Unitree RL

Train and deploy locomotion policies for Unitree robots across three stages:

• Training in IsaacGym, IsaacSim

• Sim2Sim evaluation in IsaacGym, IsaacSim, MuJoCo

• Real-world deployment (networked controller)

Well Supported robots: g1_dof29 (full-body with without hands) and g1_dof12 (lower-body).

Environment setup

Install the RL library dependency (rsl_rl v3.1.1) from source:

git clone https://github.com/leggedrobotics/rsl_rl
cd rsl_rl
git checkout v3.1.1
pip install -e .

Training (IsaacGym)

General form:

python roboverse_learn/rl/unitree_rl/main.py \
  --task <your_task> \
  --sim isaacgym \
  --num_envs 8192 \
  --robot <your_robot>

Examples:

• G1 humanoid walking (IsaacSim):

python roboverse_learn/rl/unitree_rl/main.py --task walk_g1_dof29 --sim isaacsim --num_envs 8192 --robot g1_dof29

• G1Dof12 walking (IsaacGym):

python roboverse_learn/rl/unitree_rl/main.py --task walk_g1_dof12 --sim isaacgym --num_envs 8192 --robot g1_dof12

Outputs and checkpoints are saved to:

outputs/unitree_rl/<robot>_<task>/<datetime>/

Each checkpoint is named model_<iter>.pt.

Evaluation / Play

You can evaluate trained policies in both MuJoCo, Isaacsim and IsaacGym. In evaluation, main.py also exports the jit version policy to the directory outputs/unitree_rl/<robot>_<task>/<datetime>/exported/model_exported_jit.pt, which can be further used for real-world deployment.

IsaacGym evaluation:

python roboverse_learn/rl/unitree_rl/main.py \
  --task walk_g1_dof29 \
  --sim isaacgym \
  --num_envs 1 \
  --robot g1_dof29 \
  --resume <datetime_from_outputs> \
  --checkpoint <iter> \
  --eval

MuJoCo evaluation (e.g., DOF12 with public policy):

python roboverse_learn/rl/unitree_rl/main.py \
  --checkpoint <iter> \
  --task walk_g1_dof12 \
  --sim mujoco \
  --robot g1_dof12 \
  --resume <datetime_from_outputs> \
  --eval

the --resume and --checkpoint option can also be used during training for checkpoint resume.

Real-World deployment

First please install the unitree_sdk2_python package:

cd third_party
git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
cd unitree_sdk2_python
pip install -e .

Real-world deployment entry point:

python roboverse_learn/rl/unitree_rl/deploy/deploy_real.py <network_interface> <robot_yaml>

Example:

python roboverse_learn/rl/unitree_rl/deploy/deploy_real.py eno1 g1_dof29_dex3.yaml

where you should modify the corresponding yaml file in roboverse_learn/rl/unitree_rl/deploy/configs, setting the policy_path to the exported jit policy. This will initialize the real controller and stream commands to the robot. Ensure your networking and safety interlocks are correctly configured.

Command-line arguments

The most relevant flags (see helper/utils.py):

• --task (str): Task name. CamelCase or snake_case accepted. Examples: walk_g1_dof29, walk_g1_dof12.

• --robot (str): Robot identifier. Common: g1_dof29, g1_dof12.

• --num_envs (int): Number of parallel environments.

• --sim (str): Simulator. Supported: isaacgym (training), mujoco (evaluation).

• --run_name (str): Required run tag for training logs/checkpoints.

• --learning_iterations (int): Number of learning iterations (default 15000).

• --resume (flag): Resume training from a checkpoint dir (datetime) in the specified run.

• --checkpoint (int): Which checkpoint to load. -1 loads the latest.

• --headless (flag): Headless rendering (IsaacGym).

• --jit_load (flag): Load the jit policy.

Notes:

• Checkpoints: outputs/unitree_rl/<task>/<run_name or datetime>/model_<iter>.pt

• Exported JIT model (when used): outputs/unitree_rl/<task>/<run_name or datetime>/exported/model_exported_jit.pt