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):

git clone <your-repo-url>

• Pull the robot description files from GCS

# Enter the repository root
cd RoboCup_Lab
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/robocup_lab/robocup_lab/assets/ && \
rm unitree_description.tar.gz

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

python -m pip install -e source/robocup_lab

Motion Tracking

Motion Preprocessing & Registry Setup

In order to manage the large set of motions we used in this work, we leverage the WandB registry to store and load reference motions automatically. 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
  • Sidekicks are from KungfuBot
  • Christiano Ronaldo celebration is from ASAP.
  • Balance motions are from HuB

• Log in to your WandB account; access Registry under Core on the left. Create a new registry collection with the name " Motions" and artifact type "All Types".

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

python scripts/csv_to_npz.py --input_file {motion_name}.csv --input_fps 30 --output_name {motion_name} --headless

This will automatically upload the processed motion file to the WandB registry with output name {motion_name}.

• Test if the WandB registry works properly by replaying the motion in Isaac Sim:

python scripts/replay_npz.py --registry_name={your-organization}-org/wandb-registry-motions/{motion_name}

• Debugging
  • Make sure to export WANDB_ENTITY to your organization name, not your personal username.
  • If /tmp folder is not accessible, modify csv_to_npz.py L319 & L326 to a temporary folder of your choice.

Policy Training

• Train policy by the following command:

python scripts/rsl_rl/train.py --task=Tracking-Flat-G1-v0 \
--registry_name {your-organization}-org/wandb-registry-motions/{motion_name} \
--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 --wandb_path={wandb-run-path}

The WandB run path can be located in the run overview. It follows the format {your_organization}/{project_name}/ along with a unique 8-character identifier. Note that run_name is different from run_path.

For the ball-dribble environments you can use the same script, e.g.

# Default dribble task (Unitree G1)
python scripts/rsl_rl/play.py --task=RobotLab-Isaac-Ball-Dribble-Unitree-G1-v0 \
  --num_envs=2 --wandb_path={wandb-run-path}

# Dribble + body velocity tracking variant
python scripts/rsl_rl/play.py --task=RobotLab-Isaac-Ball-Dribble-BodyTrack-Unitree-G1-v0 \
  --num_envs=2 --wandb_path={wandb-run-path}

# Booster T1 counterparts
python scripts/rsl_rl/play.py --task=RobotLab-Isaac-Ball-Dribble-Booster-T1-v0 --num_envs=2 --wandb_path={wandb-run-path}
python scripts/rsl_rl/play.py --task=RobotLab-Isaac-Ball-Dribble-BodyTrack-Booster-T1-v0 --num_envs=2 --wandb_path={wandb-run-path}

Ball-Carry Training (train.py only)

The locomotion-with-ball pipeline now trains in a single stage that blends the locomotion shaping rewards with the dribble-specific objectives:

python scripts/rsl_rl/train.py --task=RobotLab-Isaac-Ball-Dribble-Unitree-G1-v0 \
  --run_name=dribble --experiment_name=unitree_g1_ball_dribble --headless

This configuration ships with a dedicated reward mix for base stability, energy regularization, and the ball-directional bonuses, so no pretrain/resume step is required. 带球任务的奖励已独立配置，不再复用单独的 locomotion 任务。

If you also want the robot body to track the commanded planar velocity, switch the task ID to RobotLab-Isaac-Ball-Dribble-BodyTrack-Unitree-G1-v0. (Booster T1 使用 RobotLab-Isaac-Ball-Dribble-BodyTrack-Booster-T1-v0。)

The command accepts the usual flags (--logger, --log_project_name, --num_envs, etc.).

For Booster T1, swap the task ID with RobotLab-Isaac-Ball-Dribble-Booster-T1-v0 and adjust the experiment name (e.g. booster_t1_ball_dribble).

Code Structure

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

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

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

• source/robocup_lab/robocup_lab/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.