Unitree G1 humanoid climbs ladders with 92.5% success in simulation

Researchers at ETH Zurich developed Attention-based Contact Planning (ACP), a single learned policy that lets the Unitree G1 humanoid climb ladders, cross obstacles and recover from falls. In simulation, the system averaged 92.5% success across eight obstacle courses versus 58.8% for the MGMT baseline.

• ACP plans contacts for nine body parts: hands, elbows, knees, feet and pelvis
• LiDAR on a Jetson Orin NX processes 20,000 points in 30 ms at 10 Hz
• Training used 13.2 minutes of human motion data in three stages
• 92.5% vs 58.8% success is simulated, not real-world performance