Single humanoid policy beats specialized controllers across eight terrain types

Researchers from ETH Zurich with NUS, Apple and Nvidia unveiled Attention-based Contact Planning (ACP), a single learned policy that lets a humanoid climb ladders, crawl and vault obstacles. It reached a 92.5% average success rate across eight composite simulation courses, 33.7 percentage points above the leading motion-tracking baseline.

• ACP achieved 92.5% success across eight composite simulation courses
• It beat the best motion-generation baseline by 33.7 percentage points
• The planner assigns contacts to nine body parts over a 1.5-second horizon
• Two onboard networks run at 50 Hz