README.md

metadata

license: apache-2.0
tags:
  - reinforcement-learning
  - robotics
  - legged-locomotion
  - quadruped
  - mujoco
  - mjlab
  - rsl-rl
  - ppo
  - onnx
pipeline_tag: reinforcement-learning

XGO — Luwu Dynamics Quadruped Locomotion Policies

PPO locomotion policies for the Luwu Dynamics (陆吾智能, xgorobot.com) quadruped robots Mini2S, Mini2SW (wheeled variant), and Lite3, trained with RSL-RL in mjlab (Isaac-Lab-style APIs on GPU-accelerated MuJoCo Warp) via the luwu_mjlab training environment.

Policies

Folder	Task ID	Robot	Behavior	Iterations	Final mean reward
mini2s_walk/	Mini2S-Walk-Flat	Mini2S	Flat-terrain velocity-tracking walk	1500	87.1
mini2s_turn/	Mini2S-Turn-Flat	Mini2S	Fall recovery / turn-over	1500	112.9
mini2sw_walk/	Mini2SW-Walk-Flat	Mini2SW (wheeled)	Flat-terrain velocity-tracking walk	1500	89.4
mini2sw_turn/	Mini2SW-Turn-Flat	Mini2SW (wheeled)	Fall recovery / turn-over	1500	97.6
lite3_walk/	Lite3-Walk-Flat	Lite3	Flat-terrain velocity-tracking walk	1500	69.3

Each folder contains:

• policy.onnx — exported inference-only policy (observation → action), robot- and deployment-framework agnostic.
• model_1499.pt — final RSL-RL PPO checkpoint (actor-critic weights + optimizer state), for resuming training or loading in luwu_mjlab/rsl_rl.
• params/env.yaml, params/agent.yaml — the exact environment and PPO hyperparameters used for that run.

Training setup

• Trained on a single NVIDIA GPU, 4096 parallel environments, 1500 PPO iterations per policy.
• Simulator: MuJoCo 3.6.0 / mujoco-warp 3.6.0 (GPU-accelerated).
• Framework: mjlab v1.2.0 + rsl-rl-lib 5.0.1.
• Task source: luwu_mjlab src/tasks/velocity/.

Usage

With luwu_mjlab installed, evaluate a checkpoint:

python scripts/play.py Mini2S-Walk-Flat \
  --checkpoint-file /path/to/mini2s_walk/model_1499.pt

Or run the exported ONNX policy directly with onnxruntime for deployment (input/output tensor names and shapes are defined by luwu_mjlab's observation/action manager for each task — see params/env.yaml in the matching folder).

Disclaimer

These policies were trained in simulation only and have not been validated on hardware. Evaluate carefully (and add safety measures) before any real-robot deployment.