VALUE-GUIDED POLICY DISTILLATION
REPAIR.
Beyond Policy Patches:Distilling RL Specialists into
Vision-Language-Action Policies
Transferring specialist expertise into a single VLA policy.
Offline initialization. On-policy refinement. Specialist-free deployment.
REPAIR AT A GLANCE
Overview.
A narrated introduction to REPAIR: the motivation, the method, and the results.
REPAIR IN ACTION
Real-world demonstrations.
From precise insertion to coordinated manipulation. Complete task recordings play in a loop.
A CLOSER LOOK AT CRITICAL PHASES
Comparison.
SFT and RLT baselines compared with REPAIR, side by side.
SFT vs REPAIR
Precision-critical failure cases across six real-robot tasks.
Stick Pick-and-Place
Original playback speed. Phases start together; shorter clips hold their last frame until the next phase. Each pair loops together.
SFT failure modes
Successful REPAIR execution
RLT vs REPAIR
Successful dual-arm insertion, with a closer look at switching jitter.
Dual-Arm Peg Insertion
Successful rolloutsBoth supplied clips are at 5× speed. Only the latter half of the switching window (approximately 1.5–2 s) is replayed 5× slower on both sides (1× real time), with the speed labeled in-frame. Both excerpts end before gripper release; REPAIR's final frame is held while RLT continues. Playback time includes the slow-motion window.
RLT · Successful, but jittery
RLT completes the task, but exhibits jitter during the policy switch, highlighted in slow motion. Oscillations from the compact MLP specialist persist during execution and prolong task completion.
REPAIR · Faster and smoother
REPAIR completes the insertion more quickly with smoother motion. A single distilled VLA policy carries out the task without switching to a separate specialist.
THE PAPER
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 their 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 candidates 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 trials per task, REPAIR achieves 100% success on all six tasks, with lower mean end-effector jerk than specialist-assisted execution on every task. Additional visual results are available at https://repair-vla.github.io/.