Test-time verification for vision-language-action policiesExpert
VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models
Robot AI models of the pi0.5 type produce a slightly different motion every time they are sampled, and some of those motions fail. This preprint lets the robot draw several candidates only at tricky moments and pick the best one with a small predictive model, lifting real-robot task scores from 50.8% to 65.5% at a cost of about a third of a second per check.