ZEST: Zero-shot embodied skill transfer for athletic robot control

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Humanoid robot control for agile, contact-rich behaviors is challenging, requiring extensive engineering and controller tuning.
• Existing methods often rely on per-skill engineering, limiting generalization and robustness.

Purpose of the Study:

• Introduce ZEST (zero-shot embodied skill transfer), a streamlined motion-imitation framework for robot control.
• Enable zero-shot deployment of learned skills to hardware across diverse behaviors and platforms without extensive tuning.

Main Methods:

• ZEST trains policies using reinforcement learning from diverse data sources: motion capture, monocular video, and animation.
• The framework incorporates adaptive sampling and an automatic curriculum with a model-based assistive wrench for dynamic maneuvers.
• A procedure for selecting joint-level gains for closed-chain actuators is also provided.

Main Results:

• ZEST demonstrated broad generality when trained in simulation with moderate domain randomization.
• On Boston Dynamics' Atlas, ZEST learned dynamic, multicontact skills (army crawl, breakdancing) from motion capture.
• Skills were transferred from video to Atlas and Unitree G1 (dance, box climbing), and from animation to Spot quadruped (backflip).

Conclusions:

• ZEST achieves robust zero-shot deployment across heterogeneous data sources and embodiments.
• The framework establishes a scalable interface between biological movements and robotic control.
• This approach simplifies the process of imparting humanlike, agile skills to robots.