SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models

Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic control, with test-time scaling (TTS) gaining attention to enhance robustness beyond training. However, existing TTS methods for VLAs require additional training, verifiers, and multiple forward passes, making them impractical for deployment. Moreover, they intervene only at action decoding while keeping visual representations fixed—insufficient under perceptual ambiguity, where reconsidering how to perceive is as important as deciding what to do. To address these limitations, we propose SCALE, a simple inference strategy that jointly modulates visual perception and action based on 'self-uncertainty', inspired by uncertainty-driven exploration in Active Inference theory—requiring no additional training, no verifier, and only a single forward pass. SCALE broadens exploration in both perception and action under high uncertainty, while focusing on exploitation when confident—enabling adaptive execution across varying conditions. Experiments on simulated and real-world benchmarks demonstrate that SCALE improves state-of-the-art VLAs and outperforms existing TTS methods while maintaining single-pass efficiency. Our code is publicly available at https://github.com/snumprlab/scale.

Robots powered by artificial intelligence often struggle when faced with unfamiliar or confusing situations in the real world. To fix this, researchers usually make the robot evaluate multiple action choices before moving, but this requires extra training, takes too much time, and ignores the fact that the robot might simply be looking at the wrong part of the scene. We developed a new method called SCALE that helps robots adjust their behavior based on their own internal uncertainty. By measuring this 'self-uncertainty', SCALE automatically changes how the robot perceives its surroundings and how it moves, all in a single real-time step without any extra training. If the robot is confused, it looks around more broadly and explores different possible actions; if it is confident, it focuses sharply on the immediate task. In both simulations and real-world tests, robots using SCALE completed tasks much more successfully than those using previous methods. Because SCALE improves a robot's adaptability while maintaining fast, real-time efficiency, it is a highly practical step toward making reliable robots for everyday deployment.