A failed rollout can contain useful progress without making its terminal actions good imitation targets. Preserve the distinction between successful behavior, useful failure prefixes and corrective recovery segments when constructing learning labels. Keeping a whole failure recording and assigning training credit to part of it are different decisions.
The RePO-VLA paper, submitted May 10, 2026, assigns separate roles to success, recovery and failure trajectories. Its recovery initialization slices corrective segments and resets their history; a progress-aware value function supports reliability-decayed failure prefixes and value-conditioned refinement. These are author research methods, not a general rule that every failure frame deserves positive training weight.
Preserve the original episode and the selected learning segment
Imagine an illustrative placement episode with a correct grasp and lift, followed by drift and a dropped object. Its early progress may remain useful while the terminal action should not be copied as a successful continuation.
A proposed training manifest should retain the parent episode, segment boundaries, segment role, original outcome and assigned label/weight provenance. Keep rejected and unused segments in the audit record where available. A filtered training view should not overwrite the evidence needed to reconstruct its selection.
Avoid counting slices from one parent as independent evaluation episodes. Shared observations and task history can leak between training and a nominally held-out set even when the exported segments have different filenames.
Define what a recovery-history reset changes
In Recovery-Aware Initialization (RAI), the source clears the observation-history buffer at the first recovery frame and lets it refill over subsequent steps. Corrective actions are thus conditioned on the current adverse state without the preceding failure context. This reset applies only to RAI: Value-Conditioned Refinement (VCR) restores the original rolling histories, and deployment also uses the normal rolling buffer. None of these history operations physically resets the robot or object to a nominal start state.
For the RAI training view, retain the adverse-state observations available at the recovery boundary and the policy context supplied after the reset. Check whether omitting preceding history also omits information needed to interpret the current task. This is a proposed diagnostic, not an allegation about the released method.
Compare equivalent recovery examples under explicitly defined context rules. Otherwise an apparent optimization gain can include a change in the information given to the actor.
Evaluate the role assignment on unseen adverse states
For a proposed ablation, keep the parent data and evaluation set fixed while changing how success, failure-prefix and recovery segments contribute. Report nominal-task performance and adverse-state continuation separately.
Inspect whether a method learns useful early progress without repeating the terminal error. At deployment, the paper fixes the conditioning value to 1.0. That input requests behavior learned under high-value labels; it is not a measured success probability or an independent observation that the current physical state is successful.
The robot data-quality audit covers episode integrity. Add segment-role and label lineage when constructing this recovery-driven training view.
No data slicing, value model or policy optimization was performed here. Use failure experience through a traceable learning objective, preserving the difference between archival evidence and actions the actor is encouraged to reproduce. Paper version 1 checked October 15, 2026.