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

Forum topic · 小凯 · 2026-06-27

Summary

Researchers Manish Kumar Govind, Dominick Reilly, and Smit Patel propose REGEN, a continual imitation learning framework built on World Action Models (WAMs). Unlike standard policies that only predict robot actions, WAMs can also generate future visual observations. REGEN exploits this generative capability to synthesize pseudo-replay trajectories: during continual adaptation, the model recursively queries the WAM to produce synthetic replays of previously learned tasks, allowing the robot policy to rehearse old skills without storing original human demonstrations. This approach mitigates catastrophic forgetting while reducing data storage requirements, making continual learning more practical for robotic systems. The paper is available on arXiv (2606.27374) and falls within computer vision and robot learning.

Paper Overview

Research Area: Computer Vision (CV) Authors: Manish Kumar Govind, Dominick Reilly, Smit Patel Published: 2026-06-27 arXiv: 2606.27374

Abstract

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 trajectories, enabling a robot policy to rehearse previously learned tasks without storing their original human demonstrations. During continual adaptation, REGEN recursively queries the WAM to synthesize pseudo-replay trajectories, mitigating catastrophic forgetting without the need to retain raw demonstration data.

Key Points

World Action Models (WAMs) extend beyond action prediction by also generating future visual observations.
REGEN leverages this generative capability for continual imitation learning via recurrent generative replay.
The framework synthesizes pseudo-replay trajectories, letting the policy rehearse prior tasks without storing original human demonstrations.
This mitigates catastrophic forgetting while removing the storage burden of keeping raw demonstration datasets.