Robot learns complex 12-step tasks in 5 minutes with SymSkill

Researchers at the University of Pennsylvania developed SymSkill, a system that lets a robot learn basic skills from just five minutes of video demonstrations and recombine them into tasks of up to 12 steps. It combines symbolic planning with dynamical systems and can re-plan when a person interferes or a step fails.

• SymSkill learns basic skills from only five minutes of demonstrations
• The system handles tasks of up to 12 steps without full-sequence demos
• The robot re-plans when interrupted or when a step fails
• Tests ran in simulation and on a physical robotic arm