RobustWalker — RL Quadruped Locomotion

Overview

RobustWalker trains a PPO-based neural network policy to control the Unitree Go1 quadruped robot using only proprioceptive sensing (no cameras or LiDAR). The robot learns to walk robustly on rough terrain and recover from external disturbances.

The Challenge

Quadruped locomotion on uneven terrain without vision sensors requires the policy to generalize across varied physical conditions — friction, payload, motor strength, and unexpected pushes.

The Solution

Implemented domain randomization during training to randomize friction, payload, motor strength, and external forces. Used a multi-objective reward function balancing forward velocity, energy efficiency, and stability. Trained with vectorized parallel environments via Stable-Baselines3.

Key Results

• Robust blind locomotion using only joint encoders and IMU
• Sim-to-real ready through extensive domain randomization
• 12-dimensional action space controlling all joint position targets
• 57-dimensional observation space with action history for temporal reasoning
• Comprehensive reward shaping for natural gait patterns