Robotics · 2026-10-14

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

Kaipeng Fang, Hangxu Liu, Chaojian Shi, Yunsong Zhou, Hongzi Zhu, Minyi Guo, Jingkuan Song, Lianli Gao

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors propose R^2-WAM, a data-free continual learning method for world action models. A WAM regenerates rollouts of past tasks from their instructions, but raw rollouts cause forgetting due to missing task-completion signals and observation–action drift. R^2-WAM truncates rollouts at task completion and filters rollouts using physical-consistency flow scores, reducing forgetting by over 90% vs sequential fine-tuning and nearly matching ER without prior-task data.

Why it matters

Robotic continual learning is often “data-free” because the policy is deployed as weights only and past training data is unavailable. The authors show that a world action model can replay its own regenerated past tasks, and that truncation + physical-consistency filtering is key to preventing forgetting.

Method

Regenerate replay rollouts for previously learned tasks autoregressively from their instructions using a world action model.
Rectify generated rollouts via completion-aware adaptive truncation: a VLM evaluator detects task completion and truncates each rollout at the end step; rollouts that do not accomplish the task are discarded.
Rectify further via flow-based physical-consistency filtering: simulate executed actions to get Φ_sim, estimate optical flow from generated observations to get Φ_pred (Track4World), score agreement, and keep the top-ranked rollouts for replay.

Limitation

The authors attribute remaining gaps to task difficulty/long-horizon effects and interference among co-trained tasks (e.g., between visually similar targets), and note that without truncation the purposeless post-completion motion may interfere with the consistency score, making filtering unreliable.

Abstract (from arXiv)

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. In this work, we ask whether a WAM can achieve data-free continual learning by regenerating its own past. Our key observation is that a WAM can indeed autoregressively regenerate complete rollouts of previously learned tasks and replay them to mitigate forgetting. However, replaying such raw rollouts still lags far behind replaying real data, which we attribute to two issues. First, the WAM is unable to recognize task completion and keeps generating purposeless motion after the task is done. Second, under open-loop generation, the generated actions gradually drift away from the generated observations. 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. On the four LIBERO suites and five real-world tasks, R$^2$-WAM reduces forgetting by over 90% relative to sequential fine-tuning and nearly matches ER without accessing any data from previous tasks. 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.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-13 · Indexed by Pub-AI: 2026-10-14 UTC

preprint