Robots can rehearse forgotten tasks using AI-generated memory replays
REGEN lets a robot policy rehearse previously mastered tasks by generating synthetic visual trajectories, removing the need to store original human demonstrations.
REGEN is a continual imitation learning framework that lets robots retain past skills while adapting to new ones, without storing original human demonstrations. It queries a World Action Model to generate synthetic replay trajectories conditioned on prior task instructions and current observations.
The robot policy rehearses these AI-generated scenarios instead of real data. Tests in both simulation and real-world manipulation confirmed the approach.
Why it mattersRobots that avoid storing human demonstrations can scale to more tasks without growing data requirements, making continual learning more practical for real deployments.
Open this story in InSnip →
