arxivcs.ROcs.CV2026-06-25
World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays
Manish Kumar Govind, Dominick Reilly, Smit Patel, Hieu Le, Srijan Das
Going beyond predicting robot actions, World Action Models (WAMs) can also generate future visual observations. We build on this generative capability to propose Recurrent Generative Replay (REGEN), a continual imitation learning framework that synthesizes pseudo-replay trajector…