Robot Independent Intelligence: When Language Vanishes in Vision-Language-Action Models

Abstract

In vision-language-action (VLA) models, robots typically predict actions from vision and language instructions. However, this paradigm limits their reasoning and decision-making space, reducing them to task executors rather than agents with human-like “independent intelligence.” Consequently, robots struggle to make decisions and reason in scenarios without explicit instructions. To address this limitation, we introduce semantic-level vision-language entropy to quantitatively assess task uncertainty in vision-language scenarios. As entropy increases, robots require stronger independent reasoning and decision-making capabilities; without language instructions, they must rely entirely on independent intelligence. Unlike existing VLA paradigms, our approach decouples 3D visual perception, language understanding, logical reasoning, action prediction, and execution. This reduces dependence on demonstration data and model fine-tuning while enabling explicit logical reasoning. As a result, our method allows robots to autonomously make decisions and complete reasonable tasks even without language instructions. We validate our approach through experiments across varying levels of vision-language entropy. Under identical conditions, our method significantly outperforms OpenVLA and demonstrates strong robustness.