BeyondMimic Motion Tracking Code

[Website] [Arxiv] [Video]

Overview

BeyondMimic is a versatile humanoid control framework that provides highly dynamic motion tracking with the state-of-the-art motion quality on real-world deployment and steerable test-time control with guided diffusion-based controllers.

This repo covers the motion tracking training in BeyondMimic. You should be able to train any sim-to-real-ready motion in the LAFAN1 dataset, without tuning any parameters.

For sim-to-sim and sim-to-real deployment, please refer to the motion_tracking_controller.

Installation

• Install Isaac Lab v2.1.0 by following the installation guide. We recommend using the conda installation as it simplifies calling Python scripts from the terminal.

• Clone this repository separately from the Isaac Lab installation (i.e., outside the IsaacLab directory):

# Option 1: SSH
git clone git@github.com:xiaohu-art/whole_body_tracking.git

# Option 2: HTTPS
git clone https://github.com/xiaohu-art/whole_body_tracking.git

• Pull the robot description files from GCS

# Enter the repository
cd whole_body_tracking
# Rename all occurrences of whole_body_tracking (in files/directories) to your_fancy_extension_name
curl -L -o unitree_description.tar.gz https://storage.googleapis.com/qiayuanl_robot_descriptions/unitree_description.tar.gz && \
tar -xzf unitree_description.tar.gz -C source/whole_body_tracking/whole_body_tracking/assets/ && \
rm unitree_description.tar.gz

• Using a Python interpreter that has Isaac Lab installed, install the library

python -m pip install -e source/whole_body_tracking

Motion Tracking

Motion Preprocessing

Note: The reference motion should be retargeted and use generalized coordinates only.

• Gather the reference motion datasets (please follow the original licenses), we use the same convention as .csv of Unitree's dataset

  • Unitree-retargeted LAFAN1 Dataset is available on HuggingFace

    hf download lvhaidong/LAFAN1_Retargeting_Dataset --repo-type dataset --local-dir {local_path}

  • Sidekicks are from KungfuBot
  • Christiano Ronaldo celebration is from ASAP.
  • Balance motions are from HuB

• Convert retargeted motions to include the maximum coordinates information (body pose, body velocity, and body acceleration) via forward kinematics,

# example
python scripts/csv_to_npz.py --input_file LAFAN1/g1/dance1_subject1.csv --input_fps 30 --output_dir LAFAN1/g1/output --output_name dance1_subject1 --headless

python scripts/replay_npz.py --motion_file LAFAN1/g1/output/dance1_subject1.npz

• Convert PHC retargeted .pkl motions to .npz

python scripts/pkl_to_npz.py --input_file {phc retargeted motion}.pkl --input_fps 30 --output_dir phc-retarget/output --output_name {output_name} --headless

python scripts/replay_npz.py --motion_file phc-retarget/output/{output_name}.npz

• Process multiple motions from a .pkl file and save collected simulation data back to the original .pkl file

python scripts/pkl_to_pkl.py --input_file motions.pkl --input_fps 30 --output_fps 50 --headless

# Replay specific motion keys from the processed pkl file
python scripts/replay_pkl.py --input_file motions.pkl --motion_key dance1_subject1
python scripts/replay_pkl.py --input_file motions.pkl --motion_key dance1_subject1 --loop
python scripts/replay_pkl.py --input_file motions.pkl --list_motions

Policy Training

• Train policy by the following command:

python scripts/rsl_rl/train.py --task=Tracking-Flat-G1-v0 \
--motion_file LAFAN1/g1/output/dance1_subject1.npz \
--headless --logger wandb --log_project_name {project_name} --run_name {run_name}

Policy Evaluation

• Play the trained policy by the following command:

python scripts/rsl_rl/play.py --task=Tracking-Flat-G1-v0 --num_envs=2 --motion_file LAFAN1/g1/output/dance1_subject1.npz

Code Structure

Below is an overview of the code structure for this repository:

• source/whole_body_tracking/whole_body_tracking/tasks/tracking/mdp This directory contains the atomic functions to define the MDP for BeyondMimic. Below is a breakdown of the functions:

  • commands.py Command library to compute relevant variables from the reference motion, current robot state, and error computations. This includes pose and velocity error calculation, initial state randomization, and adaptive sampling.

  • rewards.py Implements the DeepMimic reward functions and smoothing terms.

  • events.py Implements domain randomization terms.

  • observations.py Implements observation terms for motion tracking and data collection.

  • terminations.py Implements early terminations and timeouts.

• source/whole_body_tracking/whole_body_tracking/tasks/tracking/tracking_env_cfg.py Contains the environment (MDP) hyperparameters configuration for the tracking task.

• source/whole_body_tracking/whole_body_tracking/tasks/tracking/config/g1/agents/rsl_rl_ppo_cfg.py Contains the PPO hyperparameters for the tracking task.

• source/whole_body_tracking/whole_body_tracking/robots Contains robot-specific settings, including armature parameters, joint stiffness/damping calculation, and action scale calculation.

• scripts Includes utility scripts for preprocessing motion data, training policies, and evaluating trained policies.

This structure is designed to ensure modularity and ease of navigation for developers expanding the project.