Tactile-JEPA: topology-aware SSL cuts robot force error 6.3%, orientation error 20.8%

Elizaveta Kovtun · hf · 2026-09-28

Elizaveta Kovtun et al. present Tactile-JEPA, a self-supervised pre-training method for distributed tactile electronic skins, whose sparse, irregularly arranged sensing elements make visual SSL methods a poor fit.

Method: a sensor connectivity graph guides spatial masking to predict embeddings of masked elements, learning topology-aware representations; dual-scale masking captures both local contact detail and global skin state.

Results: across three datasets spanning magnetic and piezoresistive sensors, different robot embodiments, and single/paired-sensor setups, Tactile-JEPA cuts force estimation error by 6.3% and in-hand orientation error by 20.8% versus the prior SOTA, with consistent gains in policy learning. The authors argue tactile sensing benefits depend critically on encoder pre-training quality. Code is open-sourced.

Original post →

More from Embodied

Embodied channel →