h1v2 Isaac

Dependencies

Overview

This repository provides the essential codebase to run custom RL policies on an Unitree H1-2 robot.

Currently the RL training uses Isaac Lab. The sim2sim validation is done with Mujoco. The sim2real deployment pipe depends on unitree_sdk2_python.

The project is structured around three main packages:

• robot_assets: Manages the models of the robot
• robot_deploy: Handles deployment-related functionalities (sim2sim and sim2real).
• robot_tasks: Contains the core definition of the training environment for the robot.

Both robot_deploy and robot_assets are adapted from the template provided for Isaac Lab Projects to facilitate integration.

Usage

This project has been packaged using UV.

To run a training:

uv run scripts/rsl_rl/train.py --task Isaac-Velocity-Flat-H12_12dof-v0

This command initiates the training process for the specified task.

To deploy a trained policy onto a real robot, use

uv run --package robot_deploy scripts/deploy/main.py

Note: the --package robot_deploy flag avoids pulling the dependencies from the robot_tasks package used for training (e.g. isaacsim, isaaclab...).

To evaluate a policy in simulation, set the use_mujoco flag to True in the config.yaml file before running the main.py script: it will spawn a MuJoCo simulator instance in the background and communicate with it through DDS, as when deploying on a real robot.

It's also possible to run the trained policy in simulation directly (i.e. without using DDS) by passing the flag --sim to the main.py script, in this case the policy is synchronously run in the simulator.

Input devices

Currently, three input methods are implemented to control the robot's behavior:

• Unitree controller: when the code is not run in simulation (i.e. no flag --sim AND use_mujoco: False in the configuration file), the default input device is the Unitree Controller
• Gamepad: when running in simulation, the code will automatically attempt to detect and use any connected gamepad
• Keyboard: if no gamepad is detected, keyboard inputs are read through the MuJoCo simulator window

The default keybindings are:

• start: initialize the robot
• select: kill the robot and activate damping mode
• L1 / R1: switch between control policies specified in the configuration file
• (in MuJoCo only) B: toggle the elastic band maintaining the robot in its standing position
• (in MuJoCo only) L2 / R2: modify the length of the elastic band

The keyboard inputs are mapped to the generic controller commands as follows:

Keyboard key	Mapped key
Enter	Start
Escape	Select
A/B/X/Y	A/B/X/Y
J	L1
K	R1
I	L2
O	R2

License

This project is licensed under the BSD 2-Clause License - see the LICENSE file for details.

Contributors

• Valentin Guillet: Core developer
• Côme Perrot: Core developer
• Constant Roux: Core developer
• Victor Lutz: Robot model integration
• Alessandro Trovatello: RSL-RL training implementation
• Olivier Stasse: Project supervisor

Citation

To cite this work in a publication:

@misc{h1v2Isaac2025,
  author = {Valentin Guillet and Côme Perrot and Constant Roux and Olivier Stasse},
  title = {h1v2-Isaac: Reinforcement Learning Framework for Unitree H1-2 Robot},
  year = {2025},
  howpublished = {\url{https://github.com/Gepetto/h1v2-Isaac}},
}