TACO uses tactile world models to boost VLA robot success rates by 44%

Vision-Language-Action (VLA) models have shown impressive generalization in robotic manipulation, but they often fail in contact-rich tasks where minor perturbations cause unrecoverable errors that vision alone cannot detect. To address this, a team of researchers from multiple institutions developed TACO (TActile world model as a Self-COrrector). This framework leverages a tactile-aware world model to intelligently correct failures during post-training, eliminating the need for expensive human annotation.

TACO operates via a Recognize-Imagine-Label loop. First, a unified progress-action model identifies states near failure using progress estimates. Then, a visuo-tactile generation model imagines local correction segments. Finally, the progress-action model labels these segments with executable corrective actions. The framework combines knowledge-insulated tactile adaptation with advantage-conditioned training, allowing the policy to learn from imagined corrections without degrading pretrained visual-language priors. In real-world experiments, TACO achieved a 44% absolute success rate improvement over the base policy and 32% over a version without tactile adaptation, demonstrating the power of tactile feedback for scalable VLA post-training.