Human-Centric Transferable Tactile Pre-Training for Dexterous Robotic Manipulation

Robotic manipulation in contact-rich or fine-grained tasks (e.g., peeling, threading, or handling fragile objects) often fails when relying solely on visual input due to occlusion and sensory ambiguity. While Vision-Language-Action (VLA) models have shown promise in general task planning, they typically lack tactile sensing capabilities. This paper asks whether large-scale, human-centric tactile data can be used to pre-train VLA models to bridge the gap between high-level semantic reasoning and low-level physical contact control.

The authors introduce H-Tac, a massive dataset containing 160 hours of egocentric human videos, 300 tasks, and 135k episodes, all annotated with tactile and action data. To utilize this data, they propose Transferable Tactile Pre-Training (TTP). The TTP framework uses a unified tactile and action space to ensure consistency between human demonstrations and robotic embodiments. The model architecture is a dual-expert system: an action expert generates future action chunks, while a tactile expert predicts future tactile signals. This dual-prediction objective forces the model to learn the underlying contact dynamics of the environment. Furthermore, the authors implement a Tactile-Action Manifold-Preserving Gating (MPG) mechanism to ensure that the model's internal context remains reliable even when faced with distribution shifts during transfer.

TTP demonstrates significant improvements in fine-grained and contact-rich manipulation tasks. In real-world experiments, the model successfully performs complex tasks such as peeling radishes, folding paper, and inserting plugs, consistently outperforming tactile-free baselines. The authors show that the tactile-based pre-training provides critical priors that allow the model to handle fragile objects with appropriate force and maintain performance despite visual occlusions. Ablation studies confirm that both the tactile prediction expert and the MPG mechanism are essential for achieving robust performance, and that the model scales effectively with larger datasets.