ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control

ABSTRACT

We introduce ZEST (Zero-shot Embodied Skill Transfer), a motion-imitation framework that trains policies via RL from diverse sources—mocap, noisy monocular video, and animation—and deploys them to hardware zero-shot. ZEST generalizes across behaviors and platforms while avoiding contact labels, reference windows, or extensive reward shaping. Its pipeline combines adaptive sampling for difficult motion segments and an automatic curriculum using a model-based assistive wrench. We execute soccer kicks and dance snippets on Atlas, achieving robust transfer from noisy video references captured with handheld phone cameras, establishing ZEST as a scalable interface between biological movements and robotic control.