Humanoid robots are getting more agile by the month, but teaching them new physical skills has traditionally required extensive task-specific training and reward engineering. A new framework called ZEST — short for Zero-shot Embodied Skill Transfer — is changing that equation.
Published in the August 2026 special issue of Science Robotics, ZEST enables humanoid robots to perform a wide range of agile movements — from continuous backflips and breakdancing to army crawling — by imitating motion-capture animations without any task-specific training or reward tuning.
The key insight is elegant: rather than painstakingly designing reward functions for each new skill (a process that often fails for highly dynamic moves), ZEST transfers skills directly from human motion-capture data to a robot's physical body in a zero-shot manner. The robot receives an animation of a desired movement and learns to replicate it on its own hardware, adapting the motion to its own physical constraints.
The framework's versatility is remarkable. In demonstrations, the same system enabled a humanoid robot to perform continuous backflips — a feat requiring precise timing, balance, and explosive power — as well as breakdancing moves and low-to-the-ground army crawls. These are movements that span vastly different dynamical regimes, from acrobatic aerial maneuvers to slow, deliberate ground-based locomotion.
The Science Robotics special issue on humanoid robots highlights this and other advances bringing humanoid control closer to real-world deployment. As the technology matures, frameworks like ZEST could dramatically reduce the cost and time required to teach robots new physical skills, accelerating the path toward robots that can adapt to diverse tasks in unstructured environments.
