Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models
SVA improves frozen Vision-Language-Action (VLA) models by adding a test-time evaluation step. It uses Monte Carlo tree search to explore potential actions in simulation, distilling this into a lightweight Q-value model. This model predicts the expected outcome of candidate actions, allowing the system to select the best action based on long-term consequences without needing simulator access during deployment.