robust-NLFRL

Project Page | 中文说明

Simulation code for certifiably robust legged locomotion on Unitree G1: probabilistic neural Lyapunov functions integrated into an Actor–Critic RL pipeline (AMP + LyaPPO).

Framework overview

Fork & Scope

This project extends the open-source repository AMP_mjlab (mjlab + rsl_rl AMP stack for Unitree G1). We thank the upstream authors for releasing their codebase.

	Upstream (ccrpRepo/AMP_mjlab)	This repo (robust-NLFRL)
Focus	AMP-based G1 locomotion infrastructure	Probabilistic neural Lyapunov + RL for robust velocity tracking
Main task	Unitree-G1-AMP-Flat / Rough	**Unitree-G1-LYA-Flat**
Added module	—	TCLF co-training, probabilistic stability regularization, certified RoA

Note: The upstream repo may include broader locomotion features. This release focuses on robust velocity-tracking locomotion.

Deployment integration from the upstream project remains in ccrpRepo/wbc_fsm (MJAmp State).

Highlights

• Probabilistic Lyapunov conditions under sub-Gaussian uncertainty (dynamics + state estimation)
• Twin Control Lyapunov Function (TCLF) co-trained with policy via LyaPPO
• State-aware gating between task reward and Lyapunov robustness penalty
• Built on mjlab + vendored rsl_rl; ONNX export supported in train/play workflows

Requirements

• Linux
• Python 3.11 (recommended)
• MuJoCo-compatible environment with GPU support

pip install "warp-lang>=1.12.0,<1.13"

Quick Start

1. Install

conda activate mjlab
cd robust-NLFRL
python -m pip install -e .
pip install -e ./rsl_rl

References: mjlab, unitree_rl_mjlab.

2. Apply mjlab patch (optional)

If you skip this patch, remove history_ordering from observation configs.

cp mjlab_patch/mjlab/managers/observation_manager.py \
  $(python -c "import mjlab, os; print(os.path.dirname(mjlab.__file__))")/managers/observation_manager.py

3. List tasks

python scripts/list_envs.py --keyword LYA

Task ID	Description
Unitree-G1-LYA-Flat	Ours — AMP + probabilistic neural Lyapunov (flat terrain)
Unitree-G1-AMP-Flat	AMP baseline (flat)
Unitree-G1-Flat	PPO velocity tracking baseline (no AMP)

Training

python scripts/train.py Unitree-G1-LYA-Flat --env.scene.num-envs=4096

Logs: logs/rsl_rl/g1_lya_locomotion/<timestamp_run>/

See scripts/README_trainArgs.md for full CLI options.

Evaluation

python scripts/play.py Unitree-G1-LYA-Flat \
  --checkpoint-file logs/rsl_rl/g1_lya_locomotion/<run_dir>/model_<iter>.pt

Motion Data

AMP/LYA tasks require motion NPZ files under src/assets/motions/g1/amp/WalkandRun.

python scripts/csv_to_npz.py --help

Repository Structure

• src/tasks/amp_lya/ — Lyapunov-augmented AMP task (Unitree-G1-LYA-Flat)
• src/tasks/amp_loco/ — AMP locomotion environment (shared with upstream)
• rsl_rl/algorithms/lya_ppo.py — LyaPPO with TCLF co-training
• rsl_rl/algorithms/neural_lyapunov/ — Twin Control Lyapunov Function
• docs/ — project page assets (GitHub Pages)

Acknowledgements

• ccrpRepo/AMP_mjlab — base G1 AMP mjlab codebase this work extends
• unitreerobotics/unitree_rl_mjlab
• Open-X-Humanoid/TienKung-Lab — AMP implementation reference in rsl_rl