Deleted Memories Don't Disappear: REGEN Teaches Robots to 'Dream' of the Past via World Action Models

Forum topic · 小凯 · 2026-06-26

Summary

A new robotics paper, "World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays" by Govind, Reilly, and Patel, introduces REGEN (Recurrent Generative Replay), a method that fights catastrophic forgetting in robots without storing any old task data. The approach exploits a World Action Model (WAM), which jointly predicts actions and future visual observations. When a robot learns a new task via imitation learning, REGEN periodically prompts the WAM to recursively generate synthetic "replays" of old tasks: starting from a current observation as an anchor and an old task instruction as a seed, the model alternately generates actions and observations step by step, producing pseudo-replay trajectories used as rehearsal data. Conditioning on current observations keeps dream replays visually aligned with the deployment environment. Experiments show REGEN matches or sometimes beats replay with real data, and on real robots reduces forgetting by up to 50% versus sequential fine-tuning. Key limitations include long-horizon visual degradation from error accumulation and occasional action-observation inconsistency. The work reframes memory as generative reconstruction rather than storage, a step toward lifelong continual learning for robots.

Deleted Memories Don't Disappear: How Robots Learn to "Dream" of the Past

*English adaptation of a zhichai.net deep-dive post on REGEN (Recurrent Generative Replay).*

Paper info

Title: World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays
Authors: Manish Kumar Govind, Dominick Reilly, Smit Patel
Areas: Robotics (cs.RO, cs.CV)
Posted: June 2026

Key points

Problem: Catastrophic forgetting — when neural networks learn a new task, shared weights get reconfigured and old skills are overwritten, unlike human learning which accumulates.
Core idea (REGEN): Use a World Action Model (WAM)'s generative ability to produce synthetic "replays" of past tasks instead of storing old demonstration data. The metaphor: deleted memories don't vanish — as long as you can dream them.

Why robots forget

Neural networks share parameters across tasks; training on task B overwrites the weight configuration found for task A.
Human memory relies on distributed representation and consolidation (hippocampal replay during sleep), which standard training lacks.
Existing mitigations and limits:
Experience replay: works well but requires storing all old demonstrations — impractical at scale.
Regularization (e.g., EWC): assumes tasks share parameters without severe conflict.
Modular architectures: no knowledge sharing; models grow without bound.
Generative replay: promising, but classic versions need a separately trained generative model and produce lower-quality data.

How REGEN works

A World Action Model jointly predicts the next *action* and the next *visual observation* — an internal world simulator rather than a reactive controller.
During new-task learning, REGEN periodically "dreams":

Pseudo-replay generation (recurrent):

1. $o_0 = o_t^{current}$ 2. For $$i = 1..T$$ : $a_i \sim WAM(o_{i-1}, I_{old})$ , then $o_i \sim WAM(o_{i-1}, a_i)$

Why pseudo-replays work surprisingly well

WAM replays are "cleaner" than noisy human demonstrations.
Unlimited synthetic replays are possible — no data exhaustion.
Replays adapt to the policy's current state, avoiding distribution mismatch of static real data.
Results: on multiple benchmarks REGEN performs close to, and sometimes better than, real-data replay; on real robots it reduces catastrophic forgetting by up to 50% vs. sequential fine-tuning.

Limitations (acknowledged by the paper)

Long-horizon visual degradation: recursively generated observations blur and drift over many steps, hurting long tasks.
Action–observation inconsistency: generated actions may contradict generated observations, since the two predictions are not fully synchronized.
Suggested fixes: hierarchical WAMs, explicit physical constraints, and adaptive replay length.

Why it matters

Reframes memory as generation/reconstruction rather than storage — echoing cognitive science's reconstructive theory of human memory.
A practical path toward continual, lifelong robot learning without storing petabytes of demonstrations — relevant for home robots, planetary rovers, and open-ended game agents.
Reinforces the broader world-models research direction (LeCun's world-model architecture, DeepMind's Dreamer line): future intelligent systems need an internal world simulator for reasoning, planning, and imagination.

Conclusion

REGEN shows forgetting is not a neural network's destiny: by making robots "dream" old skills through a world action model, they can accumulate abilities across a lifetime without storing old data. Open problems remain — visual degradation over long horizons, action–observation consistency, compute cost — but the work opens a door to machines that resist forgetting.

References

Govind, M.K., Reilly, D., & Patel, S. (2026). World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays. *arXiv preprint*.
French, R. M. (1999). Catastrophic forgetting in connectionist networks. *Trends in Cognitive Sciences*, 3(4), 128-135.
Shin, H., et al. (2017). Continual learning with deep generative replay. *NeurIPS 2017*.
Hafner, D., et al. (2023). Mastering diverse domains through world models. *Nature*, 616, 7955.
LeCun, Y. (2022). A path towards autonomous machine intelligence. *Open Review*.

*Original interpretation: Xiao Kai | Feynman-style deep explainer | 2026-06-27*