English | 中文

_{For partnership and procurement inquiries, contact opensource@limxdynamics.com}

humanoid-mujoco-sim

Part of the Oli humanoid open-source ecosystem. Where to start and the other Oli repositories: Oli quick start.

MuJoCo simulation environment for LimX humanoid robots (HU_D03, HU_D04). Use it to visualize robot models, test motion controllers, and validate RL policies in simulation before deploying to hardware.

Scope: the simulator only. It does not contain a controller or policy; run one next to it, for example humanoid-rl-deploy-ros2.

Highlights

• Same interface as the robot. The simulator talks limxsdk, so a controller written for the robot connects to 127.0.0.1 here and to 10.192.1.2 there.
• Real ankle and waist linkages. The bundled kinematic_projection maps joint commands to the push-rod motors the way the robot does.
• Virtual joystick included. robot-joystick stands in for the robot's remote controller, keyboard-driven.
• One file to read. simulator.py is about 250 lines: load the model, apply PD commands, publish joint state and IMU at the model's 1 kHz step.

Supported robots

	Model	Description	Joints	Format
	HU_D04_01	Full-size humanoid	31	MJCF (humanoid-description)
	HU_D03_03	Previous generation, discontinued	31	MJCF (humanoid-description)

How it works

Quick start

Prerequisites

• Python 3.8 or higher
• Linux (x86_64)

The bundled prebuild/kinematic_projection helper is x86_64-only. aarch64 users can install the limxsdk wheel but simulation functionality will be limited.

Install

# Clone with submodules (HTTPS)
git clone --recurse-submodules https://github.com/limxdynamics/humanoid-mujoco-sim.git
cd humanoid-mujoco-sim

# Install dependencies
pip install -r requirements.txt

# Install the LimX motion control SDK
pip install limxsdk-lowlevel/python3/amd64/limxsdk-*.whl

Set Robot Type

Available models: HU_D03_03, HU_D04_01

echo 'export ROBOT_TYPE=HU_D04_01' >> ~/.bashrc && source ~/.bashrc

Run

python simulator.py

A MuJoCo window opens with the robot. With no controller running yet, the robot is not actuated and slumps to the floor; start a controller (see Connect a controller), then click Reset in the left panel. Use mouse (right-click + drag) to rotate the view, scroll to zoom.

Connect a controller

Start simulator.py first, then your controller with robot address 127.0.0.1. With humanoid-rl-deploy-ros2:

ros2 launch robot_hw humanoid_hw_sim.launch.py

In the virtual joystick, hold Q (L1) and press I (Y) to stand, or L (B) for the greeting motion; click Reset in MuJoCo if the robot has fallen.

Writing your own controller with limxsdk (Robot(RobotType.Humanoid)):

• Fill RobotCmd.motor_names with robot.getMotorNames(). Commands without names are rejected by kinematic_projection ("joint name: is incorrect") and the robot stays limp.
• simulator.py steps the model at its own timestep (1 ms) in real time; on a slow computer the simulation runs slower than real time.

Controls

MuJoCo window

Key	Action
Tab	Toggle left/right UI panels
Space	Pause / Resume simulation
Right-click + drag	Rotate camera
Scroll	Zoom

Virtual joystick (robot-joystick, keyboard in place of the remote):

# Linux
./robot-joystick/robot-joystick
# Windows
robot-joystick/robot-joystick.exe

Related repositories

• humanoid-description — Robot model files (URDF/MuJoCo)
• humanoid-rl-deploy-python — Python RL policy deployment
• humanoid-rl-deploy-cpp — ROS-free C++ RL deployment
• humanoid-rl-isaaclab — RL training with Isaac Lab

License

Apache License 2.0; see LICENSE and NOTICE.