpet")
    breakfast_table = env.scene.object_registry("name", "breakfast_table")

    # Set camera pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.88215526, -1.40086216, 2.00311063]),
        orientation=th.tensor([0.42013364, 0.12342107, 0.25339685, 0.86258043]),
    )

    max_steps = 100 if short_exec else -1
    steps = 0

    print("\nTry dragging cloth around with CTRL + Left-Click to see the Overlaid state change:\n")

    while steps != max_steps:
        print(f"Overlaid {carpet.states[Overlaid].get_value(breakfast_table)}    ", end="\r")
        env.step(th.empty(0))
        steps += 1

    # Shut down env at the end
    og.clear()


if __name__ == "__main__":
    main()

Heat Source or Sink Demo

This demo is useful for...

  • Understanding how a heat source (or sink) is visualized in OmniGibson
  • Understanding how dynamic fire visuals are generated in real-time
python -m omnigibson.examples.object_states.heat_source_or_sink_demo

This demo loads in a stove and toggles its HeatSource on and off, showcasing the dynamic fire visuals available in OmniGibson.

heat_source_or_sink_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True


def main(random_selection=False, headless=False, short_exec=False):
    # Create the scene config to load -- empty scene with a stove object added
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": [
            {
                "type": "DatasetObject",
                "name": "stove",
                "category": "stove",
                "model": "qbjiva",
                "bounding_box": [1.611, 0.769, 1.147],
                "abilities": {
                    "heatSource": {"requires_toggled_on": True},
                    "toggleable": {},
                },
                "position": [0, 0, 0.61],
            }
        ],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Get reference to stove object
    stove = env.scene.object_registry("name", "stove")

    # Set camera to appropriate viewing pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([-0.0792399, -1.30104, 1.51981]),
        orientation=th.tensor([0.54897692, 0.00110359, 0.00168013, 0.83583509]),
    )

    # Make sure necessary object states are included with the stove
    assert object_states.HeatSourceOrSink in stove.states
    assert object_states.ToggledOn in stove.states

    # Take a few steps so that visibility propagates
    for _ in range(5):
        env.step(th.empty(0))

    # Heat source is off.
    print("Heat source is OFF.")
    heat_source_state = stove.states[object_states.HeatSourceOrSink].get_value()
    assert not heat_source_state

    # Toggle on stove, notify user
    if not short_exec:
        input("Heat source will now turn ON: Press ENTER to continue.")
    stove.states[object_states.ToggledOn].set_value(True)

    assert stove.states[object_states.ToggledOn].get_value()

    # Need to take a step to update the state.
    env.step(th.empty(0))

    # Heat source is on
    heat_source_state = stove.states[object_states.HeatSourceOrSink].get_value()
    assert heat_source_state
    for _ in range(500):
        env.step(th.empty(0))

    # Toggle off stove, notify user
    if not short_exec:
        input("Heat source will now turn OFF: Press ENTER to continue.")
    stove.states[object_states.ToggledOn].set_value(False)
    assert not stove.states[object_states.ToggledOn].get_value()
    for _ in range(200):
        env.step(th.empty(0))

    # Move stove, notify user
    if not short_exec:
        input("Heat source is now moving: Press ENTER to continue.")
    stove.set_position_orientation(position=th.tensor([0, 1.0, 0.61]))
    for i in range(100):
        env.step(th.empty(0))

    # Toggle on stove again, notify user
    if not short_exec:
        input("Heat source will now turn ON: Press ENTER to continue.")
    stove.states[object_states.ToggledOn].set_value(True)
    assert stove.states[object_states.ToggledOn].get_value()
    for i in range(500):
        env.step(th.empty(0))

    # Shutdown environment at end
    og.clear()


if __name__ == "__main__":
    main()

Temperature Demo

This demo is useful for...

  • Understanding how to dynamically sample kinematic states for BEHAVIOR dataset objects
  • Understanding how temperature changes are propagated to individual objects from individual heat sources or sinks
python -m omnigibson.examples.object_states.temperature_demo

This demo loads in various heat sources and sinks, and places an apple within close proximity to each of them. As the environment steps, each apple's temperature is printed in real-time, showcasing OmniGibson's rudimentary temperature dynamics.

temperature_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of temperature change
    Loads a stove, a microwave and an oven, all toggled on, and five frozen apples
    The user can move the apples to see them change from frozen, to normal temperature, to cooked and burnt
    This demo also shows how to load objects ToggledOn and how to set the initial temperature of an object
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Define specific objects we want to load in with the scene directly
    obj_configs = []

    # Light
    obj_configs.append(
        dict(
            type="LightObject",
            light_type="Sphere",
            name="light",
            radius=0.01,
            intensity=1e8,
            position=[-2.0, -2.0, 1.0],
        )
    )

    # Stove
    obj_configs.append(
        dict(
            type="DatasetObject",
            name="stove",
            category="stove",
            model="yhjzwg",
            bounding_box=[1.185, 0.978, 1.387],
            position=[0, 0, 0.69],
        )
    )

    # Microwave
    obj_configs.append(
        dict(
            type="DatasetObject",
            name="microwave",
            category="microwave",
            model="hjjxmi",
            bounding_box=[0.384, 0.256, 0.196],
            position=[2.5, 0, 0.10],
        )
    )

    # Oven
    obj_configs.append(
        dict(
            type="DatasetObject",
            name="oven",
            category="oven",
            model="wuinhm",
            bounding_box=[1.075, 0.926, 1.552],
            position=[-1.25, 0, 0.88],
        )
    )

    # Tray
    obj_configs.append(
        dict(
            type="DatasetObject",
            name="tray",
            category="tray",
            model="xzcnjq",
            bounding_box=[0.319, 0.478, 0.046],
            position=[-0.25, -0.12, 1.26],
        )
    )

    # Fridge
    obj_configs.append(
        dict(
            type="DatasetObject",
            name="fridge",
            category="fridge",
            model="hivvdf",
            bounding_box=[1.065, 1.149, 1.528],
            abilities={
                "coldSource": {
                    "temperature": -100.0,
                    "requires_inside": True,
                }
            },
            position=[1.25, 0, 0.81],
        )
    )

    # 5 Apples
    for i in range(5):
        obj_configs.append(
            dict(
                type="DatasetObject",
                name=f"apple{i}",
                category="apple",
                model="agveuv",
                bounding_box=[0.065, 0.065, 0.077],
                position=[0, i * 0.1, 5.0],
            )
        )

    # Create the scene config to load -- empty scene with desired objects
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": obj_configs,
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Get reference to relevant objects
    stove = env.scene.object_registry("name", "stove")
    microwave = env.scene.object_registry("name", "microwave")
    oven = env.scene.object_registry("name", "oven")
    tray = env.scene.object_registry("name", "tray")
    fridge = env.scene.object_registry("name", "fridge")
    apples = list(env.scene.object_registry("category", "apple"))

    # Set camera to appropriate viewing pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.46938863, -3.97887141, 1.64106008]),
        orientation=th.tensor([0.63311689, 0.00127259, 0.00155577, 0.77405359]),
    )

    # Let objects settle
    for _ in range(25):
        env.step(th.empty(0))

    # Turn on all scene objects
    stove.states[object_states.ToggledOn].set_value(True)
    microwave.states[object_states.ToggledOn].set_value(True)
    oven.states[object_states.ToggledOn].set_value(True)

    # Set initial temperature of the apples to -50 degrees Celsius, and move the apples to different objects
    for apple in apples:
        apple.states[object_states.Temperature].set_value(-50)
    apples[0].states[object_states.Inside].set_value(oven, True)
    apples[1].set_position_orientation(
        position=stove.states[object_states.HeatSourceOrSink].link.get_position_orientation()[0]
        + th.tensor([0, 0, 0.1])
    )
    apples[2].states[object_states.OnTop].set_value(tray, True)
    apples[3].states[object_states.Inside].set_value(fridge, True)
    apples[4].states[object_states.Inside].set_value(microwave, True)

    steps = 0
    max_steps = -1 if not short_exec else 1000

    # Main recording loop
    locations = [f"{loc:>20}" for loc in ["Inside oven", "On stove", "On tray", "Inside fridge", "Inside microwave"]]
    print()
    print(f"{'Apple location:':<20}", *locations)
    while steps != max_steps:
        env.step(th.empty(0))
        temps = [f"{apple.states[object_states.Temperature].get_value():>20.2f}" for apple in apples]
        print(f"{'Apple temperature:':<20}", *temps, end="\r")
        steps += 1

    # Always close env at the end
    og.clear()


if __name__ == "__main__":
    main()

Heated Demo

This demo is useful for...

  • Understanding how temperature modifications can cause objects' visual changes
  • Understanding how dynamic steam visuals are generated in real-time
python -m omnigibson.examples.object_states.heated_state_demo

This demo loads in three bowls, and immediately sets their temperatures past their Heated threshold. Steam is generated in real-time from these objects, and then disappears once the temperature of the objects drops below their Heated threshold.

heated_state_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True


def main(random_selection=False, headless=False, short_exec=False):
    # Define object configurations for objects to load -- we want to load a light and three bowls
    obj_configs = []

    obj_configs.append(
        dict(
            type="LightObject",
            light_type="Sphere",
            name="light",
            radius=0.01,
            intensity=1e8,
            position=[-2.0, -2.0, 1.0],
        )
    )

    for i, (scale, x) in enumerate(zip([0.5, 1.0, 2.0], [-0.6, 0, 0.8])):
        obj_configs.append(
            dict(
                type="DatasetObject",
                name=f"bowl{i}",
                category="bowl",
                model="ajzltc",
                bounding_box=th.tensor([0.329, 0.293, 0.168]) * scale,
                abilities={"heatable": {}},
                position=[x, 0, 0.2],
            )
        )

    # Create the scene config to load -- empty scene with light object and bowls
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": obj_configs,
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Set camera to appropriate viewing pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.182103, -2.07295, 0.14017]),
        orientation=th.tensor([0.77787037, 0.00267566, 0.00216149, 0.62841535]),
    )

    # Dim the skybox so we can see the bowls' steam effectively
    og.sim.skybox.intensity = 100.0

    # Grab reference to objects of relevance
    objs = list(env.scene.object_registry("category", "bowl"))

    def report_states(objs):
        for obj in objs:
            print("=" * 20)
            print("object:", obj.name)
            print("temperature:", obj.states[object_states.Temperature].get_value())
            print("obj is heated:", obj.states[object_states.Heated].get_value())

    # Report default states
    print("==== Initial state ====")
    report_states(objs)

    if not short_exec:
        # Notify user that we're about to heat the object
        input("Objects will be heated, and steam will slowly rise. Press ENTER to continue.")

    # Heated.
    for obj in objs:
        obj.states[object_states.Temperature].set_value(50)
    env.step(th.empty(0))
    report_states(objs)

    # Take a look at the steam effect.
    # After a while, objects will be below the Steam temperature threshold.
    print("==== Objects are now heated... ====")
    print()
    max_iterations = 2000 if not short_exec else 100
    for _ in range(max_iterations):
        env.step(th.empty(0))
        # Also print temperatures
        temps = [f"{obj.states[object_states.Temperature].get_value():>7.2f}" for obj in objs]
        print("obj temps:", *temps, end="\r")
    print()

    # Objects are not heated anymore.
    print("==== Objects are no longer heated... ====")
    report_states(objs)

    if not short_exec:
        # Close environment at the end
        input("Demo completed. Press ENTER to shutdown environment.")

    og.clear()


if __name__ == "__main__":
    main()

Onfire Demo

This demo is useful for...

  • Understanding how changing onfire state can cause objects' visual changes
  • Understanding how onfire can be triggered by nearby onfire objects
python -m omnigibson.examples.object_states.onfire_demo

This demo loads in a stove (toggled on) and two apples. The first apple will be ignited by the stove first, then the second apple will be ignited by the first apple.

onfire_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of on fire state.
    Loads a stove (toggled on), and two apples.
    The first apple will be ignited by the stove first, then the second apple will be ignited by the first apple.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Define specific objects we want to load in with the scene directly
    obj_configs = []

    # Light
    obj_configs.append(
        dict(
            type="LightObject",
            light_type="Sphere",
            name="light",
            radius=0.01,
            intensity=1e8,
            position=[-2.0, -2.0, 1.0],
        )
    )

    # Stove
    obj_configs.append(
        dict(
            type="DatasetObject",
            name="stove",
            category="stove",
            model="yhjzwg",
            bounding_box=[1.185, 0.978, 1.387],
            position=[0, 0, 0.69],
        )
    )

    # 2 Apples
    for i in range(2):
        obj_configs.append(
            dict(
                type="DatasetObject",
                name=f"apple{i}",
                category="apple",
                model="agveuv",
                bounding_box=[0.065, 0.065, 0.077],
                position=[0, i * 0.07, 2.0],
                abilities={"flammable": {"ignition_temperature": 100, "distance_threshold": 0.5}},
            )
        )

    # Create the scene config to load -- empty scene with desired objects
    cfg = {
        "scene": {
            "type": "Scene",
        },
        "objects": obj_configs,
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Get reference to relevant objects
    stove = env.scene.object_registry("name", "stove")
    apples = list(env.scene.object_registry("category", "apple"))

    # Set camera to appropriate viewing pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([-0.42246569, -0.34745704, 1.56810353]),
        orientation=th.tensor([0.50083786, -0.10407796, -0.17482619, 0.84128772]),
    )

    # Let objects settle
    for _ in range(10):
        env.step(th.empty(0))

    # Turn on the stove
    stove.states[object_states.ToggledOn].set_value(True)

    # The first apple will be affected by the stove
    apples[0].set_position_orientation(
        position=stove.states[object_states.HeatSourceOrSink].link.get_position_orientation()[0]
        + th.tensor([0.11, 0, 0.1])
    )

    # The second apple will NOT be affected by the stove, but will be affected by the first apple once it's on fire.
    apples[1].set_position_orientation(
        position=stove.states[object_states.HeatSourceOrSink].link.get_position_orientation()[0]
        + th.tensor([0.32, 0, 0.1])
    )

    steps = 0
    max_steps = -1 if not short_exec else 1000

    # Main recording loop
    while steps != max_steps:
        env.step(th.empty(0))
        temps = [f"{apple.states[object_states.Temperature].get_value():>20.2f}" for apple in apples]
        print(f"{'Apple temperature:':<20}", *temps, end="\r")
        steps += 1

    # Always close env at the end
    og.clear()


if __name__ == "__main__":
    main()

Particle Applier and Remover Demo

This demo is useful for...

  • Understanding how a ParticleRemover or ParticleApplier object can be generated
  • Understanding how particles can be dynamically generated on objects
  • Understanding different methods for applying and removing particles via the ParticleRemover or ParticleApplier object
python -m omnigibson.examples.object_states.particle_applier_remover_demo

This demo loads in a washtowel and table and lets you choose the ability configuration to enable the washtowel with. The washtowel will then proceed to either remove and generate particles dynamically on the table while moving.

particle_applier_remover_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
import torch as th

import omnigibson as og
from omnigibson.macros import gm, macros
from omnigibson.object_states import Covered, ToggledOn
from omnigibson.utils.ui_utils import choose_from_options

# Set macros for this example
macros.object_states.particle_modifier.VISUAL_PARTICLES_REMOVAL_LIMIT = 1000
macros.object_states.particle_modifier.PHYSICAL_PARTICLES_REMOVAL_LIMIT = 8000
macros.object_states.particle_modifier.MAX_VISUAL_PARTICLES_APPLIED_PER_STEP = 4
macros.object_states.particle_modifier.MAX_PHYSICAL_PARTICLES_APPLIED_PER_STEP = 40
macros.object_states.covered.MAX_VISUAL_PARTICLES = 300

# Make sure object states and GPU dynamics are enabled (GPU dynamics needed for fluids)
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True
gm.ENABLE_HQ_RENDERING = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of ParticleApplier and ParticleRemover object states, which enable objects to either apply arbitrary
    particles and remove arbitrary particles from the simulator, respectively.

    Loads an empty scene with a table, and starts clean to allow particles to be applied or pre-covers the table
    with particles to be removed. The ParticleApplier / ParticleRemover state is applied to an imported cloth object
    and allowed to interact with the table, applying / removing particles from the table.

    NOTE: The key difference between ParticleApplier/Removers and ParticleSource/Sinks is that Applier/Removers
    requires contact (if using ParticleProjectionMethod.ADJACENCY) or overlap
    (if using ParticleProjectionMethod.PROJECTION) in order to spawn / remove particles, and generally only spawn
    particles at the contact points. ParticleSource/Sinks are special cases of ParticleApplier/Removers that
    always use ParticleProjectionMethod.PROJECTION and always spawn / remove particles within their projection volume,
    irregardless of overlap with other objects!
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Choose what configuration to load
    modifier_type = choose_from_options(
        options={
            "particleApplier": "Demo object's ability to apply particles in the simulator",
            "particleRemover": "Demo object's ability to remove particles from the simulator",
        },
        name="particle modifier type",
        random_selection=random_selection,
    )
    particle_types = ["salt", "water"]
    particle_type = choose_from_options(
        options={name: f"{name} particles will be applied or removed from the simulator" for name in particle_types},
        name="particle type",
        random_selection=random_selection,
    )

    table_cfg = dict(
        type="DatasetObject",
        name="table",
        category="breakfast_table",
        model="kwmfdg",
        bounding_box=[3.402, 1.745, 1.175],
        position=[0, 0, 0.98],
    )
    tool_cfg = dict(
        type="DatasetObject",
        name="tool",
        visual_only=True,
        position=[0, 0.3, 5.0],
    )

    if modifier_type == "particleRemover":
        if particle_type == "salt":
            # only ask this question if the modifier type is salt particleRemover
            method_type = choose_from_options(
                options={
                    "Adjacency": "Close proximity to the object will be used to determine whether particles can be applied / removed",
                    "Projection": "A Cone or Cylinder shape protruding from the object will be used to determine whether particles can be applied / removed",
                },
                name="modifier method type",
                random_selection=random_selection,
            )
        else:
            # If the particle type is water, the remover is always adjacency type
            method_type = "Adjacency"
        if method_type == "Adjacency":
            # use dishtowel to remove adjacent particles
            tool_cfg["category"] = "dishtowel"
            tool_cfg["model"] = "dtfspn"
            tool_cfg["bounding_box"] = [0.34245, 0.46798, 0.07]
        elif method_type == "Projection":
            # use vacuum to remove projections particles
            tool_cfg["category"] = "vacuum"
            tool_cfg["model"] = "wikhik"
    else:
        # If the modifier type is particleApplier, the applier is always projection type
        method_type = "Projection"

        if particle_type == "salt":
            # use salt shaker to apply salt particles
            tool_cfg["category"] = "salt_shaker"
            tool_cfg["model"] = "iomwtn"
        else:
            # use water atomizer to apply water particles
            tool_cfg["category"] = "water_atomizer"
            tool_cfg["model"] = "lfarai"

    # Create the scene config to load -- empty scene with a light and table
    cfg = {
        "env": {
            "rendering_frequency": 60,  # for HQ rendering
        },
        "scene": {
            "type": "Scene",
        },
        "objects": [table_cfg, tool_cfg],
    }

    # Sanity check inputs: Remover + Adjacency + Fluid will not work because we are using a visual_only
    # object, so contacts will not be triggered with this object

    # Load the environment, then immediately stop the simulator since we need to add in the modifier object
    env = og.Environment(configs=cfg)
    og.sim.stop()

    # Grab references to table and tool
    table = env.scene.object_registry("name", "table")
    tool = env.scene.object_registry("name", "tool")

    # Set the viewer camera appropriately
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([-1.61340969, -1.79803028, 2.53167412]),
        orientation=th.tensor([0.46291845, -0.12381886, -0.22679218, 0.84790371]),
    )

    # Play the simulator and take some environment steps to let the objects settle
    og.sim.play()
    for _ in range(25):
        env.step(th.empty(0))

    # If we're removing particles, set the table's covered state to be True
    if modifier_type == "particleRemover":
        table.states[Covered].set_value(env.scene.get_system(particle_type), True)
        # Take a few steps to let particles settle
        for _ in range(25):
            env.step(th.empty(0))

    # If the particle remover/applier is projection type, set the turn on shaker
    if method_type == "Projection":
        tool.states[ToggledOn].set_value(True)

    # Enable camera teleoperation for convenience
    og.sim.enable_viewer_camera_teleoperation()

    tool.keep_still()

    # Set the modifier object to be in position to modify particles
    if modifier_type == "particleRemover" and method_type == "Projection":
        tool.set_position_orientation(
            position=[0, 0.3, 1.45],
            orientation=[0, 0, 0, 1.0],
        )
    elif modifier_type == "particleRemover" and method_type == "Adjacency":
        tool.set_position_orientation(
            position=[0, 0.3, 1.175],
            orientation=[0, 0, 0, 1.0],
        )
    elif modifier_type == "particleApplier" and particle_type == "water":
        tool.set_position_orientation(
            position=[0, 0.3, 1.4],
            orientation=[0.3827, 0, 0, 0.9239],
        )
    else:
        tool.set_position_orientation(
            position=[0, 0.3, 1.5],
            orientation=[0.7071, 0, 0.7071, 0],
        )

    # Move object in square around table
    deltas = [
        [130, th.tensor([-0.01, 0, 0])],
        [60, th.tensor([0, -0.01, 0])],
        [130, th.tensor([0.01, 0, 0])],
        [60, th.tensor([0, 0.01, 0])],
    ]
    for t, delta in deltas:
        for _ in range(t):
            tool.set_position_orientation(position=tool.get_position_orientation()[0] + delta)
            env.step(th.empty(0))

    # Always shut down environment at the end
    og.clear()


if __name__ == "__main__":
    main()

Particle Source and Sink Demo

This demo is useful for...

  • Understanding how a ParticleSource or ParticleSink object can be generated
  • Understanding how particles can be dynamically generated and destroyed via such objects
python -m omnigibson.examples.object_states.particle_source_sink_demo

This demo loads in a sink, which is enabled with both the ParticleSource and ParticleSink states. The sink's particle source is located at the faucet spout and spawns a continuous stream of water particles, which is then destroyed ("sunk") by the sink's particle sink located at the drain.

Difference between ParticleApplier/Removers and ParticleSource/Sinks

The key difference between ParticleApplier/Removers and ParticleSource/Sinks is that Applier/Removers requires contact (if using ParticleProjectionMethod.ADJACENCY) or overlap (if using ParticleProjectionMethod.PROJECTION) in order to spawn / remove particles, and generally only spawn particles at the contact points. ParticleSource/Sinks are special cases of ParticleApplier/Removers that always use ParticleProjectionMethod.PROJECTION and always spawn / remove particles within their projection volume, irregardless of overlap with other objects.

particle_source_sink_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm
from omnigibson.utils.constants import ParticleModifyCondition

# Make sure object states are enabled and GPU dynamics are used
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True
gm.ENABLE_HQ_RENDERING = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of ParticleSource and ParticleSink object states, which enable objects to either spawn arbitrary
    particles and remove arbitrary particles from the simulator, respectively.

    Loads an empty scene with a sink, which is enabled with both the ParticleSource and ParticleSink states.
    The sink's particle source is located at the faucet spout and spawns a continuous stream of water particles,
    which is then destroyed ("sunk") by the sink's particle sink located at the drain.

    NOTE: The key difference between ParticleApplier/Removers and ParticleSource/Sinks is that Applier/Removers
    requires contact (if using ParticleProjectionMethod.ADJACENCY) or overlap
    (if using ParticleProjectionMethod.PROJECTION) in order to spawn / remove particles, and generally only spawn
    particles at the contact points. ParticleSource/Sinks are special cases of ParticleApplier/Removers that
    always use ParticleProjectionMethod.PROJECTION and always spawn / remove particles within their projection volume,
    irregardless of overlap with other objects!
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene
    cfg = {
        "env": {
            "rendering_frequency": 60,  # for HQ rendering
        },
        "scene": {
            "type": "Scene",
        },
    }

    # Define objects to load into the environment
    sink_cfg = dict(
        type="DatasetObject",
        name="sink",
        category="sink",
        model="egwapq",
        bounding_box=[2.427, 0.625, 1.2],
        abilities={
            "toggleable": {},
            "particleSource": {
                "conditions": {
                    "water": [
                        (ParticleModifyCondition.TOGGLEDON, True)
                    ],  # Must be toggled on for water source to be active
                },
                "initial_speed": 0.0,  # Water merely falls out of the spout
            },
            "particleSink": {
                "conditions": {
                    "water": [],  # No conditions, always sinking nearby particles
                },
            },
        },
        position=[0.0, 0, 0.42],
    )

    cfg["objects"] = [sink_cfg]

    # Create the environment!
    env = og.Environment(configs=cfg)

    # Set camera to ideal angle for viewing objects
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([0.37860532, -0.65396566, 1.4067066]),
        orientation=th.tensor([0.49909498, 0.15201752, 0.24857062, 0.81609284]),
    )

    # Take a few steps to let the objects settle, and then turn on the sink
    for _ in range(10):
        env.step(th.empty(0))  # Empty action since no robots are in the scene

    sink = env.scene.object_registry("name", "sink")
    assert sink.states[object_states.ToggledOn].set_value(True)

    # Take a step, and save the state
    env.step(th.empty(0))
    initial_state = og.sim.dump_state()

    # Main simulation loop.
    max_steps = 1000
    max_iterations = -1 if not short_exec else 1
    iteration = 0

    try:
        while iteration != max_iterations:
            # Keep stepping until table or bowl are clean, or we reach 1000 steps
            steps = 0
            while steps != max_steps:
                steps += 1
                env.step(th.empty(0))
            og.log.info("Max steps reached; resetting.")

            # Reset to the initial state
            og.sim.load_state(initial_state)

            iteration += 1

    finally:
        # Always shut down environment at the end
        og.clear()


if __name__ == "__main__":
    main()

Kinematics Demo

This demo is useful for...

  • Understanding how to dynamically sample kinematic states for BEHAVIOR dataset objects
  • Understanding how to import additional objects after the environment is created
python -m omnigibson.examples.object_states.sample_kinematics_demo

This demo procedurally generates a mini populated scene, spawning in a cabinet and placing boxes in its shelves, and then generating a microwave on a cabinet with a plate and apples sampled both inside and on top of it.

sample_kinematics_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo to use the raycasting-based sampler to load objects onTop and/or inside another
    Loads a cabinet, a microwave open on top of it, and two plates with apples on top, one inside and one on top of the cabinet
    Then loads a shelf and cracker boxes inside of it
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Create the scene config to load -- empty scene
    cfg = {
        "scene": {
            "type": "Scene",
        },
    }

    # Define objects we want to sample at runtime
    microwave_cfg = dict(
        type="DatasetObject",
        name="microwave",
        category="microwave",
        model="hjjxmi",
        bounding_box=[0.768, 0.512, 0.392],
    )

    cabinet_cfg = dict(
        type="DatasetObject",
        name="cabinet",
        category="bottom_cabinet",
        model="bamfsz",
        bounding_box=[1.075, 1.131, 1.355],
    )

    plate_cfgs = [
        dict(
            type="DatasetObject",
            name=f"plate{i}",
            category="plate",
            model="iawoof",
            bounding_box=th.tensor([0.20, 0.20, 0.05]),
        )
        for i in range(2)
    ]

    apple_cfgs = [
        dict(
            type="DatasetObject",
            name=f"apple{i}",
            category="apple",
            model="agveuv",
            bounding_box=[0.065, 0.065, 0.077],
        )
        for i in range(4)
    ]

    shelf_cfg = dict(
        type="DatasetObject",
        name="shelf",
        category="shelf",
        model="pkgbcp",
        bounding_box=th.tensor([1.0, 0.4, 2.0]),
    )

    box_cfgs = [
        dict(
            type="DatasetObject",
            name=f"box{i}",
            category="box_of_crackers",
            model="cmdigf",
            bounding_box=th.tensor([0.2, 0.05, 0.3]),
        )
        for i in range(5)
    ]

    # Compose objects cfg
    objects_cfg = [
        microwave_cfg,
        cabinet_cfg,
        *plate_cfgs,
        *apple_cfgs,
        shelf_cfg,
        *box_cfgs,
    ]

    # Update their spawn positions so they don't collide immediately
    for i, obj_cfg in enumerate(objects_cfg):
        obj_cfg["position"] = [100 + i, 100 + i, 100 + i]

    cfg["objects"] = objects_cfg

    # Create the environment
    env = og.Environment(configs=cfg)
    env.step([])

    # Sample microwave and boxes
    sample_boxes_on_shelf(env)
    sample_microwave_plates_apples(env)

    max_steps = 100 if short_exec else -1
    step = 0
    while step != max_steps:
        env.step(th.empty(0))
        step += 1

    # Always close environment at the end
    og.clear()


def sample_microwave_plates_apples(env):
    microwave = env.scene.object_registry("name", "microwave")
    cabinet = env.scene.object_registry("name", "cabinet")
    plates = list(env.scene.object_registry("category", "plate"))
    apples = list(env.scene.object_registry("category", "apple"))

    # Place the cabinet at a pre-determined location on the floor
    og.log.info("Placing cabinet on the floor...")
    cabinet.set_orientation([0, 0, 0, 1.0])
    env.step(th.empty(0))
    offset = cabinet.get_position_orientation()[0][2] - cabinet.aabb_center[2]
    cabinet.set_position_orientation(position=th.tensor([1.0, 0, cabinet.aabb_extent[2] / 2]) + offset)
    env.step(th.empty(0))

    # Set microwave on top of the cabinet, open it, and step 100 times
    og.log.info("Placing microwave OnTop of the cabinet...")
    assert microwave.states[object_states.OnTop].set_value(cabinet, True)
    assert microwave.states[object_states.Open].set_value(True)
    og.log.info("Microwave placed.")
    for _ in range(50):
        env.step(th.empty(0))

    og.log.info("Placing plates")
    n_apples_per_plate = int(len(apples) / len(plates))
    for i, plate in enumerate(plates):
        # Put the 1st plate in the microwave
        if i == 0:
            og.log.info(f"Placing plate {i} Inside the microwave...")
            assert plate.states[object_states.Inside].set_value(microwave, True)
        else:
            og.log.info(f"Placing plate {i} OnTop the microwave...")
            assert plate.states[object_states.OnTop].set_value(microwave, True)

        og.log.info(f"Plate {i} placed.")
        for _ in range(50):
            env.step(th.empty(0))

        og.log.info(f"Placing {n_apples_per_plate} apples OnTop of the plate...")
        for j in range(n_apples_per_plate):
            apple_idx = i * n_apples_per_plate + j
            apple = apples[apple_idx]
            assert apple.states[object_states.OnTop].set_value(plate, True)
            og.log.info(f"Apple {apple_idx} placed.")
            for _ in range(50):
                env.step(th.empty(0))


def sample_boxes_on_shelf(env):
    shelf = env.scene.object_registry("name", "shelf")
    boxes = list(env.scene.object_registry("category", "box_of_crackers"))
    # Place the shelf at a pre-determined location on the floor
    og.log.info("Placing shelf on the floor...")
    shelf.set_orientation([0, 0, 0, 1.0])
    env.step(th.empty(0))
    offset = shelf.get_position_orientation()[0][2] - shelf.aabb_center[2]
    shelf.set_position_orientation(position=th.tensor([-1.0, 0, shelf.aabb_extent[2] / 2]) + offset)
    env.step(th.empty(0))  # One step is needed for the object to be fully initialized

    og.log.info("Shelf placed.")
    for _ in range(50):
        env.step(th.empty(0))

    og.log.info("Placing boxes...")
    for i, box in enumerate(boxes):
        box.states[object_states.Inside].set_value(shelf, True)
        og.log.info(f"Box {i} placed.")

        for _ in range(50):
            env.step(th.empty(0))


if __name__ == "__main__":
    main()

Attachment Demo

This demo is useful for...

  • Understanding how to leverage the Attached state
  • Understanding how to enable objects to be attachable

python -m omnigibson.examples.object_states.attachment_demo
This demo loads an assembled shelf, and showcases how it can be manipulated to attach and detach parts.

attachment_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
import torch as th
import yaml

import omnigibson as og
from omnigibson.macros import gm

# Make sure object states are enabled
gm.ENABLE_OBJECT_STATES = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Demo of attachment of different parts of a shelf
    """
    cfg = yaml.load(open(f"{og.example_config_path}/default_cfg.yaml", "r"), Loader=yaml.FullLoader)
    # Add objects that we want to create
    obj_cfgs = []
    obj_cfgs.append(
        dict(
            type="LightObject",
            name="light",
            light_type="Sphere",
            radius=0.01,
            intensity=5000,
            position=[0, 0, 1.0],
        )
    )

    base_z = 0.2
    delta_z = 0.01

    idx = 0
    obj_cfgs.append(
        dict(
            type="DatasetObject",
            name="shelf_back_panel",
            category="shelf_back_panel",
            model="gjsnrt",
            position=[0, 0, 0.01],
            abilities={"attachable": {}},
        )
    )
    idx += 1

    obj_cfgs.append(
        dict(
            type="DatasetObject",
            name="shelf_side_left",
            category="shelf_side",
            model="bxfkjj",
            position=[-0.4, 0, base_z + delta_z * idx],
            abilities={"attachable": {}},
        )
    )
    idx += 1

    obj_cfgs.append(
        dict(
            type="DatasetObject",
            name="shelf_side_right",
            category="shelf_side",
            model="yujrmw",
            position=[0.4, 0, base_z + delta_z * idx],
            abilities={"attachable": {}},
        )
    )
    idx += 1

    ys = [-0.93, -0.61, -0.29, 0.03, 0.35, 0.68]
    for i in range(6):
        obj_cfgs.append(
            dict(
                type="DatasetObject",
                name=f"shelf_shelf_{i}",
                category="shelf_shelf",
                model="ymtnqa",
                position=[0, ys[i], base_z + delta_z * idx],
                abilities={"attachable": {}},
            )
        )
        idx += 1

    obj_cfgs.append(
        dict(
            type="DatasetObject",
            name="shelf_top_0",
            category="shelf_top",
            model="pfiole",
            position=[0, 1.0, base_z + delta_z * idx],
            abilities={"attachable": {}},
        )
    )
    idx += 1

    obj_cfgs.append(
        dict(
            type="DatasetObject",
            name="shelf_baseboard",
            category="shelf_baseboard",
            model="hlhneo",
            position=[0, -10.97884506, base_z + delta_z * idx],
            abilities={"attachable": {}},
        )
    )
    idx += 1

    cfg["objects"] = obj_cfgs

    env = og.Environment(configs=cfg)

    # Set viewer camera pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([-1.689292, -2.11718198, 0.93332228]),
        orientation=th.tensor([0.57687967, -0.22995655, -0.29022759, 0.72807814]),
    )

    for _ in range(10):
        env.step([])

    shelf_baseboard = env.scene.object_registry("name", "shelf_baseboard")
    shelf_baseboard.set_position_orientation(position=[0, -0.979, 0.21], orientation=[0, 0, 0, 1])
    shelf_baseboard.keep_still()
    # Lower the mass of the baseboard - otherwise, the gravity will create enough torque to break the joint
    shelf_baseboard.root_link.mass = 0.1

    input(
        "\n\nShelf parts fall to their correct poses and get automatically attached to the back panel.\n"
        "You can try to drag (Shift + Left-CLICK + Drag) parts of the shelf to break it apart (you may need to zoom out and drag with a larger force).\n"
        "Press [ENTER] to continue.\n"
    )

    for _ in range(5000):
        og.sim.step()

    og.shutdown()


if __name__ == "__main__":
    main()

Object Texture Demo

This demo is useful for...

  • Understanding how different object states can result in texture changes
  • Understanding how to enable objects with texture-changing states
  • Understanding how to dynamically modify object states
python -m omnigibson.examples.object_states.object_state_texture_demo

This demo loads in a single object, and then dynamically modifies its state so that its texture changes with each modification.

object_state_texture_demo.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
import torch as th

import omnigibson as og
from omnigibson import object_states
from omnigibson.macros import gm, macros
from omnigibson.utils.constants import ParticleModifyMethod

# Make sure object states are enabled, we're using GPU dynamics, and HQ rendering is enabled
gm.ENABLE_OBJECT_STATES = True
gm.USE_GPU_DYNAMICS = True
gm.ENABLE_HQ_RENDERING = True


def main(random_selection=False, headless=False, short_exec=False):
    # Create the scene config to load -- empty scene plus a cabinet
    cfg = {
        "env": {
            "rendering_frequency": 60,  # for HQ rendering
        },
        "scene": {
            "type": "Scene",
            "floor_plane_visible": True,
        },
        "objects": [
            {
                "type": "DatasetObject",
                "name": "cabinet",
                "category": "bottom_cabinet",
                "model": "zuwvdo",
                "bounding_box": [1.595, 0.537, 1.14],
                "abilities": {
                    "freezable": {},
                    "cookable": {},
                    "burnable": {},
                    "saturable": {},
                    "particleRemover": {
                        "method": ParticleModifyMethod.ADJACENCY,
                        "conditions": {
                            # For a specific particle system, this specifies what conditions are required in order for the
                            # particle applier / remover to apply / remover particles associated with that system
                            # The list should contain functions with signature condition() --> bool,
                            # where True means the condition is satisfied
                            # In this case, we only allow our cabinet to absorb water, with no conditions needed.
                            # This is needed for the Saturated ("saturable") state so that we can modify the texture
                            # according to the water.
                            # NOTE: This will only change color if gm.ENABLE_HQ_RENDERING and gm.USE_GPU_DYNAMICS is
                            # enabled!
                            "water": [],
                        },
                    },
                },
                "position": [0, 0, 0.59],
            },
        ],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Set camera to appropriate viewing pose
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([1.7789, -1.68822, 1.13551]),
        orientation=th.tensor([0.57065614, 0.20331904, 0.267029, 0.74947212]),
    )

    # Grab reference to object of interest
    obj = env.scene.object_registry("name", "cabinet")

    # Make sure all the appropriate states are in the object
    assert object_states.Frozen in obj.states
    assert object_states.Cooked in obj.states
    assert object_states.Burnt in obj.states
    assert object_states.Saturated in obj.states

    def report_states():
        # Make sure states are propagated before printing
        for i in range(5):
            env.step(th.empty(0))

        print("=" * 20)
        print("temperature:", obj.states[object_states.Temperature].get_value())
        print("obj is frozen:", obj.states[object_states.Frozen].get_value())
        print("obj is cooked:", obj.states[object_states.Cooked].get_value())
        print("obj is burnt:", obj.states[object_states.Burnt].get_value())
        print("obj is soaked:", obj.states[object_states.Saturated].get_value(env.scene.get_system("water")))

    # Report default states
    print("==== Initial state ====")
    report_states()

    # Notify user that we're about to freeze the object, and then freeze the object
    if not short_exec:
        input("\nObject will be frozen. Press ENTER to continue.")
    obj.states[object_states.Temperature].set_value(-50)
    report_states()

    # Notify user that we're about to cook the object, and then cook the object
    if not short_exec:
        input("\nObject will be cooked. Press ENTER to continue.")
    obj.states[object_states.Temperature].set_value(100)
    report_states()

    # Notify user that we're about to burn the object, and then burn the object
    if not short_exec:
        input("\nObject will be burned. Press ENTER to continue.")
    obj.states[object_states.Temperature].set_value(250)
    report_states()

    # Notify user that we're about to reset the object to its default state, and then reset state
    if not short_exec:
        input("\nObject will be reset to default state. Press ENTER to continue.")
    obj.states[object_states.Temperature].set_value(macros.object_states.temperature.DEFAULT_TEMPERATURE)
    obj.states[object_states.MaxTemperature].set_value(macros.object_states.temperature.DEFAULT_TEMPERATURE)
    report_states()

    # Notify user that we're about to soak the object, and then soak the object
    if not short_exec:
        input("\nObject will be saturated with water. Press ENTER to continue.")
    obj.states[object_states.Saturated].set_value(env.scene.get_system("water"), True)
    report_states()

    # Notify user that we're about to unsoak the object, and then unsoak the object
    if not short_exec:
        input("\nObject will be unsaturated with water. Press ENTER to continue.")
    obj.states[object_states.Saturated].set_value(env.scene.get_system("water"), False)
    report_states()

    # Close environment at the end
    if not short_exec:
        input("Demo completed. Press ENTER to shutdown environment.")
    og.clear()


if __name__ == "__main__":
    main()

Robots

These examples showcase how to interact and leverage robot objects in OmniGibson.

Robot Visualizer Demo

This demo is useful for...

  • Understanding how to load a robot into OmniGibson after an environment is created
  • Accessing all OmniGibson robot models
  • Viewing robots' low-level joint motion
python -m omnigibson.examples.robots.all_robots_visualizer

This demo iterates over all robots in OmniGibson, loading each one into an empty scene and randomly moving its joints for a brief amount of time.

all_robots_visualizer.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
import torch as th

import omnigibson as og
from omnigibson.robots import REGISTERED_ROBOTS


def main(random_selection=False, headless=False, short_exec=False):
    """
    Robot demo
    Loads all robots in an empty scene, generate random actions
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)
    # Create empty scene with no robots in it initially
    cfg = {
        "scene": {
            "type": "Scene",
        }
    }

    env = og.Environment(configs=cfg)

    og.sim.stop()

    # Iterate over all robots and demo their motion
    for robot_name, robot_cls in REGISTERED_ROBOTS.items():
        # Create and import robot
        robot = robot_cls(
            name=robot_name,
            obs_modalities=[],  # We're just moving robots around so don't load any observation modalities
        )
        env.scene.add_object(robot)

        # At least one step is always needed while sim is playing for any imported object to be fully initialized
        og.sim.play()
        og.sim.step()

        # Reset robot and make sure it's not moving
        robot.reset()
        robot.keep_still()

        # Log information
        og.log.info(f"Loaded {robot_name}")
        og.log.info(f"Moving {robot_name}")

        if not headless:
            # Set viewer in front facing robot
            og.sim.viewer_camera.set_position_orientation(
                position=th.tensor([2.69918369, -3.63686664, 4.57894564]),
                orientation=th.tensor([0.39592411, 0.1348514, 0.29286304, 0.85982]),
            )

        og.sim.enable_viewer_camera_teleoperation()

        # Hold still briefly so viewer can see robot
        for _ in range(100):
            og.sim.step()

        # Then apply random actions for a bit
        if robot_name not in ["BehaviorRobot"]:
            for _ in range(30):
                action_lo, action_hi = -0.1, 0.1
                action = th.rand(robot.action_dim) * (action_hi - action_lo) + action_lo
                if robot_name == "Tiago":
                    tiago_lo, tiago_hi = -0.1, 0.1
                    action[robot.base_action_idx] = (
                        th.rand(len(robot.base_action_idx)) * (tiago_hi - tiago_lo) + tiago_lo
                    )
                for _ in range(10):
                    env.step(action)

        # Stop the simulator and remove the robot
        og.sim.stop()
        env.scene.remove_object(obj=robot)

    # Always shut down the environment cleanly at the end
    og.clear()


if __name__ == "__main__":
    main()

Robot Control Demo

This demo is useful for...

  • Understanding how different controllers can be used to control robots
  • Understanding how to teleoperate a robot through external commands
python -m omnigibson.examples.robots.robot_control_example

This demo lets you choose a robot and the set of controllers to control the robot, and then lets you teleoperate the robot using your keyboard.

robot_control_example.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
"""
Example script demo'ing robot control.

Options for random actions, as well as selection of robot action space
"""

import torch as th

import omnigibson as og
import omnigibson.lazy as lazy
from omnigibson.macros import gm
from omnigibson.robots import REGISTERED_ROBOTS
from omnigibson.utils.ui_utils import KeyboardRobotController, choose_from_options

CONTROL_MODES = dict(
    random="Use autonomous random actions (default)",
    teleop="Use keyboard control",
)

SCENES = dict(
    Rs_int="Realistic interactive home environment (default)",
    empty="Empty environment with no objects",
)

# Don't use GPU dynamics and use flatcache for performance boost
gm.USE_GPU_DYNAMICS = False
gm.ENABLE_FLATCACHE = True


def choose_controllers(robot, random_selection=False):
    """
    For a given robot, iterates over all components of the robot, and returns the requested controller type for each
    component.

    :param robot: BaseRobot, robot class from which to infer relevant valid controller options
    :param random_selection: bool, if the selection is random (for automatic demo execution). Default False

    :return dict: Mapping from individual robot component (e.g.: base, arm, etc.) to selected controller names
    """
    # Create new dict to store responses from user
    controller_choices = dict()

    # Grab the default controller config so we have the registry of all possible controller options
    default_config = robot._default_controller_config

    # Iterate over all components in robot
    controller_names = robot.controller_order
    for controller_name in controller_names:
        controller_options = default_config[controller_name]
        # Select controller
        options = list(sorted(controller_options.keys()))
        choice = choose_from_options(
            options=options,
            name=f"{controller_name} controller",
            random_selection=random_selection,
        )

        # Add to user responses
        controller_choices[controller_name] = choice

    return controller_choices


def main(random_selection=False, headless=False, short_exec=False, quickstart=False):
    """
    Robot control demo with selection
    Queries the user to select a robot, the controllers, a scene and a type of input (random actions or teleop)
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Choose scene to load
    scene_model = "Rs_int"
    if not quickstart:
        scene_model = choose_from_options(options=SCENES, name="scene", random_selection=random_selection)

    # Choose robot to create
    robot_name = "Fetch"
    if not quickstart:
        robot_name = choose_from_options(
            options=list(sorted(REGISTERED_ROBOTS.keys())), name="robot", random_selection=random_selection
        )

    scene_cfg = dict()
    if scene_model == "empty":
        scene_cfg["type"] = "Scene"
    else:
        scene_cfg["type"] = "InteractiveTraversableScene"
        scene_cfg["scene_model"] = scene_model

    # Add the robot we want to load
    robot0_cfg = dict()
    robot0_cfg["type"] = robot_name
    robot0_cfg["obs_modalities"] = ["rgb"]
    robot0_cfg["action_type"] = "continuous"
    robot0_cfg["action_normalize"] = True

    # Compile config
    cfg = dict(scene=scene_cfg, robots=[robot0_cfg])

    # Create the environment
    env = og.Environment(configs=cfg)

    # Choose robot controller to use
    robot = env.robots[0]
    controller_choices = {
        "base": "DifferentialDriveController",
        "arm_0": "InverseKinematicsController",
        "gripper_0": "MultiFingerGripperController",
        "camera": "JointController",
    }
    if not quickstart:
        controller_choices = choose_controllers(robot=robot, random_selection=random_selection)

    # Choose control mode
    if random_selection:
        control_mode = "random"
    elif quickstart:
        control_mode = "teleop"
    else:
        control_mode = choose_from_options(options=CONTROL_MODES, name="control mode")

    # Update the control mode of the robot
    controller_config = {component: {"name": name} for component, name in controller_choices.items()}
    robot.reload_controllers(controller_config=controller_config)

    # Because the controllers have been updated, we need to update the initial state so the correct controller state
    # is preserved
    env.scene.update_initial_file()

    # Update the simulator's viewer camera's pose so it points towards the robot
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([1.46949, -3.97358, 2.21529]),
        orientation=th.tensor([0.56829048, 0.09569975, 0.13571846, 0.80589577]),
    )

    # Reset environment and robot
    env.reset()
    robot.reset()

    # Create teleop controller
    action_generator = KeyboardRobotController(robot=robot)

    # Register custom binding to reset the environment
    action_generator.register_custom_keymapping(
        key=lazy.carb.input.KeyboardInput.R,
        description="Reset the robot",
        callback_fn=lambda: env.reset(),
    )

    # Print out relevant keyboard info if using keyboard teleop
    if control_mode == "teleop":
        action_generator.print_keyboard_teleop_info()

    # Other helpful user info
    print("Running demo.")
    print("Press ESC to quit")

    # Loop control until user quits
    max_steps = -1 if not short_exec else 100
    step = 0

    while step != max_steps:
        action = (
            action_generator.get_random_action() if control_mode == "random" else action_generator.get_teleop_action()
        )
        env.step(action=action)
        step += 1

    # Always shut down the environment cleanly at the end
    og.clear()


if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="Teleoperate a robot in a BEHAVIOR scene.")

    parser.add_argument(
        "--quickstart",
        action="store_true",
        help="Whether the example should be loaded with default settings for a quick start.",
    )
    args = parser.parse_args()
    main(quickstart=args.quickstart)

Robot Grasping Demo

This demo is useful for...

  • Understanding the difference between physical and sticky grasping
  • Understanding how to teleoperate a robot through external commands
python -m omnigibson.examples.robots.grasping_mode_example

This demo lets you choose a grasping mode and then loads a Fetch robot and a cube on a table. You can then teleoperate the robot to grasp the cube, observing the difference is grasping behavior based on the grasping mode chosen. Here, physical means natural friction is required to hold objects, while sticky means that objects are constrained to the robot's gripper once contact is made.

grasping_mode_example.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
"""
Example script demo'ing robot manipulation control with grasping.
"""

import torch as th

import omnigibson as og
from omnigibson.macros import gm
from omnigibson.sensors import VisionSensor
from omnigibson.utils.ui_utils import KeyboardRobotController, choose_from_options

GRASPING_MODES = dict(
    sticky="Sticky Mitten - Objects are magnetized when they touch the fingers and a CLOSE command is given",
    assisted="Assisted Grasping - Objects are magnetized when they touch the fingers, are within the hand, and a CLOSE command is given",
    physical="Physical Grasping - No additional grasping assistance applied",
)

# Don't use GPU dynamics and Use flatcache for performance boost
gm.USE_GPU_DYNAMICS = False
gm.ENABLE_FLATCACHE = True


def main(random_selection=False, headless=False, short_exec=False):
    """
    Robot grasping mode demo with selection
    Queries the user to select a type of grasping mode
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Choose type of grasping
    grasping_mode = choose_from_options(options=GRASPING_MODES, name="grasping mode", random_selection=random_selection)

    # Create environment configuration to use
    scene_cfg = dict(type="Scene")
    robot0_cfg = dict(
        type="Fetch",
        obs_modalities=["rgb"],  # we're just doing a grasping demo so we don't need all observation modalities
        action_type="continuous",
        action_normalize=True,
        grasping_mode=grasping_mode,
    )

    # Define objects to load
    table_cfg = dict(
        type="DatasetObject",
        name="table",
        category="breakfast_table",
        model="lcsizg",
        bounding_box=[0.5, 0.5, 0.8],
        fixed_base=True,
        position=[0.7, -0.1, 0.6],
        orientation=[0, 0, 0.707, 0.707],
    )

    chair_cfg = dict(
        type="DatasetObject",
        name="chair",
        category="straight_chair",
        model="amgwaw",
        bounding_box=None,
        fixed_base=False,
        position=[0.45, 0.65, 0.425],
        orientation=[0, 0, -0.9990215, -0.0442276],
    )

    box_cfg = dict(
        type="PrimitiveObject",
        name="box",
        primitive_type="Cube",
        rgba=[1.0, 0, 0, 1.0],
        size=0.05,
        position=[0.53, -0.1, 0.97],
    )

    # Compile config
    cfg = dict(scene=scene_cfg, robots=[robot0_cfg], objects=[table_cfg, chair_cfg, box_cfg])

    # Create the environment
    env = og.Environment(configs=cfg)

    # Reset the robot
    robot = env.robots[0]
    robot.set_position_orientation(position=[0, 0, 0])
    robot.reset()
    robot.keep_still()

    # Make the robot's camera(s) high-res
    for sensor in robot.sensors.values():
        if isinstance(sensor, VisionSensor):
            sensor.image_height = 720
            sensor.image_width = 720

    # Update the simulator's viewer camera's pose so it points towards the robot
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([-2.39951, 2.26469, 2.66227]),
        orientation=th.tensor([-0.23898481, 0.48475231, 0.75464013, -0.37204802]),
    )

    # Create teleop controller
    action_generator = KeyboardRobotController(robot=robot)

    # Print out relevant keyboard info if using keyboard teleop
    action_generator.print_keyboard_teleop_info()

    # Other helpful user info
    print("Running demo with grasping mode {}.".format(grasping_mode))
    print("Press ESC to quit")

    # Loop control until user quits
    max_steps = -1 if not short_exec else 100
    step = 0
    while step != max_steps:
        action = action_generator.get_random_action() if random_selection else action_generator.get_teleop_action()
        for _ in range(10):
            env.step(action)
            step += 1

    # Always shut down the environment cleanly at the end
    og.clear()


if __name__ == "__main__":
    main()

Advanced: IK Demo

This demo is useful for...

  • Understanding how to construct your own IK functionality using omniverse's native lula library without explicitly utilizing all of OmniGibson's class abstractions
  • Understanding how to manipulate the simulator at a lower-level than the main Environment entry point
python -m omnigibson.examples.robots.advanced.ik_example

This demo loads in Fetch robot and a IK solver to control the robot, and then lets you teleoperate the robot using your keyboard.

ik_example.py 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
import argparse
import time

import torch as th

import omnigibson as og
import omnigibson.lazy as lazy
from omnigibson.objects import PrimitiveObject
from omnigibson.utils.control_utils import IKSolver


def main(random_selection=False, headless=False, short_exec=False):
    """
    Minimal example of usage of inverse kinematics solver

    This example showcases how to construct your own IK functionality using omniverse's native lula library
    without explicitly utilizing all of OmniGibson's class abstractions, and also showcases how to manipulate
    the simulator at a lower-level than the main Environment entry point.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    # Assuming that if random_selection=True, headless=True, short_exec=True, we are calling it from tests and we
    # do not want to parse args (it would fail because the calling function is pytest "testfile.py")
    if not (random_selection and headless and short_exec):
        parser = argparse.ArgumentParser()
        parser.add_argument(
            "--programmatic",
            "-p",
            dest="programmatic_pos",
            action="store_true",
            help="if the IK solvers should be used with the GUI or programmatically",
        )
        args = parser.parse_args()
        programmatic_pos = args.programmatic_pos
    else:
        programmatic_pos = True

    # Import scene and robot (Fetch)
    scene_cfg = {"type": "Scene"}
    # Create Fetch robot
    # Note that since we only care about IK functionality, we fix the base (this also makes the robot more stable)
    # (any object can also have its fixed_base attribute set to True!)
    # Note that since we're going to be setting joint position targets, we also need to make sure the robot's arm joints
    # (which includes the trunk) are being controlled using joint positions
    robot_cfg = {
        "type": "Fetch",
        "fixed_base": True,
        "controller_config": {
            "arm_0": {
                "name": "NullJointController",
                "motor_type": "position",
            }
        },
    }
    cfg = dict(scene=scene_cfg, robots=[robot_cfg])
    env = og.Environment(configs=cfg)

    # Update the viewer camera's pose so that it points towards the robot
    og.sim.viewer_camera.set_position_orientation(
        position=th.tensor([4.32248, -5.74338, 6.85436]),
        orientation=th.tensor([0.39592, 0.13485, 0.29286, 0.85982]),
    )

    robot = env.robots[0]

    # Set robot base at the origin
    robot.set_position_orientation(position=th.tensor([0, 0, 0]), orientation=th.tensor([0, 0, 0, 1]))
    # At least one simulation step while the simulator is playing must occur for the robot (or in general, any object)
    # to be fully initialized after it is imported into the simulator
    og.sim.play()
    og.sim.step()
    # Make sure none of the joints are moving
    robot.keep_still()
    # Since this demo aims to showcase how users can directly control the robot with IK,
    # we will need to disable the built-in controllers in OmniGibson
    robot.control_enabled = False

    # Create the IK solver -- note that we are controlling both the trunk and the arm since both are part of the
    # controllable kinematic chain for the end-effector!
    control_idx = th.cat([robot.trunk_control_idx, robot.arm_control_idx[robot.default_arm]])
    ik_solver = IKSolver(
        robot_description_path=robot.robot_arm_descriptor_yamls[robot.default_arm],
        robot_urdf_path=robot.urdf_path,
        reset_joint_pos=robot.get_joint_positions()[control_idx],
        eef_name=robot.eef_link_names[robot.default_arm],
    )

    # Define a helper function for executing specific end-effector commands using the ik solver
    def execute_ik(pos, quat=None, max_iter=100):
        og.log.info("Querying joint configuration to current marker position")
        # Grab the joint positions in order to reach the desired pose target
        joint_pos = ik_solver.solve(
            target_pos=pos,
            target_quat=quat,
            tolerance_pos=0.002,
            tolerance_quat=0.01,
            weight_pos=20.0,
            weight_quat=0.05,
            max_iterations=max_iter,
            initial_joint_pos=robot.get_joint_positions()[control_idx],
        )
        if joint_pos is not None:
            og.log.info("Solution found. Setting new arm configuration.")
            robot.set_joint_positions(joint_pos, indices=control_idx, drive=True)
        else:
            og.log.info("EE position not reachable.")
        og.sim.step()

    if programmatic_pos or headless:
        # Sanity check IK using pre-defined hardcoded positions
        query_positions = [[1, 0, 0.8], [1, 1, 1], [0.5, 0.5, 0], [0.5, 0.5, 0.5]]
        for query_pos in query_positions:
            execute_ik(query_pos)
            time.sleep(2)
    else:
        # Create a visual marker to be moved by the user, representing desired end-effector position
        marker = PrimitiveObject(
            relative_prim_path="/marker",
            name="marker",
            primitive_type="Sphere",
            radius=0.03,
            visual_only=True,
            rgba=[1.0, 0, 0, 1.0],
        )
        env.scene.add_object(marker)

        # Get initial EE position and set marker to that location
        command = robot.get_eef_position()
        marker.set_position_orientation(position=command)
        og.sim.step()

        # Setup callbacks for grabbing keyboard inputs from omni
        exit_now = False

        def keyboard_event_handler(event, *args, **kwargs):
            nonlocal command, exit_now
            # Check if we've received a key press or repeat
            if (
                event.type == lazy.carb.input.KeyboardEventType.KEY_PRESS
                or event.type == lazy.carb.input.KeyboardEventType.KEY_REPEAT
            ):
                if event.input == lazy.carb.input.KeyboardInput.ENTER:
                    # Execute the command
                    execute_ik(pos=command)
                elif event.input == lazy.carb.input.KeyboardInput.ESCAPE:
                    # Quit
                    og.log.info("Quit.")
                    exit_now = True
                else:
                    # We see if we received a valid delta command, and if so, we update our command and visualized
                    # marker position
                    delta_cmd = input_to_xyz_delta_command(inp=event.input)
                    if delta_cmd is not None:
                        command = command + delta_cmd
                        marker.set_position_orientation(position=command)
                        og.sim.step()

            # Callback must return True if valid
            return True

        # Hook up the callback function with omni's user interface
        appwindow = lazy.omni.appwindow.get_default_app_window()
        input_interface = lazy.carb.input.acquire_input_interface()
        keyboard = appwindow.get_keyboard()
        input_interface.subscribe_to_keyboard_events(keyboard, keyboard_event_handler)

        # Print out helpful information to the user
        print_message()

        # Loop until the user requests an exit
        while not exit_now:
            og.sim.step()

    # Always shut the simulation down cleanly at the end
    og.clear()


def input_to_xyz_delta_command(inp, delta=0.01):
    mapping = {
        lazy.carb.input.KeyboardInput.W: th.tensor([delta, 0, 0]),
        lazy.carb.input.KeyboardInput.S: th.tensor([-delta, 0, 0]),
        lazy.carb.input.KeyboardInput.DOWN: th.tensor([0, 0, -delta]),
        lazy.carb.input.KeyboardInput.UP: th.tensor([0, 0, delta]),
        lazy.carb.input.KeyboardInput.A: th.tensor([0, delta, 0]),
        lazy.carb.input.KeyboardInput.D: th.tensor([0, -delta, 0]),
    }

    return mapping.get(inp)


def print_message():
    print("*" * 80)
    print("Move the marker to a desired position to query IK and press ENTER")
    print("W/S: move marker further away or closer to the robot")
    print("A/D: move marker to the left or the right of the robot")
    print("UP/DOWN: move marker up and down")
    print("ESC: quit")


if __name__ == "__main__":
    main()

Simulator

These examples showcase useful functionality from OmniGibson's monolithic Simulator object.

What's the difference between Environment and Simulator?

The Simulator class is a lower-level object that:

  • handles importing scenes and objects into the actual simulation
  • directly interfaces with the underlying physics engine

The Environment class thinly wraps the Simulator's core functionality, by:

  • providing convenience functions for automatically importing a predefined scene, object(s), and robot(s) (via the cfg argument), as well as a task
  • providing a OpenAI Gym interface for stepping through the simulation

While most of the core functionality in Environment (as well as more fine-grained physics control) can be replicated via direct calls to Simulator (og.sim), it requires deeper understanding of OmniGibson's infrastructure and is not recommended for new users.

State Saving and Loading Demo

This demo is useful for...

  • Understanding how to interact with objects using the mouse
  • Understanding how to save the active simulator state to a file
  • Understanding how to restore the simulator state from a given file
python -m omnigibson.examples.simulator.sim_save_load_example

This demo loads a stripped-down scene with the Turtlebot robot, and lets you interact with objects to modify the scene. The state is then saved, written to a .json file, and then restored in the simulation.

sim_save_load_example.py 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
import os

import torch as th

import omnigibson as og
import omnigibson.lazy as lazy
from omnigibson.utils.ui_utils import KeyboardEventHandler

TEST_OUT_PATH = ""  # Define output directory here.


def main(random_selection=False, headless=False, short_exec=False):
    """
    Prompts the user to select whether they are saving or loading an environment, and interactively
    shows how an environment can be saved or restored.
    """
    og.log.info(f"Demo {__file__}\n    " + "*" * 80 + "\n    Description:\n" + main.__doc__ + "*" * 80)

    cfg = {
        "scene": {
            "type": "InteractiveTraversableScene",
            "scene_model": "Rs_int",
            "load_object_categories": ["floors", "walls", "bed", "bottom_cabinet", "chair"],
        },
        "robots": [
            {
                "type": "Turtlebot",
                "obs_modalities": ["rgb", "depth"],
            },
        ],
    }

    # Create the environment
    env = og.Environment(configs=cfg)

    # Set the camera to a good angle
    def set_camera_pose():
        og.sim.viewer_camera.set_position_orientation(
            position=th.tensor([-0.229375, -3.40576, 7.26143]),
            orientation=th.tensor([0.27619733, -0.00230233, -0.00801152, 0.9610648]),
        )

    set_camera_pose()

    # Give user instructions, and then loop until completed
    completed = short_exec
    if not short_exec and not random_selection:
        # Notify user to manipulate environment until ready, then press Z to exit
        print()
        print("Modify the scene by SHIFT + left clicking objects and dragging them. Once finished, press Z.")

        # Register callback so user knows to press space once they're done manipulating the scene
        def complete_loop():
            nonlocal completed
            completed = True

        KeyboardEventHandler.add_keyboard_callback(lazy.carb.input.KeyboardInput.Z, complete_loop)
    while not completed:
        action_lo, action_hi = -1, 1
        env.step(th.rand(env.robots[0].action_dim) * (action_hi - action_lo) + action_lo)

    print("Completed scene modification, saving scene...")
    save_path = os.path.join(TEST_OUT_PATH, "saved_stage.json")
    og.sim.save(json_paths=[save_path])

    print("Re-loading scene...")
    og.clear()
    og.sim.restore(scene_files=[save_path])

    # env is no longer valid after og.clear()
    del env

    # Take a sim step and play
    og.sim.step()
    og.sim.play()
    set_camera_pose()

    # Loop until user terminates
    completed = short_exec
    if not short_exec and not random_selection:
        # Notify user to manipulate environment until ready, then press Z to exit
        print()
        print("View reloaded scene. Once finished, press Z.")
        # Register callback so user knows to press space once they're done manipulating the scene
        KeyboardEventHandler.add_keyboard_callback(lazy.carb.input.KeyboardInput.Z, complete_loop)
    while not completed:
        og.sim.step()


if __name__ == "__main__":
    main()

Rendering

These examples showcase how to change renderer settings in OmniGibson.

Renderer Settings Demo

This demo is useful for...

  • Understanding how to use RendererSettings class
python -m omnigibson.examples.renderer_settings.renderer_settings_example

This demo iterates over different renderer settings of and shows how they can be programmatically set with OmniGibson interface.

renderer_settings_example.py 
1