arushbisht12/unitree-g1-stepdown-safety

End-to-End Safety Locomotion Policy for Unitree G1 Humanoid

This model is a high-performance safety-first neural locomotion policy for the Unitree G1 Humanoid Robot, trained end-to-end with a 4-level terrain curriculum.

Key Capabilities

• Dynamic Bipedal Walking: Learns natural stepping limit cycles via feet_air_time rewards.
• Rear-Leg CoM Anchoring: Stabilizes center of mass on the rear stance leg during edge traversal.
• Adaptive Step vs. Retraction:
  • Drops $\le 12\text{ cm}$: Safely descends with compliant force loading onto the lower platform.
  • Drops $> 12\text{ cm}$: Detects unsafe drop, halts forward momentum, pulls lead leg back, and maintains upright standing posture.

Deployment with sim.py / view_stepdown.py

# 1. Interactive 3D Viewer on Step-Down
mjpython view_stepdown.py --drop 0.08 --policy safety_motion.pt

# 2. Interactive 3D Viewer on Cliff (Retract & Hold)
mjpython view_stepdown.py --drop 0.20 --policy safety_motion.pt

# 3. Standard Simulation
python sim.py --policy safety_motion.pt