Verifier-free Test-Time Sampling for Vision Language Action Models
Suhyeok Jang, Dongyoung Kim, Changyeon Kim, Youngsuk Kim, Jinwoo Shin
2025-10-08
Summary
This paper focuses on improving how well robots can perform tasks based on both what they see and what they're told to do, specifically when those tasks require a lot of accuracy.
What's the problem?
Current robots using 'Vision-Language-Action' models are good at following instructions and reacting to their surroundings, but they struggle with tasks needing precise movements. Existing methods to improve this precision often require extra training data or don't work well in new, unexpected situations. Basically, they aren't adaptable enough and need a lot of help to be accurate.
What's the solution?
The researchers developed a new technique called 'MG-Select' that helps robots choose the best action without needing any additional training or external tools. It works by having the robot assess how confident it is in its own potential actions. To do this, it creates a 'reference' action distribution by slightly scrambling the information it receives (like blurring the image or changing the wording of the instructions). This scrambled input forces the robot to consider a wider range of possibilities, and then it compares those possibilities to its original plan. The action that aligns best with the 'reference' distribution is chosen. They also trained the robot in a way that helps it better understand both specific instructions and general situations.
Why it matters?
This research is important because it allows robots to be more reliable and adaptable in real-world scenarios. By improving precision without needing extra training, robots can handle more complex tasks and work effectively even when things aren't exactly as expected. The significant performance gains, especially in challenging 'out-of-distribution' tasks, show this method could be a big step towards more capable and versatile robots.
Abstract
Vision-Language-Action models (VLAs) have demonstrated remarkable performance in robot control. However, they remain fundamentally limited in tasks that require high precision due to their single-inference paradigm. While test-time scaling approaches using external verifiers have shown promise, they require additional training and fail to generalize to unseen conditions. We propose Masking Distribution Guided Selection (MG-Select), a novel test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules. Our approach utilizes KL divergence from a reference action token distribution as a confidence metric for selecting the optimal action from multiple candidates. We introduce a reference distribution generated by the same VLA but with randomly masked states and language conditions as inputs, ensuring maximum uncertainty while remaining aligned with the target task distribution. Additionally, we propose a joint training strategy that enables the model to learn both conditional and unconditional distributions by applying dropout to state and language conditions, thereby further improving the quality of the reference distribution. Our experiments demonstrate that MG-Select achieves significant performance improvements, including a 28%/35% improvement in real-world in-distribution/out-of-distribution tasks, along with a 168% relative gain on RoboCasa pick-and-place tasks trained with 30 demonstrations.