World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays

As robot foundation models move toward general-purpose control, they face the challenge of catastrophic forgetting—the tendency to lose previously learned skills when fine-tuned on new tasks. Traditional solutions rely on storing and replaying real human demonstrations, which is increasingly impractical due to data privacy and storage constraints. This paper asks: can the generative capabilities of World Action Models (WAMs) be leveraged to 'hallucinate' past tasks, thereby enabling continual learning without storing any original demonstration data?

The authors propose Recurrent Generative Replay (REGEN), a framework that uses a pretrained WAM as a native generative memory. When adapting to a new task, REGEN generates pseudo-trajectories for previous tasks by conditioning the WAM on the original task's language instruction and an initial observation from the current task. The model then recursively feeds its own predicted future observations back into itself to synthesize complete trajectories. These synthetic trajectories are mixed with real demonstrations from the new task during fine-tuning, allowing the model to rehearse old behaviors without access to the original datasets.

Experiments in both simulated (LIBERO) and real-world manipulation environments show that REGEN significantly mitigates catastrophic forgetting compared to naive sequential fine-tuning. Specifically, it reduces negative backward transfer (a measure of forgetting) by over 50% while maintaining strong forward transfer to new tasks. While it does not fully reach the performance of privileged methods that have access to real replay buffers, it serves as a highly effective, memory-efficient alternative for scenarios where storing past demonstrations is impossible.

This work establishes that WAMs are not just control policies but also potential memory systems for robots. By demonstrating that robots can 'imagine' their past to preserve skills, the authors provide a scalable path for lifelong robot learning that respects data privacy and storage limitations. The identified bottlenecks—specifically visual drift and action-observation inconsistency—provide a clear roadmap for future improvements in generative robot foundation models.