A robot can spot a wine glass on a shelf and reach for it. What it cannot tell, until something breaks, is how hard it is squeezing — or that the glass has already started to slip. For years the field treated this as a minor gap. It is turning out to be the gap.
The big bet in manipulation right now is not a smarter camera but a sense of touch. Researchers and startups are racing to give robot hands the feedback humans take for granted, and the work is finally producing numbers instead of promises.
Key takeaways
- Vision-language-action (视觉—语言—动作) models lifted robots' general skills, but they still fumble fine contact tasks like plugging in a cable.
- A Berkeley dataset called T-Rex added 100 hours of tactile data and roughly doubled success rates on tricky manipulation tasks.
- China's DAIMON Robotics builds vision-based tactile sensors with about 110,000 sensing units and proposes a VTLA framework that folds touch into the VLA recipe.
- The hard part is no longer sensing touch but reacting fast enough, and sharing data across very different hands and sensors.
Why vision alone stalls
Most modern robots lean on cameras and a class of models called vision-language-action (视觉—语言—动作), or VLA. Trained on huge piles of images, video, and text, then fine-tuned on a bit of teleoperation, these models let a robot fold laundry or tidy a room from a video feed and a plain instruction. They are genuinely useful.
But ask a robot to plug in a USB cable, turn a key, or handle a raw egg and it often fails. As UC Berkeley's Trevor Darrell puts it, most dexterous manipulation can be done by a person with their eyes closed — force, slip, and precise grip are things vision simply cannot see. Without touch, a robot either slows to a crawl, retries endlessly, or drops the object and hands the mess back to a human.
The T-Rex evidence
The bottleneck has been data. Tactile signals look nothing like the internet-scale images VLAs train on, and good tactile datasets have been rare. Berkeley's T-Rex project attacked this directly: the team first pretrained on existing data, then trained a specialist submodel on about 100 hours of high-quality tactile demonstrations using more than 200 household objects.
The clever part was speed. Touch has to correct a grip in milliseconds, far faster than a typical language-vision model thinks. So the system splits the job — a slower "action expert" plans the motion, while a faster "tactile expert" adjusts it four times more often based on what the hand feels. Fine-tuned on about 100 teleoperated demos of tasks like screwing in a bulb or moving an egg, the model hit roughly 65 percent success across 12 tasks — nearly double the best VLA baseline.
Tsinghua's Chengbo Yuan took a different tack, pooling more than 3,000 hours of tactile data across 21 sensor types into a hardware-agnostic model. The lesson: touch data is only as useful as it is shareable.
A Chinese player: DAIMON and the VTLA bet
DAIMON Robotics, a company founded by HKUST robotics veteran Wang Yu, builds vision-based tactile sensors that capture the deformation of a fingertip surface as images — over 110,000 effective sensing units per sensor. Because the output is visual, it slides naturally into the VLA framework, turning it into what the company calls VTLA: vision-tactile-language-action.
That framing matters. The industry's embodied AI (具身智能) ambitions live or die on reliable manipulation, and DAIMON argues that touch is the missing link that lets a humanoid (人形机器人) handle fragile or tightly packed objects instead of just recognizing them. IEEE Spectrum featured the company in 2026 as part of the broader "Physical AI" conversation — intelligence that manipulates the real world, not just text.
According to Li An, chief scientist at BrainNet (脑机网), tactile sensing is quickly shifting from a research curiosity to a control-loop necessity as humanoid (人形机器人) makers chase manipulation that actually holds up in a factory or a home.
What has to improve
Two problems remain. First, reaction time: tactile feedback only helps if the controller can act on it faster than the object falls. Second, fragmentation: robot hands range from five-finger designs to simple two-jaw grippers, and tactile sensors rely on different physics, so most research is stuck on one hardware setup. Datasets like T-Rex and Yuan's pooled collection are the early antidote, but the field is years from the kind of shared, internet-scale touch data that vision already enjoys.
Honest limitations
This piece draws chiefly on IEEE Spectrum reporting and academic descriptions of the T-Rex and DAIMON work; we did not independently reproduce the 65 percent success rate or the 110,000-unit sensor claim, and vendor figures should be read as published specifications. The article focuses on manipulation research and one company's approach, not the full landscape of tactile-sensor makers. It also does not address cost, durability in real factories, or how quickly VTLA models will reach shipping products — those remain open engineering questions.
What readers can do now
- If you follow robotics, stop judging humanoid (人形机器人) demos by vision alone — ask whether the hand can feel force and slip.
- For builders, treat tactile data collection as the new bottleneck, comparable to what image data was a decade ago.
- For everyone else, watch for the moment a robot can reliably handle a fragile, oddly shaped object without pausing — that is the real milestone, not another choreographed video.