Tactile sensing is essential for robots to understand physical interactions, but collecting dense tactile measurements is expensive and often limited in coverage. A new paper on arXiv proposes Touch-JEPA, a model that learns tactile representations by combining hand-object video with only sparse tactile supervision. The approach is designed to capture contact and tactile magnitude without relying on exhaustive sensor data.
The method leverages the rich visual cues in video—such as hand pose and object motion—to infer tactile properties, using the limited tactile signals as a guide. This reduces the burden of tactile data collection while still enabling the model to build a useful representation of contact. According to the abstract, this is a step toward physically intelligent systems that can reason about manipulation tasks.
The paper does not include experimental details in the abstract, so the reported effectiveness of Touch-JEPA is not yet available. Still, the direction is notable: it suggests that video can serve as a scalable complement to tactile sensing, potentially making tactile learning more practical for real-world robotics.