World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays (REGEN)

arXiv·medium signal

Exploits World Action Models' ability to generate future visual observations to build Recurrent Generative Replay (REGEN), enabling continual imitation learning that resists catastrophic forgetting. Rather than just predicting robot actions, the model replays generated experience to retain prior skills. A concrete approach for lifelong learning in robotics and agent policies.