pretrained-vision-language-action-policy-fine-tuned-sequentially-stream-real-robot-manipulation-tasks CLM-0470mechanismsingle paperReviewed by AI
When a pretrained vision-language-action policy is fine-tuned sequentially on a stream of real-robot manipulation tasks, the choice of action-normalization statistics is itself a determinant of continual-learning success: statistics that are task-derived, first-task-derived, or recomputed each stage produce forgetting or poor transfer, while statistics estimated once from a small task-independent calibration set covering the embodiment's motion range and then frozen give the best acquisition, retention and task-order stability.
Capability: Turning plans into physical action
Observed on
Replay-based continual fine-tuning of a pretrained VLA (pi-0.5) on four real-robot streams of five tasks each, single-arm and bimanual, with percentile-based normalization; 10 roll.
Sources
- supportsFAN: Foresight Action Normalization for Continual Adaptation of Vision-Language-Action ModelsValidated by: Average final score across four real-robot streams: 95.3 for FAN vs 77.0 for the best baseline (first-task statistics), 62.0, 59.0 and 25.9 for the others; forward transfer positive on all four streams (avg 9.1 vs -12.7 to -75.6); mean absolute matched-task difference across stream orders 4.2 points vs 17.5-48.5 for baselines. Ablation: narrowing or widening (2.5x) the frozen range cut average score by roughly 24 points on two streams.. Measured on one bimanual platform with one base policy, four streams of five tasks, 10 rollouts per cell; [truncated]
Either way, bring evidence: a link, or what you saw.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For post-training a pretrained vision-language-action manipulation policy under a fixed per-round frame budget, collecting handheld demonstrations only at states where the human action chunk diverges from the policy's sampled action distribution, and weighting them by a learned task-progress advantage, raises task progress more than collecting the same number of frames of ordinary demonstrations and fine-tuning on them.Turning plans into physical action · reviewed by AI
- When adapting a pretrained VLA manipulation policy with only a handful of target demonstrations, retrieving prior demonstrations at the level of LLM-decomposed subtasks (then reranking by behavior-feature similarity) raises success more than retrieval by visual, optical-flow, or state-action similarity, and the margin is concentrated on long-horizon composite tasks rather than single-skill pick-and-place.Turning plans into physical action · reviewed by AI
- For manipulation policies pretrained on mixed egocentric-human and robot data, adding egocentric data helps only when the embodiment gap is bridged explicitly: a three-stage curriculum (ego priors, then task-matched human/robot alignment in a shared end-effector frame, then robot-only) raised real-robot success, while separate ego/robot action heads gave only a marginal in-distribution gain and shared future-video prediction in a world-action model made success worse than a robot-only baseline.Turning plans into physical action · reviewed by AI
- For vision-language-action and world-action manipulation policies evaluated on LIBERO-style tasks, near-ceiling one-shot task success does not predict the ability to resume a task from a failure state that arose during the model's own execution: all six tested policies lost over half their success rate on recovery scenarios, rankings inverted between the two settings, and recovery success fell steeply as the failure altered object or environment state rather than just the last action.Turning plans into physical action · reviewed by AI
- For real-world online RL post-training of large pretrained vision-language-action policies on high-precision manipulation, optimizing the action head toward a paired advantage margin against a frozen reference action, rather than maximizing absolute Q values, yields higher autonomous success on long multi-step tasks, because absolute-Q maximization amplifies early optimistic value errors into policy drift that behavior-cloning regularization alone does not stop.Turning plans into physical action · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A continual VLA adaptation study on comparable real-robot streams where frozen calibration-set statistics score no better than task-specific or dynamically recomputed statistics, or where the coverage width of the frozen range has no effect on later-task acquisition.