This is my first working pass of a standalone repo to train and deploy a basic locomotion policy for the 23dof Unitree G1.
It is adapted from the G1 29dof code distributed throughout the Mujoco Playground repo along with the deployment code from the Unitree RL repo.
Note: The policy is quite jank (i.e. the robot stumbles drunkenly around), but it's a full train-test-deploy loop and the robot actually balances and moves around IRL!
IMPORTANT: For this to work, be sure to use the old 4.3.8 version of jax as mentioned in this Mujoco Playground issue: (google-deepmind/mujoco_playground#112)
Setup/pip install the following repos (https://github.com/unitreerobotics/unitree_sdk2_python) (https://github.com/google-deepmind/mujoco_playground) (https://github.com/google/brax)
To train, run python train/main.py
To convert pkl to onnx, update the ckpt_path and run all the cells in train/brax_network_to_onnx.ipynb
To test using sim2sim with mujoco, run python deploy/deploy_sim.py with updated ONNX_PATH string
To deploy irl, setup your G1 and run python deploy/deploy_real.py with updated ONNX_PATH and NETWORK_CARD_NAME strings