Berkeley, California – A new framework enables a humanoid robot to smoothly switch between versatile, humanlike movements and perform agile maneuvers such as cartwheels and spin kicks.

Called BeyondMimic, the motion-tracking, reinforcement learning-based framework is trained on human motions and can scale to diverse and new skills without task-specific training.

The framework converts human motion data into continuous representations and produces movement trajectories without task-specific fine-tuning. It can also classify prompts for new or previously unseen tasks without retraining.

After training on about 2.5 hours of human motion data, researchers tested BeyondMimic in simulation before conducting experiments with a physical humanoid robot.

The robot performed cartwheels, tennis swings and soccer goal celebrations with humanlike agility. It also completed complex martial arts-inspired movements and acrobatics outdoors on uneven ground.

It could walk and run with humanlike naturalness while smoothly transitioning between movements without stiffness.

“These components enable scalable acquisition of humanlike motor skills from human motion and motion synthesis that generalizes and adapts beyond the training setup,” Qiayuan Liao and colleagues write.

The authors propose that future versions could learn directly from human demonstrations and enable versatile movements in different environments.

Podcast: Play in new window | Download