A robot model fine-tuned on about 100 teleoperated demonstrations averaged a 65 percent success rate across 12 manipulation tasks, nearly double the best vision-language-action model. The tasks included screwing in a light bulb, applying toothpaste to a toothbrush and transferring an egg between trays. The model came from a team at the University of California, Berkeley, and it depends on robot tactile data: recordings of what a robot hand feels as it works. Several groups are now building much larger collections of that data, and each is testing a different way to use it.
Why vision alone falls short
Vision-language-action (VLA) models are pretrained on large volumes of images, video and text, then fine-tuned on a smaller set of teleoperated robot demonstrations. They have improved how robots handle unfamiliar objects and environments. They now guide robots through jobs such as folding laundry, tidying living rooms and operating kitchen gadgets, using only a video feed and natural language instructions.
They still struggle when a task needs fine hand control. Handling deformable materials, plugging in a USB cable and turning a key in a lock are examples. VLAs do not use tactile feedback, which humans rely on in these situations. Trevor Darrell, a professor of computer science at Berkeley, says most dexterous manipulation can be done by humans with their eyes closed. He says force, slip and precise grasping are not handled well by traditional vision sensors.
The obstacle is data. Tactile signals look very different from the images VLAs are trained on, and tactile datasets are far smaller than the internet-scale collections behind vision and language models.
Berkeley splits planning from touch
Darrell's team first pretrained a model on existing datasets. It then added a tactile submodel trained on 100 hours of purpose-collected data. That data covers actions such as wiping, grasping, twisting and pouring, performed with more than 200 household objects.
Reaction time was the central design problem. Correcting a grip as it happens requires a response faster than most vision-language models can deliver. The team used two submodels, called experts. The action expert produces motion plans. The tactile expert runs four times faster and adjusts the plan in real time based on what the robot feels.
Darrell acknowledges one limit. All of the data comes from a single robot hardware setup.
Pooling data across sensors
Hand designs vary from fully articulated five-finger hands to simple pincer grippers. Tactile sensors also differ in their physics, from measuring changes in resistance to imaging a soft gel pad as it deforms. Chengbo Yuan, a master's student at Tsinghua University in Beijing, says this makes most tactile AI research sensor-specific and makes it hard for groups to share data or transfer what they learn.
Yuan aggregated more than 3,000 hours of tactile robot data from public datasets. The data covers 21 sensor types and a range of robot embodiments. His team converts each sensor's output into a shared format and maps it onto labeled positions on a template of a human hand. The resulting hardware-agnostic model performed much better than a baseline, including on hardware it had not seen in training. Yuan attributes this to the model acquiring some kind of common sense of tactile knowledge.
His group is now leading a collaboration of 80 institutions to collect a larger set of teleoperated demonstrations using standardized tactile data collection and processing.
Larger collections and touch inferred from images
Fudan University in Shanghai and its spin-out NeoteAI have built a tactile dataset an order of magnitude larger than earlier efforts. It uses a proprietary sensor mounted on various robotic arms and on a handheld gripper operated by humans. It holds more than 30,000 hours of demonstrations with synchronized visual and tactile data. The team trained a model that predicts what the robot should be feeling, then uses that prediction to guide and assess its actions. Shunlin Lu, a postdoctoral researcher at Fudan and CTO of NeoteAI, says the results show that large, diverse tactile data brings significant performance gains. He estimates that closer to 100,000 hours, gathered in varied real-world settings, could unlock new capabilities.
Researchers at the University of Southern California took a different route. They trained a model on more than 2,700 everyday manipulation demonstrations recorded with a handheld gripper that captures both tactile data and camera images. The model links images of the gripper contacting objects to the pressure the sensors measured at that moment. That gives robots without tactile sensors a rudimentary sense of touch, which the researchers found useful for contact-rich tasks. Their longer-term aim is to add generated tactile data to existing vision datasets.
Long Cheng of the Chinese Academy of Sciences in Beijing argues that more data is not enough. Vision supplies a continuous, high-bandwidth stream of pixels, while tactile signals are sparse and intermittent, so models learn to ignore them. His model predicts what the robot should feel from vision alone and compares that with the real tactile input. A large gap means the sensor has detected something the robot would otherwise miss, so those signals are amplified and predictable ones are dampened. Across five contact-rich tasks it averaged 62.8 percent success, against 28.2 percent for the same model without touch. The work is slated for presentation at IROS 2026.
What the results do and do not show
The strongest point in these results is that three different levers each produced gains. Berkeley's gains came from architecture, a fast tactile expert paired with a slower planner. Cheng's came from a training signal that stops models from ignoring touch. Fudan's came from scale. These are separate contributions, and nothing in the reported work shows how they combine. A large dataset used with a model that discounts touch may not deliver the gains that either piece shows alone.
The numbers also cannot be compared directly. The 65 percent figure covers 12 tasks and the 62.8 percent figure covers five, with different robots, tasks and baselines. Berkeley's comparison is to the best VLA model, and Cheng's is to the same model without touch. Neither result says which method is stronger.
Generalization is the open question. Berkeley's data comes from one hardware setup, and Yuan's work exists because tactile research is tied to specific sensors. His shared-format model handled unseen hardware better than a baseline, but his pooled data is far smaller than Fudan's 30,000 hours, and Fudan's data comes from a proprietary sensor. Whether a large single-sensor collection transfers to other sensors is not established by the reported results.
The scaling claim is also unproven. Lu's 100,000-hour figure is his own guess, and Yuan says tactile training has so far mainly made robots more efficient at tasks already within reach, such as picking and placing. He suspects tasks that are truly impossible without touch may need new algorithms. That is a narrower result than the framing of touch as the missing sense for dexterity. The field has early evidence that touch helps. It does not yet have evidence that touch unlocks tasks vision-only systems cannot do at all.
The next useful evidence would be shared benchmarks. The 80-institution standardized collection effort is the most direct step toward comparisons that hold across labs, because it would put different methods on common data and common processing.