Pre-training tactile encoders on sensor layout cuts force error by 6.3%

single source· 1 articles · confidence: high · first seen 2026-09-20 20:00 UTC

What this means for you

Nothing to migrate to — this is a training method with a code release, not a product. If you build contact-rich manipulation, the masking scheme is the transferable idea, but the 6.3% and 20.8% gains come from three datasets in the authors' setup. Reproduce them on your own sensors before switching.

A paper posted to arXiv on 20 September describes Tactile-JEPA, a pre-training method for the encoders that read distributed tactile sensors — the sparse, irregularly placed elements in robot electronic skins. It hides some sensing elements and trains the encoder to predict their embeddings from the rest, using the sensor connectivity graph to guide which are hidden, at two scales. Across three datasets covering magnetic and piezoresistive sensors, the authors report 6.3% lower force estimation error and 20.8% lower in-hand orientation error than the prior state of the art. Code is released; no evaluation date is given.

Key facts

  • ·The encoder is pre-trained by masking some sensing elements and predicting their embeddings from the unmasked remainder, with the sensor connectivity graph guiding which are hidden and masking applied at two scales. source
  • ·Reported gains over the prior state of the art are 6.3% lower force estimation error and 20.8% lower in-hand orientation error. source
  • ·Evaluation covers three datasets spanning magnetic and piezoresistive sensors, different robot embodiments, and single- and paired-sensor configurations. source
  • ·Code is available at github.com/E-Kovtun/tactile. source
  • ·The paper was posted to arXiv on 20 September 2026 and gives no evaluation date for the comparisons. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire