beyondAMP: One-Step Integration of AMP into IsaacLab & mjlab

Overview

beyondAMP provides a unified pipeline to integrate Adversarial Motion Priors (AMP) into any IsaacLab robot setup, with minimal modifications and full compatibility with custom robot designs. The same AMP machinery (discriminator, motion dataset, observation groups, on-policy runner) is also available against the mjlab (MuJoCo-Warp) backend — see mjlab Backend below. 中文README

Supported backends:

Backend	Tasks package	Train entry	Reference task
IsaacLab (PhysX)	source/amp_tasks	scripts/factoryIsaac/train.py	beyondAMP-DemoPunch-G1-BasicAMP
mjlab (MuJoCo-Warp)	source/amp_tasks_mjlab	scripts/factoryMjlab/train.py	Mjlab-AMP-Velocity-Flat-Unitree-G1

🚀 Fast Setup

cd beyondAMP
bash scripts/setup_ext.sh
# Downloads assets, robot configs, and installs dependencies

Optional VSCode workspace setup:

python scripts/setup_vscode.py

📌 How to Use

Quick Start

• Basic environment: source/amp_tasks/amp_tasks/amp
• PPO config for G1 robot: source/amp_tasks/amp_tasks/amp/robots/g1/rsl_rl_ppo_cfg.py

Training can be launched with:

python scripts/factoryIsaac/train.py --task beyondAMP-DemoPunch-G1-BasicAMP --headless
# python scripts/factoryIsaac/train.py --task beyondAMP-DemoPunch-G1-SoftAMPTrack --headless
# python scripts/factoryIsaac/train.py --task beyondAMP-DemoPunch-G1-HardAMPTrack --headless

To evaluate or visualize a trained checkpoint:

python scripts/factoryIsaac/play.py --headless --target <path to your ckpt.pt> --video --num_envs 32

AMP Punch (3k)	Motion Tracking Punck (30k)	AMP Dog Move	AMP Knee Walk

Dataset Preparation

The dataset follows the same structure and conventions used in BeyondMimic(whole_body_tracking). All motion sequences should be stored as *.npz files and placed under data/datasets/, maintaining a consistent directory layout with the reference pipeline.

For motion retargeting and preprocessing, GMR is recommended for generating high-quality retargeted mocap data. TrackerLab may be used to perform forward kinematics checks and robot-specific adjustments, ensuring the motions remain physically plausible for your robot model.

With these tools, the dataset organization naturally aligns with the conventions established in BeyondMimic(whole_body_tracking), enabling seamless integration with the AMP training pipeline.

Following the dataset pipeline of BeyondMimic:

• Motion files: place *.npz into data/datasets/
• Recommended tools:
  • GMR for retargeted motion
  • TrackerLab for FK validation & robot-specific preprocessing

AMP Integration Details

• AMP observation group added via a new amp observation config
• RSL-RL integration: source/rsl_rl/rsl_rl/env/isaaclab/amp_wrapper.py
• Default transition builder: source/beyondAMP/beyondAMP/amp_obs.py

For full tutorial and customization, see docs/tutorial.md.

🧩 mjlab Backend

beyondAMP also ships an mjlab-backed pipeline that mirrors the IsaacLab path. mjlab adopts IsaacLab's manager-based API on top of MuJoCo-Warp, so the AMP discriminator, motion dataset, and on-policy runner are reused — only the env wrapper and observation hooks differ.

Layout

Path	Purpose
source/amp_tasks_mjlab/	mjlab-side AMP task package (registers Mjlab-AMP-* tasks)
source/beyondAMP/beyondAMP/mjlab/	mjlab-specific obs groups, mdp helpers, and rsl_rl wrappers
scripts/factoryMjlab/	Training entry point for mjlab AMP tasks (see its README)

Install

scripts/setup_ext.sh currently installs the IsaacLab task package only. For the mjlab backend, additionally install:

pip install mjlab                       # MuJoCo-Warp backend
pip install -e source/amp_tasks_mjlab   # registers Mjlab-AMP-* tasks

Train

uv run python scripts/factoryMjlab/train.py \
    Mjlab-AMP-Velocity-Flat-Unitree-G1 \
    --agent.amp-data.motion-files data/demo/punch_000.npz

Tasks are looked up via mjlab.tasks.registry; the env is wrapped with beyondAMP.mjlab.rsl_rl.AMPEnvWrapper and trained with rsl_rl_amp.runners.AMPOnPolicyRunner. Registered out of the box:

• Mjlab-AMP-Velocity-Flat-Unitree-G1
• Mjlab-AMP-Velocity-Rough-Unitree-G1

To add your own mjlab AMP task, follow the pattern in source/amp_tasks_mjlab/amp_tasks_mjlab/velocity/g1/__init__.py and call register_mjlab_task(..., runner_cls=None) (the AMP wrapper replaces mjlab's stock RslRlVecEnvWrapper).

Additional Notes

• Fully modular AMP observation builder
• Compatible with IsaacLab 4.5+
• Designed for rapid experimentation across robot morphologies

🙏 Acknowledgement

Referenced Repositories

Repository	Purpose
robotlib	Robot configurations
assetslib	Asset storage
TrackerLab	Data organization & retargeting tools
AMP_for_hardware	AMP implementation reference
BeyondMimic	Dataset format & tracking comparison

---

📘 Citation

@software{zheng2025@beyondAMP,
  author = {Ziang Zheng},
  title = {beyondAMP: One step unify IsaacLab with AMP.},
  url = {https://github.com/Renforce-Dynamics/beyondAMP},
  year = {2025}
}

@software{zheng2025@trackerLab,
  author = {Ziang Zheng},
  title = {TrackerLab: One step unify IsaacLab with multi-mode whole-body control.},
  url = {https://github.com/interval-package/trackerLab},
  year = {2025}
}

@INPROCEEDINGS{Escontrela@amphardware,
  author={Escontrela, Alejandro and Peng, Xue Bin and Yu, Wenhao and Zhang, Tingnan and Iscen, Atil and Goldberg, Ken and Abbeel, Pieter},
  booktitle={2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={Adversarial Motion Priors Make Good Substitutes for Complex Reward Functions}, 
  year={2022}
  }

@article{liao2025beyondmimic,
  title={Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion},
  author={Liao, Qiayuan and Truong, Takara E and Huang, Xiaoyu and Tevet, Guy and Sreenath, Koushil and Liu, C Karen},
  journal={arXiv preprint arXiv:2508.08241},
  year={2025}
}

Community

Renforce Dynamics	Join our WeChat Group