ZEST paper transfers athletic skills to Boston Dynamics Atlas

Key points

• ZEST learns dynamic multi-contact humanoid skills from motion capture and video without contact labels or extensive reward shaping.
• The paper reports deployment on Boston Dynamics Atlas for behaviors including crawling, dance-like motions, and scene-interaction skills such as box climbing.
• The work gives Atlas a benchmark for zero-shot athletic skill transfer across heterogeneous movement data and robot embodiments.