Researchers have introduced Tactile-JEPA, a self-supervised pre-training method designed specifically for distributed tactile sensors such as electronic skins. The work addresses a gap in robot learning: while pre-trained image encoders are now standard, tactile encoders are still often trained from scratch on raw, noisy signals, which can limit their expressiveness.
Existing self-supervised learning approaches for touch have focused mainly on vision-based tactile sensors, leaving distributed electronic skins largely unaddressed. These skins have a distinctive property: their sensing elements are sparse and irregularly arranged across the surface, making direct reuse of visual SSL methods suboptimal.
Tactile-JEPA tackles this by learning topology-aware representations from the spatial arrangement of the sensors. It is trained to predict the embeddings of masked sensing elements from the unmasked remainder, with the sensor connectivity graph guiding how masking is applied. The authors use dual-scale masking so that representations capture both local contact details and the global state of the tactile surface.
Across three datasets spanning magnetic and piezoresistive sensors, different robot embodiments, and single- and paired-sensor configurations, the method reduced force estimation error by 6.3% and in-hand orientation error by 20.8% compared with the prior state-of-the-art, with consistent gains in other downstream applications including policy learning.
The authors conclude that the benefit of tactile sensing depends critically on the quality of encoder pre-training, and present Tactile-JEPA as a direct answer to that problem. Code is available at github.com/E-Kovtun/tactile.