Episode Details

VLS steers frozen diffusion and flow-matching robot policies using vision-language models to handle out-of-distribution changes such as novel obstacles and lighting conditions, requiring no additional training data. This advances the adaptability of robot foundation models at inference time without retraining.