generic-robot-env 0.0.2a0

A generic MuJoCo-based robot environment generator for LeRobot or Gymnasium-style experiments.

generic_robot_env

A generic MuJoCo-based robot environment generator for LeRobot or Gymnasium-style experiments.

This library is heavily inspired by and depends on the gym_hil project which implemented the first variant of these configurations.

This package provides

• RobotConfig: A version of the gym_hil MujocoGymEnv that can extract its configuration from a well-formed mujoco scene.xml file.
• GenericRobotEnv, a configurable robot-control base environment around a MuJoCo XML model, with reusable robot methods (apply_action, get_robot_state, reset_robot, render, get_gripper_pose).
• GenericTaskEnv, a task-oriented layer on top of GenericRobotEnv that adds Panda-pick-style task behavior (task reset, environment state observation, reward, success/termination).

Features

• Auto-detects joints, actuators, end-effector site, cameras and (optionally) a home keyframe from a MuJoCo XML file.
• Two control modes: osc (end-effector operational-space control) and joint (direct actuator control).
• Separation of concerns: robot-control APIs in GenericRobotEnv, task APIs in GenericTaskEnv.
• Optional image observations from model cameras.
• Returns structured observations compatible with Gymnasium spaces.Dict.

Installation

This project depends on MuJoCo and Gymnasium. Install prerequisites in your Python environment (example using pip):

pip install generic_robot_env

Note: The repository includes typed stubs under typings/ for local development and the project expects a working MuJoCo installation accessible to the mujoco Python package.

Usage examples

• Simple loop: Create a configuration from a scene XML and run a quick random-action loop. mujoco_menagerie is a good repository of scenes.

from pathlib import Path
import numpy as np
from generic_robot_env.generic_robot_env import RobotConfig, GenericTaskEnv

xml = Path("mujoco_menagerie/aloha/scene.xml")  # point to a model in the repo
config = RobotConfig.from_xml(xml, robot_name="aloha")
env = GenericTaskEnv(config, control_mode="osc", image_obs=False)

obs, _ = env.reset()
for _ in range(200):
	action = env.action_space.sample()
	obs, reward, terminated, truncated, info = env.step(action)
	if terminated:
		break
env.close()

• Image observations: Enable image_obs=True to receive pixel arrays under obs['pixels'] keyed by camera name (when cameras exist in the MJCF).

config = RobotConfig.from_xml(Path("mujoco_menagerie/aloha/aloha.xml"), robot_name="aloha")
env = GenericTaskEnv(config, control_mode="osc", image_obs=True)
obs, _ = env.reset()
# obs['pixels'] -> dict(camera_name -> HxWx3 uint8 array)

Notes:

• If the MuJoCo model contains a keyframe named home, the environment will try to use that as the default joint pose.
• If cameras are present and image_obs=True, pixel observations are returned under obs['pixels'] keyed by camera name.

Files

• src/generic_robot_env/generic_robot_env.py — main environment implementation (this module).

See src/generic_robot_env/generic_robot_env.py for the full implementation.

Observation and action spaces

GenericRobotEnv observations are returned as a Gymnasium spaces.Dict with primary keys under agent_pos:

• joint_pos: Joint positions for the detected robot joints (array)
• joint_vel: Joint velocities (array)
• tcp_pose: End-effector pose as [x, y, z, qx, qy, qz, qw]
• tcp_vel: End-effector linear and angular velocity (6D)
• gripper_pose (optional): Single-value gripper state when a gripper actuator is detected

When image_obs=True, pixels is included and contains a spaces.Dict of camera-name -> RGB image arrays.

GenericTaskEnv uses Panda-pick-style observations:

• State mode: {"agent_pos": <vector>, "environment_state": <object position>}
• Image mode: {"pixels": <camera dict>, "agent_pos": <vector>}

Action spaces depend on control_mode:

• osc: Continuous Box controlling end-effector delta in position (x,y,z) and rotation (rx,ry,rz). If a gripper is available, an extra dimension for gripper command is appended.
• joint: Continuous Box mapped directly to actuator control values. When a gripper actuator exists, it is appended to the action vector.

Implementation notes

• The environment auto-resolves joint and actuator ids using the MuJoCo model and maps qpos/qvel indices for direct data access.
• For OSC control, a simple opspace solver (from gym_hil.controllers.opspace) is used each simulation substep to compute actuator torques that track a desired end-effector target.
• Gripper mapping tries to respect actuator control ranges defined in the model (actuator_ctrlrange) when present.

Running tests

Run the test suite with pytest from the repository root:

pytest tests

Local presubmit hooks

Install local git hooks so linting/formatting runs before commit and tests run before push:

uv run pre-commit install --hook-type pre-commit --hook-type pre-push

Run hooks manually (optional):

uv run pre-commit run --all-files
uv run pre-commit run --hook-stage pre-push

• Faster experiments: Many models include camera/site names that the environment will auto-detect (end-effector sites, cameras, and optional home keyframes). If your chosen model provides a home keyframe the environment will attempt to use it as the default reset pose.

• Example with a bundled task: Some pre-made gym-style wrappers (for example in gym_hil) subclass the same base utilities; you can switch between GenericRobotEnv and those wrappers by pointing both at the same XML and configuration.

Tips and troubleshooting

• Missing end-effector/site detection: If the end-effector isn't found automatically, open the XML and add a site with a common name like ee, end_effector, tcp or attachment_site so auto-detection can find it.
• Gripper mapping: If the model exposes a gripper actuator with a control range, the environment will append a gripper command dimension to the action space and will respect actuator_ctrlrange when possible.

Contributing

Contributions, bug reports and improvements are welcome. When adding new robots or tasks, prefer adding descriptive camera and site names in the MJCF so the auto-detection heuristics can find the end-effector and camera frames.

License

This project inherits the license of the repository. Ensure you follow the licensing terms of MuJoCo and any third-party dependencies.