Researchers Develop New Framework to Teach Three Robots Dynamic Movements

Researchers from the Institute for Robotics and Artificial Intelligence and Boston Dynamics have recently developed an imitation learning framework called ZEST (Zero-shot Embodied Skill Transfer). This framework leverages reinforcement learning to train control policies from diverse sources—including high-fidelity motion capture, noisy monocular videos, and non-physically constrained animations—and enables zero-shot deployment of these policies onto real robotic hardware. The findings have been published in the journal *Science Robotics* and successfully demonstrated on two humanoid robots and one quadruped robot, enabling them to learn new dynamic motions.