Imitation Learning

DeepMimic

We have implemented DeepMimic for Unitree G1 robot. To use it, you can follow below steps

1. Prepare Retargetted Data

We have validated retargetted data from GMR, you can follow its instructions to generate retargetted reference motion for g1_29dof.

Then you should paste the retargetted data to LeggedGymEx/resources/reference_motion/.

2. Process Retargetted Data

To use the retargetted data in our framework, you should process it using legged_gym/scripts/process_reference_motion.py:

Process a single motion file:

python legged_gym/scripts/process_reference_motion.py --task=g1_motion_vis --motion_file=name_of_your_refenrece_motion.pkl

# for example
python legged_gym/scripts/process_reference_motion.py --task=g1_motion_vis --motion_file=unitree_g1/02_01_walk_stageii_60hz.pkl

Process all motion files in a directory:

python legged_gym/scripts/process_reference_motion.py --task=g1_motion_vis --motion_file=path/to/motion_directory --motion_out_dir=output_directory

# for example
python legged_gym/scripts/process_reference_motion.py --task=g1_motion_vis --motion_file=unitree_g1 --motion_out_dir=processed/unitree_g1

When processing, the program will visualize the motion in the simulator. By default, we use reference motion at 60Hz and the control frequency of the policy is 50Hz.

The processed motion will be saved as a .pkl file under LeggedGymEx/resources/reference_motion/. The simulator name in its name indicate the simulator where it is generated. If you specify --motion_out_dir, the processed files will be saved to LeggedGymEx/resources/reference_motion/{motion_out_dir}/.

3. Training the policy

Then you can start training by executing python legged_gym/scripts/train.py --task=g1_motion_vis --headless --motion_file=name_of_your_processed_motion.pkl (For example: python legged_gym/scripts/train.py --task=g1_motion_vis --headless --motion_file=unitree_g1/02_01_walk_stageii_60hz_isaacgym.pkl).

Note

Because link sequences in IsaacGym/Genesis/IsaacLab are different, please make sure you use the reference motion generated from the same simulator when training.

After the training is over, you can see the result using python legged_gym/scripts/play.py --task=g1_motion_vis --motion_file=name_of_your_processed_motion.pkl --load_run=loaded_training_session. Below are some demos: