Dream to Remember: Continual Steering of Generalist Robot Policies with World Models

Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience. We ask a question the continual-RL literature has assumed an answer to but never measured: which component forgets? Under never-clear replay, pre-registered component-level probes (n=3 seeds throughout) show that the world model retains essentially everything measurable about old tasks -- reward discrimination (retention ratio ~1.0), value estimates, and termination structure -- while the actor's behavior collapses. Forgetting in this regime is a channel problem, not a memory problem. We demonstrate this by intervention: with the world model frozen and identical imagined rollouts, reinforcement learning in imagination fails to recover a lost skill (0/3 seeds), while supervised self-imitation on the world model's own graded dreams recovers it on 3/3 seeds with zero environment interaction. Interleaved during training, this graded dream rehearsal yields a task-label-free, parameter-constant continual learner: 3/3 four-task chains retained where plain replay passes 0/3, 3/3 eight-task chains, and consistent gains over matched real-episode cloning (paired difference +0.13, bootstrap 95% CI [0.07, 0.24], complete seed separation). The dream-grading step is load-bearing: we characterize two scoring failure modes, provide an offline selection gauge that caught both before they contaminated results, and give a realized-first grading rule that closes them. All experiments were pre-registered with committed protocols; every refuted hypothesis is reported.

We introduce DreamFormer, a model-based agent that acquires language-conditioned, multi-task skills by imitating expert demonstrations within the latent imagination of a learned world model. DreamFormer first learns a task-agnostic Transformer world model from unstructured play data, then acquires task-specific behaviors by optimizing an intrinsic reward that aligns agent-generated rollouts with expert demonstrations in latent space. Since the policy is trained on-policy inside imagination, it is exposed to its own errors during training, mitigating the covariate shift inherent to offline behavioral cloning. To make long-horizon imagination affordable, DreamFormer encodes a high-resolution multi-view robotic observation into a single input token, avoiding both spatial downsampling and the multi-token representations used by prior Transformer world models. On the long-horizon CALVIN benchmark, DreamFormer outperforms LUMOS, the comparable model-based agent, on single-environment evaluation (2.52 vs 2.34 average tasks completed per chain of five) while keeping imagination rollouts tractable. Against HULC, the behavior cloning baseline, it nearly doubles performance on zero-shot transfer to an unseen environment (1.30 vs 0.67), indicating that dynamics learned by the world model transfer more readily than a directly cloned policy. This is consistent with the role attributed to internal models in biological agents, where a model of environment dynamics supports behavior in situations not previously encountered.

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 conditioned only on prior task instructions and current-task observations. Experiments in both simulation and real-world manipulation settings show that REGEN reduces catastrophic forgetting by up to $50\%$ relative to sequential fine-tuning, while approaching the performance of privileged experience replay methods that require access to real replay data. Finally, we analyze the factors limiting generated replay, identifying long-horizon visual degradation and action-observation inconsistency as the primary bottlenecks. Our results establish WAMs as a promising foundation for continual robot learning without stored demonstrations.