josabb/G1-humanoid-6dof-hands-locomotion-rl

G1 Humanoid 6DoF Hands Locomotion (RL)

Overview

Reinforcement Learning locomotion policy for the Unitree G1 humanoid equipped with 6DoF Inspire robotic hands.

The policy was trained using RSL-RL within the MJLab framework, built on top of the MuJoCo physics engine.

<video src="https://cdn-uploads.huggingface.co/production/uploads/677802f589b3bbc27873cb18/AXme91ks4ksKSE2nGyyo0.mp4" controls width="700"></video>

Environment

• —Physics engine: MuJoCo
• —Framework: MJLab
• —RL library: RSL-RL
• —Task: Commanded velocity tracking

Framework Architecture

MJLab uses a modular manager-based RL environment:

• —Observation manager
• —Reward manager
• —Curriculum manager
• —Command manager
• —Termination manager

This design enables scalable task composition, rapid reward iteration, and clean sim-to-real transfer pipelines.

Algorithm

• —PPO
• —On-policy runner
• —Adaptive KL schedule

Observations

• —Base orientation
• —Base angular velocity
• —Joint positions
• —Joint velocities
• —Commanded velocity

Actions

• —Joint position targets

Network Architecture

Actor

• —[512, 256, 128]
• —ELU activation

Critic

• —[512, 256, 128]
• —ELU activation

Training Details

• —Iterations: 4000
• —Steps per environment: 24
• —Gamma: 0.99
• —Lambda: 0.95
• —Learning rate: 1e-3
• —Mini-batches: 4
• —Epochs per update: 5
• —Seed: 42

Files

• —checkpoints/model_4000.pt → trained policy
• —onnx/policy.onnx → deployment-ready model
• —configs/ → training configuration

Author

Josué Abad

Lab

NONHUMAN Robotics – Perú