Creating a Modular Environment

Creating a Modular Environment#

This tutorial demonstrates how to create sophisticated robotic environments using EmbodiChain’s modular architecture. You’ll learn how to use the advanced envs.EmbodiedEnv class with configuration-driven setup, event managers, observation managers, and randomization systems.

The Code#

The tutorial corresponds to the modular_env.py script in the scripts/tutorials/gym directory.

Code for modular_env.py
  1# ----------------------------------------------------------------------------
  2# Copyright (c) 2021-2025 DexForce Technology Co., Ltd.
  3#
  4# Licensed under the Apache License, Version 2.0 (the "License");
  5# you may not use this file except in compliance with the License.
  6# You may obtain a copy of the License at
  7#
  8#     http://www.apache.org/licenses/LICENSE-2.0
  9#
 10# Unless required by applicable law or agreed to in writing, software
 11# distributed under the License is distributed on an "AS IS" BASIS,
 12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 13# See the License for the specific language governing permissions and
 14# limitations under the License.
 15# ----------------------------------------------------------------------------
 16
 17import torch
 18
 19from typing import List, Dict, Any
 20
 21import embodichain.lab.gym.envs.managers.randomization as rand
 22import embodichain.lab.gym.envs.managers.events as events
 23import embodichain.lab.gym.envs.managers.observations as obs
 24
 25from embodichain.lab.gym.envs.managers import (
 26    EventCfg,
 27    SceneEntityCfg,
 28    ObservationCfg,
 29)
 30from embodichain.lab.gym.envs import EmbodiedEnv, EmbodiedEnvCfg
 31from embodichain.lab.gym.utils.registration import register_env
 32from embodichain.lab.sim.robots import DexforceW1Cfg
 33from embodichain.lab.sim.sensors import StereoCameraCfg, SensorCfg
 34from embodichain.lab.sim.shapes import MeshCfg
 35from embodichain.lab.sim.cfg import (
 36    LightCfg,
 37    ArticulationCfg,
 38    RobotCfg,
 39    RigidObjectCfg,
 40    RigidBodyAttributesCfg,
 41)
 42from embodichain.data import get_data_path
 43from embodichain.utils import configclass
 44
 45
 46@configclass
 47class ExampleEventCfg:
 48
 49    replace_obj: EventCfg = EventCfg(
 50        func=events.replace_assets_from_group,
 51        mode="reset",
 52        params={
 53            "entity_cfg": SceneEntityCfg(
 54                uid="fork",
 55            ),
 56            "folder_path": get_data_path("TableWare/tableware/fork/"),
 57        },
 58    )
 59
 60    randomize_fork_mass: EventCfg = EventCfg(
 61        func=rand.randomize_rigid_object_mass,
 62        mode="reset",
 63        params={
 64            "entity_cfg": SceneEntityCfg(
 65                uid="fork",
 66            ),
 67            "mass_range": (0.1, 2.0),
 68        },
 69    )
 70
 71    randomize_light: EventCfg = EventCfg(
 72        func=rand.randomize_light,
 73        mode="interval",
 74        interval_step=5,
 75        params={
 76            "entity_cfg": SceneEntityCfg(
 77                uid="point",
 78            ),
 79            "position_range": [[-0.5, -0.5, 2], [0.5, 0.5, 2]],
 80            "color_range": [[0.6, 0.6, 0.6], [1, 1, 1]],
 81            "intensity_range": [50.0, 100.0],
 82        },
 83    )
 84
 85    randomize_table_mat: EventCfg = EventCfg(
 86        func=rand.randomize_visual_material,
 87        mode="interval",
 88        interval_step=10,
 89        params={
 90            "entity_cfg": SceneEntityCfg(
 91                uid="table",
 92            ),
 93            "random_texture_prob": 0.5,
 94            "texture_path": get_data_path("CocoBackground/coco"),
 95            "base_color_range": [[0.2, 0.2, 0.2], [1.0, 1.0, 1.0]],
 96        },
 97    )
 98
 99
100@configclass
101class ObsCfg:
102
103    obj_pose: ObservationCfg = ObservationCfg(
104        func=obs.get_rigid_object_pose,
105        mode="add",
106        name="fork_pose",
107        params={"entity_cfg": SceneEntityCfg(uid="fork")},
108    )
109
110
111@configclass
112class ExampleCfg(EmbodiedEnvCfg):
113
114    # Define the robot configuration using DexforceW1Cfg
115    robot: RobotCfg = DexforceW1Cfg.from_dict(
116        {
117            "uid": "dexforce_w1",
118            "version": "v021",
119            "arm_kind": "anthropomorphic",
120            "init_pos": [0.0, 0, 0.0],
121        }
122    )
123
124    # Define the sensor configuration using StereoCameraCfg
125    sensor: List[SensorCfg] = [
126        StereoCameraCfg(
127            uid="eye_in_head",
128            width=960,
129            height=540,
130            enable_mask=True,
131            enable_depth=True,
132            left_to_right_pos=(0.06, 0, 0),
133            intrinsics=(450, 450, 480, 270),
134            intrinsics_right=(450, 450, 480, 270),
135            extrinsics=StereoCameraCfg.ExtrinsicsCfg(
136                parent="eyes",
137            ),
138        )
139    ]
140
141    light: EmbodiedEnvCfg.EnvLightCfg = EmbodiedEnvCfg.EnvLightCfg(
142        direct=[
143            LightCfg(
144                uid="point",
145                light_type="point",
146                color=(1.0, 1.0, 1.0),
147                intensity=50.0,
148                init_pos=(0, 0, 2),
149            )
150        ]
151    )
152
153    background: List[RigidObjectCfg] = [
154        RigidObjectCfg(
155            uid="table",
156            shape=MeshCfg(
157                fpath=get_data_path("CircleTableSimple/circle_table_simple.ply"),
158                compute_uv=True,
159            ),
160            attrs=RigidBodyAttributesCfg(
161                mass=10.0,
162                static_friction=0.95,
163                dynamic_friction=0.85,
164                restitution=0.01,
165            ),
166            body_type="kinematic",
167            init_pos=(0.80, 0, 0.8),
168            init_rot=(0, 90, 0),
169        ),
170    ]
171
172    rigid_object: List[RigidObjectCfg] = [
173        RigidObjectCfg(
174            uid="fork",
175            shape=MeshCfg(
176                fpath=get_data_path("TableWare/tableware/fork/standard_fork_scale.ply"),
177            ),
178            body_scale=(0.75, 0.75, 1.0),
179            init_pos=(0.8, 0, 1.0),
180        ),
181    ]
182
183    articulation_cfg: List[ArticulationCfg] = [
184        ArticulationCfg(
185            uid="drawer",
186            fpath="SlidingBoxDrawer/SlidingBoxDrawer.urdf",
187            init_pos=(0.5, 0.0, 0.85),
188        )
189    ]
190
191    events = ExampleEventCfg()
192
193    observations = ObsCfg()
194
195
196@register_env("ModularEnv-v1", max_episode_steps=100, override=True)
197class ModularEnv(EmbodiedEnv):
198    """
199    An example of a modular environment that inherits from EmbodiedEnv
200    and uses custom event and observation managers.
201    """
202
203    def __init__(self, cfg: EmbodiedEnvCfg, **kwargs):
204        super().__init__(cfg, **kwargs)
205
206
207if __name__ == "__main__":
208    import gymnasium as gym
209    import argparse
210
211    from embodichain.lab.sim import SimulationManagerCfg
212
213    parser = argparse.ArgumentParser()
214    parser.add_argument("--enable_rt", action="store_true", help="Enable ray tracing")
215    args = parser.parse_args()
216
217    env_cfg = ExampleCfg(sim_cfg=SimulationManagerCfg(enable_rt=args.enable_rt))
218
219    # Create the Gym environment
220    env = gym.make("ModularEnv-v1", cfg=env_cfg)
221
222    while True:
223        obs, info = env.reset()
224
225        for i in range(100):
226            action = torch.zeros(env.action_space.shape, dtype=torch.float32)
227            obs, reward, done, truncated, info = env.step(action)

The Code Explained#

This tutorial showcases EmbodiChain’s most powerful environment creation approach using the envs.EmbodiedEnv class. Unlike the basic environment tutorial, this approach uses declarative configuration classes and manager systems for maximum flexibility and reusability.

Event Configuration#

Events define automated behaviors that occur during simulation. There are three types of supported modes:

  • startup: triggers once when the environment is initialized

  • reset: triggers every time the environment is reset

  • interval: triggers at fixed step intervals during simulation

The ExampleEventCfg demonstrates three types of events:

    LightCfg,
    ArticulationCfg,
    RobotCfg,
    RigidObjectCfg,
    RigidBodyAttributesCfg,
)
from embodichain.data import get_data_path
from embodichain.utils import configclass


@configclass
class ExampleEventCfg:

    replace_obj: EventCfg = EventCfg(
        func=events.replace_assets_from_group,
        mode="reset",
        params={
            "entity_cfg": SceneEntityCfg(
                uid="fork",
            ),
            "folder_path": get_data_path("TableWare/tableware/fork/"),
        },
    )

    randomize_fork_mass: EventCfg = EventCfg(
        func=rand.randomize_rigid_object_mass,
        mode="reset",
        params={
            "entity_cfg": SceneEntityCfg(
                uid="fork",
            ),
            "mass_range": (0.1, 2.0),
        },
    )

    randomize_light: EventCfg = EventCfg(
        func=rand.randomize_light,
        mode="interval",
        interval_step=5,
        params={
            "entity_cfg": SceneEntityCfg(

Asset Replacement Event

The replace_obj event demonstrates dynamic asset swapping:

Light Randomization Event

The randomize_light event creates dynamic lighting conditions:

  • Function: envs.managers.randomization.rendering.randomize_light()

  • Mode: "interval" - triggers every 5 steps

  • Parameters: Randomizes position, color, and intensity within specified ranges

Material Randomization Event

The randomize_table_mat event varies visual appearance:

  • Function: envs.managers.randomization.rendering.randomize_visual_material()

  • Mode: "interval" - triggers every 10 steps

  • Features: Random textures from COCO dataset and base color variations

for more randomization events, please refer

Observation Configuration#

The default observation from envs.EmbodiedEnv includes: - robot: robot proprioceptive data (joint positions, velocities, efforts) - sensor: all available sensor data (images, depth, segmentation, etc.)

However, users always need to define some custom observation for specified learning tasks. To handle this, the observation manager system allows users to declaratively specify additional observations.

            "position_range": [[-0.5, -0.5, 2], [0.5, 0.5, 2]],
            "color_range": [[0.6, 0.6, 0.6], [1, 1, 1]],
            "intensity_range": [50.0, 100.0],
        },
    )

    randomize_table_mat: EventCfg = EventCfg(
        func=rand.randomize_visual_material,
        mode="interval",

This configuration:

For details documentation, see envs.managers.cfg.ObservationCfg.

Environment Configuration#

The main environment configuration inherits from envs.EmbodiedEnvCfg and defines all scene components:

Robot Configuration

    robot: RobotCfg = DexforceW1Cfg.from_dict(
        {
            "uid": "dexforce_w1",
            "version": "v021",
            "arm_kind": "anthropomorphic",
            "init_pos": [0.0, 0, 0.0],
        }
    )

Uses the pre-configured DexforceW1Cfg with customizations:

  • Version: Specific robot variant (v021)

  • Arm Type: Anthropomorphic configuration

  • Position: Initial placement in the scene

Sensor Configuration

        func=obs.get_rigid_object_pose,
        mode="add",
        name="fork_pose",
        params={"entity_cfg": SceneEntityCfg(uid="fork")},
    )


@configclass
class ExampleCfg(EmbodiedEnvCfg):

    # Define the robot configuration using DexforceW1Cfg
    robot: RobotCfg = DexforceW1Cfg.from_dict(
        {
            "uid": "dexforce_w1",
            "version": "v021",

Configures a stereo camera system using StereoCameraCfg:

  • Resolution: 960x540 pixels for realistic visual input

  • Features: Depth sensing and segmentation masks enabled

  • Stereo Setup: 6cm baseline between left and right cameras

  • Mounting: Attached to robot’s “eyes” frame

Lighting Configuration

            "init_pos": [0.0, 0, 0.0],
        }
    )

    # Define the sensor configuration using StereoCameraCfg
    sensor: List[SensorCfg] = [
        StereoCameraCfg(
            uid="eye_in_head",
            width=960,
            height=540,
            enable_mask=True,

Defines scene illumination with controllable point lights:

  • Type: Point light for realistic shadows

  • Properties: Configurable color, intensity, and position

  • UID: Named reference for event system manipulation

Rigid Objects

            left_to_right_pos=(0.06, 0, 0),
            intrinsics=(450, 450, 480, 270),
            intrinsics_right=(450, 450, 480, 270),
            extrinsics=StereoCameraCfg.ExtrinsicsCfg(
                parent="eyes",
            ),
        )
    ]

    light: EmbodiedEnvCfg.EnvLightCfg = EmbodiedEnvCfg.EnvLightCfg(
        direct=[
            LightCfg(
                uid="point",
                light_type="point",
                color=(1.0, 1.0, 1.0),
                intensity=50.0,
                init_pos=(0, 0, 2),
            )
        ]
    )

    background: List[RigidObjectCfg] = [
        RigidObjectCfg(
            uid="table",
            shape=MeshCfg(
                fpath=get_data_path("CircleTableSimple/circle_table_simple.ply"),

Multiple objects demonstrate different physics properties:

Table Configuration:

  • Shape: Custom PLY mesh with UV mapping

  • Physics: Kinematic body (movable but not affected by forces)

  • Material: Friction and restitution properties for realistic contact

Fork Configuration:

  • Shape: Detailed mesh from asset library

  • Scale: Proportionally scaled for scene consistency

  • Physics: Dynamic body affected by gravity and collisions

Articulated Objects

            ),
            attrs=RigidBodyAttributesCfg(
                mass=10.0,
                static_friction=0.95,
                dynamic_friction=0.85,
                restitution=0.01,
            ),
            body_type="kinematic",
            init_pos=(0.80, 0, 0.8),
            init_rot=(0, 90, 0),
        ),

Demonstrates complex mechanisms with moving parts:

  • URDF: Sliding drawer with joints and constraints

  • Positioning: Placed on table surface for interaction

Environment Implementation#

The actual environment class is remarkably simple due to the configuration-driven approach:

@register_env("ModularEnv-v1", max_episode_steps=100, override=True)
class ModularEnv(EmbodiedEnv):
    """
    An example of a modular environment that inherits from EmbodiedEnv
    and uses custom event and observation managers.
    """

    def __init__(self, cfg: EmbodiedEnvCfg, **kwargs):
        super().__init__(cfg, **kwargs)

The envs.EmbodiedEnv base class automatically:

  • Loads all configured scene components

  • Sets up observation and action spaces

  • Initializes event and observation managers

  • Handles environment lifecycle (reset, step, etc.)

The Code Execution#

To run the modular environment:

cd /path/to/embodichain
python scripts/tutorials/gym/modular_env.py

The script demonstrates the complete workflow:

  1. Configuration: Creates an instance of ExampleCfg

  2. Registration: Uses the registered environment ID

  3. Execution: Runs episodes with zero actions to observe automatic behaviors

Manager System Benefits#

The manager-based architecture provides several key advantages:

Event Managers

  • Modularity: Reusable event functions across environments

  • Timing Control: Flexible scheduling (reset, interval, condition-based)

  • Parameter Binding: Type-safe configuration with validation

  • Extensibility: Easy to add custom event behaviors

Observation Managers

  • Flexible Data: Any simulation data can become an observation

  • Processing Pipeline: Built-in normalization and transformation

  • Dynamic Composition: Runtime observation space modification

  • Performance: Efficient data collection and GPU acceleration

Key Features Demonstrated#

This tutorial showcases the most advanced features of EmbodiChain environments:

  1. Configuration-Driven Design: Declarative environment specification

  2. Manager Systems: Modular event and observation handling

  3. Asset Management: Dynamic loading and randomization

  4. Sensor Integration: Realistic camera systems with stereo vision

  5. Physics Simulation: Complex articulated and rigid body dynamics

  6. Visual Randomization: Automated domain randomization

  7. Extensible Architecture: Easy customization and extension points

This tutorial demonstrates the full power of EmbodiChain’s modular environment system, providing the foundation for creating sophisticated robotic learning scenarios.

Contents