Data-Free Continual Learning for World Action Model with Rectified Generative Replay

Key points

• Experience replay (ER) mitigates such forgetting by rehearsing data of previous tasks, yet this data is often unavailable in robotics, where a policy is deployed as weights only and its training data is not shared along with it.1
• In this work, we ask whether a WAM can achieve data-free continual learning by regenerating its own past.1
• To address both issues, we present rectified generative replay (R$^2$-WAM), which truncates generated rollouts at task completion with an agentic evaluator and filters out rollouts whose actions are physically inconsistent with the observations.1
• Our analysis further reveals that forgetting in WAMs concentrates in inverse dynamics modeling: a forgotten WAM still produces visually plausible observations of previous tasks while its generated actions fail.1

Sources (1)

• [1]Data-Free Continual Learning for World Action Model with Rectified Generative Replay
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 09:27 AM

  World action models (WAMs) built on pretrained video generation models adapt quickly to new robotic tasks, but they forget previously learned tasks under sequential fine-tuning.

  Experience replay (ER) mitigates such forgetting by rehearsing data of previous tasks, yet this data is often unavailable in robotics, where a policy is deployed as weights only and its training data is not shared along with it.

Extractive summary: sentences quoted from the sources.