Simulation — Isaac Lab & MuJoCo

1. Isaac Lab Simulation

Overview

Welcome to the OpenArm Isaac Lab Simulation Documentation. OpenArm is officially integrated into NVIDIA's Isaac Sim / Isaac Lab ecosystem, enabling development and evaluation of reinforcement learning, imitation learning, and foundation model-based approaches.

The OpenArm model is released under an open-source license (Apache 2.0) and the team aims to contribute practical simulation assets, training workflows, and benchmarks that can be directly applied to real-world robotic systems.

Version Information

Available RL Demo Environments

Four reinforcement learning environments are publicly available in the openarm_isaac_lab repository:

Detailed implementation code, configuration files, and usage instructions can be found in the openarm_isaac_lab GitHub repository.

Isaac Lab Setup

Follow the official Isaac Lab installation guide at docs.isaacsim.omniverse.nvidia.com, then clone the OpenArm extension:

# After installing Isaac Lab, clone OpenArm extension
cd /path/to/isaaclab
git clone https://github.com/enactic/openarm_isaac_lab.git \
  source/extensions/openarm_isaac_lab

# Install the extension
./isaaclab.sh -i source/extensions/openarm_isaac_lab

# Run a demo environment (example: reaching task)
./isaaclab.sh -p source/extensions/openarm_isaac_lab/scripts/train.py \
  --task OpenArm-Reach-v0 \
  --num_envs 4096

Coming Soon

• Teleoperation code — VR and joystick teleoperation in simulation
• Imitation Learning Code — Dataset recording and behavior cloning setup
• Sim2Real Code — Transfer pipeline from simulation to real hardware

2. MuJoCo Simulation

Overview

MuJoCo (Multi-Joint dynamics with Contact) is a physics engine maintained by Google DeepMind. MJCFs (MuJoCo Description Files) are XML files describing all moving scene components, facilitating algorithm experimentation in reproducible environments.

The OpenArm MuJoCo model is available in the openarm_mujoco repository.

Setup & Quick Start

1. Download and install MuJoCo.

   # MuJoCo 3.3.4 on x86 Linux
   MUJOCO_VERSION="3.3.4"
   wget -q --show-progress \
     "https://github.com/google-deepmind/mujoco/releases/download/${MUJOCO_VERSION}/mujoco-${MUJOCO_VERSION}-linux-x86_64.tar.gz"
   tar --extract --gzip --verbose \
     --file="mujoco-${MUJOCO_VERSION}-linux-x86_64.tar.gz"
   rm "mujoco-${MUJOCO_VERSION}-linux-x86_64.tar.gz"

2. Launch the MuJoCo simulate viewer.

   cd "mujoco-${MUJOCO_VERSION}/bin"
   ./simulate

3. Clone the OpenArm MuJoCo repository.

   git clone https://github.com/enactic/openarm_mujoco.git

4. Load the OpenArm model. In the MuJoCo simulate window, open the file explorer and drag v1/openarm_bimanual.xml into the window. The bimanual OpenArm model will load in the physics simulation.

Technical Details

Actuator Model

The MuJoCo model uses torque control for actuators, enabling realistic simulation while requiring client-side control management. This matches the real hardware's motor control mode.

Geometry Groups

Physics simulation uses the convex hull of collision geometries, providing an efficient approximation for contact computation.

MJCF Composition

The MJCF files are composable — you can load individual arm components or the full bimanual setup:

• v1/openarm_bimanual.xml — Full bimanual setup (both arms)
• Individual arm components can be loaded independently for single-arm tasks

Python API Usage

# Basic MuJoCo Python usage with OpenArm
import mujoco
import mujoco.viewer
import numpy as np

# Load the model
model = mujoco.MjModel.from_xml_path("v1/openarm_bimanual.xml")
data = mujoco.MjData(model)

# Apply torque control to joints
with mujoco.viewer.launch_passive(model, data) as viewer:
    while viewer.is_running():
        # Set torque commands (torque control mode)
        data.ctrl[:] = np.zeros(model.nu)

        # Step simulation
        mujoco.mj_step(model, data)
        viewer.sync()

ROS2 Bridge (Coming Soon)

A ROS2 package enabling MuJoCo integration as a mock hardware interface will be released shortly after the v1 MuJoCo model release. This will allow the full ROS2 control stack to be tested against MuJoCo simulation before deployment to real hardware.