T-Rex and 4 Rivals Are Racing to Teach Robots to Feel

A robot arm can watch millions of images of a coffee mug and still crush the handle the first time it actually grips one, because almost none of the data that trains it records what a grip feels like. Five research groups published tactile-manipulation datasets in 2026 to close that gap, and the group that collected the most data says robots still need more than 3 times as much again before the gap closes.

The clearest demonstration comes from the University of California, Berkeley. A team led by computer science professor Trevor Darrell built T-Rex, a system trained on 100 hours of tactile-rich recordings across more than 200 household objects, covering motions like wiping, grasping, twisting and pouring, according to the paper posted to arXiv on June 15. The paper reports that T-Rex beats the strongest prior baseline by more than 30 percent on average across 12 manipulation tasks, a result IEEE Spectrum's reporting put at a 65 percent average success rate, describing it as nearly double what comparable vision-language-action models achieve.

Why vision alone is not enough

Most robot manipulation models today train on video and language, the same two data types that built large language models, because cameras and text are cheap to collect at scale. Touch is not: a tactile sensor has to be built into a gripper, calibrated to a specific hardware design, and worn on a robot doing the exact task being recorded, which is why the field has comparatively little of it. A robot that has only ever seen a mug, and never felt one slip, has no data telling it how much force keeps a handle from cracking versus how much force lets it fall.

The other four datasets

Tsinghua University in Beijing took a different approach: rather than collecting new touch data, a team led by master's student Chengbo Yuan aggregated more than 3,000 hours of existing tactile recordings spanning 21 different sensor types into one hardware-agnostic model, an effort modeled on the Open X-Embodiment project that did the same for robot vision data. NeoteAI, a spin-out from Shanghai's Fudan University, built the largest dataset of the group, more than 30,000 hours of synchronized visual and tactile demonstrations captured with its own proprietary sensor on robotic arms and handheld grippers. A team at the University of Southern California trained a model on 2,700 demonstrations that infers touch from vision alone, lifting success on contact-rich tasks to 62.8 percent from a 28.2 percent baseline. Researcher Long Cheng at the Chinese Academy of Sciences presented a fifth approach at the IROS 2026 conference, a model that predicts what a surface should feel like from what it looks like, then checks that prediction against real sensor input.

Dataset	Group	Tactile data
T-Rex	UC Berkeley	100 hours
Aggregated model	Tsinghua University	3,000+ hours, 21 sensor types
n0-twam	NeoteAI (Fudan University)	30,000+ hours
Vision-to-touch	University of Southern California	2,700 demonstrations

How much data would actually be enough

Shunlin Lu, a Fudan University postdoc and NeoteAI's chief technology officer, told IEEE Spectrum the field needs roughly 100,000 hours of tactile data before robot manipulation reaches the kind of reliability that made large language models useful outside a lab. NeoteAI's own 30,000-hour dataset, the largest published so far, covers less than a third of that. None of the five groups has said what happens to model performance between today's data and Lu's 100,000-hour mark, because nobody has collected enough to test it, which is the gap every one of these projects is racing to fill first.

1. T-Rex (UC Berkeley)100 hours
2. Tsinghua aggregated model3000 hours
3. NeoteAI n0-twam30K hours
4. Estimated hours needed (Lu)100K hours

Source: IEEE Spectrum, Sept. 2026, and arXiv:2606.17055

What the T-Rex paper does not claim is that its 100-hour dataset generalizes past the more than 200 household objects it was recorded on, and its authors frame the release as a demonstration of a data-collection method rather than a finished model for a factory floor. That distinction matters for anyone reading the 65 percent figure as a production benchmark: it is a lab result on a fixed object set, the same caveat that applies to Arm's physical AI framework, which standardizes the compute a robot runs on without saying anything about what data trains it. Cheap, open-source robot arms already selling for $6,500 can run any of these models once trained, which means the bottleneck now sits entirely on the data side, not the hardware.