Vision tells a robot what something is. It doesn't tell it what's about to happen in the next fifty milliseconds of contact - whether a grip is holding, whether an object has started to slip, whether a surface is softer or rougher than it looks. That information only exists at the point of contact, and most manipulation systems still treat it as an afterthought: a binary touch/no-touch bit, if they use it at all.
Treating touch as a signal, not a switch
A few recent projects are pushing tactile sensing toward something closer to how vision or audio get treated - a continuous, information-rich stream that a policy learns from directly, rather than a threshold that trips a flag.
Visuo-Skin (ViSk), out of NYU, makes a specific and slightly contrarian bet: instead of the high-resolution optical tactile sensors most tactile-manipulation work uses (cameras looking at a deformable gel), they use low-dimensional magnetic skin sensors and feed that signal into a transformer policy as additional tokens alongside vision - no separate encoder, no dimensionality-reduction step. On four contact-rich tasks (credit card swiping, plug insertion, USB insertion, bookshelf retrieval) it outperforms both vision-only and optical-tactile baselines, with the biggest gains showing up specifically in spatial generalization. The interesting claim isn't “tactile sensing helps” - that's expected - it's that a cheap, low-dimensional signal, integrated the right way, can beat a more expensive high-resolution one.
eFlesh builds on the same magnetic sensing family from the hardware side - customizable, cut-cell microstructure touch sensing designed to be easy to fabricate and attach - and demonstrates it directly on slip detection and the same kind of visuotactile policy learning ViSk introduced, treating fingertip sensing as something you design into the hardware from the start rather than retrofit.
VibeAct takes a different physical channel entirely: instead of pressure or deformation, it reads contact and slip from vibro-acoustic signal - piezoelectric microphones embedded in a robot hand, picking up the sound and vibration of contact itself. It bridges real and simulated data by replaying the same recordings through a calibrated digital twin to automatically label per-finger contact and slip, so a tactile estimator can learn to predict contact and slip from real microphone waveforms while the manipulation policy trains in simulation on the exact same representation.
Why this is a signal, not yet a trend
We want to be straight about this one: it's a smaller, thinner body of work than the other three posts on this list. There's real depth in slip-detection research generally - the survey literature on tactile sensors for friction estimation and slip goes back over a decade - but the specific idea we're pointing at here, treating raw tactile signal as a first-class input a policy learns from directly rather than a hand-engineered feature, is still a handful of papers, not a wave. ViSk and eFlesh come out of the same lab lineage; VibeAct is a single paper exploring a genuinely different sensing channel. This is worth watching closely, not citing as settled consensus.
Where this connects for us
This is the sensing-side half of the full-duplex problem from our first post. An async control loop that's fast but blind to contact has nothing useful to react to quickly - the fast pathway needs a signal that updates as fast as it does, and vision alone doesn't provide that at the point of contact. Cheap, continuous tactile signal, read the way ViSk and VibeAct read it, is what would actually feed that loop. We think this is an early enough area that there's real room to contribute something original here, rather than just applying someone else's result.