This was an amazing collaboration at
@aiatmeta.bsky.social &
@cmurobotics.bsky.social : @carohiguera.bsky.social ,
@mukadammh, @francois_hogan
and many others
Code coming soon! Checkout our paper: arxiv.org/abs/2505.11420 &
Website: akashsharma02.github.io/sparsh-skin-... for more details
6/6

And we also show improvement in many tactile perception tasks such as force estimation, pose estimation and full-hand joystick state estimation.
5/6

With this we see a 75% improvement in real-world tactile plug insertion over end-to-end using vision and tactile:
4/6

We pretrain Sparsh-skin with 4 hours of unlabeled data via self-distillation, and make several changes to get highly performant reps:

Decorrelate signals by tokenizing 1s window of tactile data.
Condition the encoder on robot hand configurations via sensor positions as input
3/6

Sparsh-skin is an approach to pretrain encoders for magnetic skin sensors on a dexterous robot hand.

It improves tactile tasks by over 56% in end-to-end methods and by over 41% in prior work.
It is trained via self-supervision for the Xela sensor, so no labeled data needed!
2/6

I might sound salty, but I never got how 'outstanding reviewers' are chosen. Til then, part of the 'mediocre reviewers' gang it is 🤣

Check out my work to know more:

1. Sparsh: tactile reps for vision based sensors sparsh-ssl.github.io

2. [Releasing soon] Sparsh-skin: Tactile reps for full hand magnetic skins

3. [Coming soon] Reps for multimodal touch fusing tactile-images, audio, motion and pressure

We took a matter of fact approach for a robotics conference, and it backfired too.

Congrats Eric, really cool stuff!

At first I thought poisson or log-normal, but after a bit of searching, maybe Gumbel distribution: en.m.wikipedia.org/wiki/Gumbel_...

Now, see how life changes when you swap control and caps lock! 😆