Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors
The paper reports lower force and orientation errors from pretraining tactile encoders on sensor layout, not just raw signals.
Tactile-JEPA masks parts of a distributed tactile sensor graph and trains the model to predict the missing embeddings from the remaining sensors. The authors say the method uses dual-scale masking to capture both local contact detail and the global state of the tactile surface. Across three datasets, it cuts force estimation error by 6.3% and in-hand orientation error by 20.8% versus the prior state of the art. HF Daily Papers' note
Tactile-JEPA masks parts of a distributed tactile sensor graph and trains the model to predict the missing embeddings from the remaining sensors. The authors say the method uses dual-scale masking to capture both local contact detail and the global state of the tactile surface. Across three datasets, it cuts force estimation error by 6.3% and in-hand orientation error by 20.8% versus the prior state of the art. HF Daily Papers' note
score 5