A new inference-time framework improves robot task success without touching the underlying model.
The paper, posted to arXiv, introduces Guided Action Flow, a method that keeps a pretrained vision-language-action model called SmolVLA completely frozen and bolts on a separately trained critic. That critic learns from real recordings of the robot succeeding and failing, then nudges the model's sampling process toward better action sequences at runtime. The whole thing runs during inference only — no gradient updates to the base policy, no retraining loop. Tests ran on LIBERO, a standard set of robotic manipulation benchmarks.
The numbers move in the right direction. A critic trained on a single task pushed success rates from 68% to 82% on one test window and from 82% to 86% on another. A broader critic covering multiple task types lifted validation success from 46% to 56%, though gains on the held-out test set were narrower — from 65% to 67.5%. That gap between validation and test performance is the honest headline: the critic generalizes, but not as much as you'd want it to.
The appeal here is architectural. Flow-matching policies generate actions through an iterative process that naturally invites steering, much like how image diffusion models have been guided by classifiers without retraining. Applying the same logic to robot control is sensible, and the LIBERO results confirm the idea works. But the researchers are candid that critic generalization and knowing when the critic is uncertain are the two problems standing between this and practical deployment — which is a polite way of saying the hard part is still unsolved.