Beyond Policy Patches: Distilling RL Specialists into Vision-Language-Action Policies
Beyond Policy Patches: Distilling RL Specialists into Vision-Language-Action Policies
This research addresses failures in vision-language-action (VLA) policies during short, precision-critical phases by introducing REPAIR, a value-guided distillation framework.
The paper introduces REPAIR, a value-guided distillation framework that transfers capabilities from lightweight reinforcement learning specialists into a single vision-language-action policy by combining offline initialization with on-policy refinement using weighted flow-matching objectives.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
number of trials per task
30
Reported trials
View original evidence
ndidates and weights them using its critic to construct a supervision distribution. A shared weighted flow-matching objective trains the VLA first on observations from specialist-assisted rollouts, then on observations collected through its own execution. Specialists provide supervision without taking control during on-policy collection and are removed at deployment. Across six real-robot tasks with 30 evaluation tri
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ndidates and weights them using its critic to construct a supervision distribution. A shared weighted flow-matching objective trains the VLA first on observations from specialist-assisted rollouts, then on observations collected through its own execution. Specialists provide supervision without taking control during on-policy collection and are removed at deployment. Across six real-robot tasks with 30 evaluation tri
Open source S4
arXiv:2611.02777v1 Announce Type: new Abstract: Vision-language-action (VLA) policies can handle noncritical stages of long-horizon manipulation yet fail during short precision-critical phases. Lightweight reinforcement learning (RL) specialists can address these failures without directly applying RL to the full VLA, but retain auxiliary controllers and switching mechanisms at deployment. We study how to transfer the
Open source S2
als per task, REPAIR achieves 100% success on all six tasks, with lower mean end-effector jerk than specialist-assisted execution on every task.
Open source S5
ir capabilities into a single flow-matching VLA. Direct imitation of specialist rollouts can reproduce action jitter and leave the VLA unprepared for states encountered during its own execution. We introduce REPAIR, a value-guided distillation framework that combines offline initialization with on-policy refinement. At each critical-phase observation, REPAIR queries the corresponding specialist for multiple action ca
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org