Summary

TACO is a tactile-aware world-model framework for scalable VLA post-training in contact-rich manipulation. It uses a Recognize–Imagine–Label loop: real-world failure-adjacent states are recognized, a visuo-tactile world model imagines local correction segments (joint RGB+force denoising), and corrective actions are labeled for VLA post-training — converting failures into self-supervised supervisory signal.

Key Contributions

  • Visuo-tactile generation model: jointly denoises video frames and force trajectories for local correction segment synthesis
  • Unified progress-action model: estimates task progress and predicts corrective actions from generated segments
  • Recognize–Imagine–Label loop: closes the failure-to-correction pipeline with no human annotation
  • Scalable post-training: generates corrective supervision from real-world failures without large-scale tactile pretraining datasets

Significance

Bridges the gap between rich tactile sensing and VLA post-training by turning contact failures into training data automatically — a scalable route to contact-rich manipulation that does not require purpose-built tactile pretraining corpora.