Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors
Tactile-JEPA pre-trains robot touch encoders by using the sensor layout itself as structure.
The method masks parts of a distributed tactile skin and predicts their embeddings from the remaining sensors, guided by the connectivity graph of the sensing elements. The authors argue that these skins are sparse and irregular, making vision-style self-supervised methods a poor fit. Across three datasets, they report lower force estimation error by 6.3% and lower in-hand orientation error by 20.8% versus prior state of the art, with gains also in policy learning. ArXiv · AI/CL/LG's note
The method masks parts of a distributed tactile skin and predicts their embeddings from the remaining sensors, guided by the connectivity graph of the sensing elements. The authors argue that these skins are sparse and irregular, making vision-style self-supervised methods a poor fit. Across three datasets, they report lower force estimation error by 6.3% and lower in-hand orientation error by 20.8% versus prior state of the art, with gains also in policy learning. ArXiv · AI/CL/LG's note
score 5