Pick up a grape with your fingers and you already know, before you look, whether you’ve crushed it. That feedback loop — mechanoreceptors firing, motor commands adjusting in milliseconds — is so deeply embedded in human movement that we barely notice it. Robots have been essentially numb by comparison. Not anymore.
Tactile sensing is having a serious moment, and the results are starting to look genuinely astonishing. The core problem for embodied AI has never been vision alone. It’s that contact with the world produces information that cameras simply cannot capture: the shear forces when a screwdriver bites into a slot, the micro-slip that tells you a glass is about to fall, the precise distribution of pressure that distinguishes a firm grip from a damaging one. Without that signal, manipulation is fundamentally guesswork dressed up as motor control.
The hardware side has made a real leap. GelSight-style sensors — soft elastomers backed by cameras that image deformation in real time — can now resolve surface texture at sub-millimeter scale and measure tangential force vectors simultaneously. DIGIT, developed at Meta AI Research and now widely used in research labs, packages this into a compact fingertip that produces a rich tactile image at 60 frames per second. More recent derivatives push that further, adding shear estimation and embedding the sensor directly into compliant robot fingers rather than bolting it on as an afterthought.
But hardware alone doesn’t close the loop. The interesting question is what you do with that torrent of tactile data, and this is where the embodied AI story gets exciting. The current generation of tactile-aware manipulation systems is training visuotactile policies — neural networks that jointly process RGB images and tactile sensor streams — using a combination of human demonstration, simulation, and self-supervised contact experience. The results are qualitatively different from vision-only systems. Robots trained with tactile feedback can reliably handle deformable objects, reorient parts in-hand without dropping them, and adjust grip force on the fly when an object starts to slip. These are things that vision-only systems still fail at regularly, regardless of how good their spatial reasoning has gotten.
Stanford’s robotics group and Carnegie Mellon’s RPAD lab have both published work in the past year showing that even sparse tactile signals — just a handful of pressure points — produce dramatic improvements in downstream manipulation success rates on cluttered, contact-rich tasks. The gains aren’t marginal. In some task categories, adding tactile input roughly doubles success rates compared to matched vision-only baselines. That’s the kind of number that changes what engineers think is worth building.
Humanoid robot developers are paying close attention. Fingers are no longer an afterthought. Companies building general-purpose humanoids have started treating the hand, not the arm, as the critical design challenge — because the hand is where the robot meets the actual world in all its physical complexity. Getting sensing right at the fingertip level is what separates a robot that can demo in a controlled setting from one that can work reliably in a factory or a kitchen.
There’s a deeper point here too. Touch is not just an additional sensor modality — it’s a fundamentally different kind of information about the world. It’s local, immediate, and causal in a way that vision is not. A robot that can feel is a robot that can learn from contact, build physical intuition through experience, and develop something that starts to resemble genuine manual skill rather than pre-programmed motion sequences.
We are still early. But the trajectory is clear and the pace is accelerating. Within a few years, tactile-aware manipulation will likely be a baseline expectation for capable robot hands, the way stereo vision is now. What robots will be able to do with that sense of touch — in surgery, in manufacturing, in the everyday physical world — is only starting to come into focus.