T-Rex: Panasonic and Berkeley unveil a bimanual robot that corrects its movements in real time via touch

T-Rex: Panasonic and Berkeley present a bimanual robot that corrects its movements in real time through touch

🔎 Robots know how to see. They don't know how to touch.

That is the observation that motivated T-Rex, a dexterous manipulation system presented on November 5, 2026 by Panasonic HD, developed with UC Berkeley, NVIDIA, Stanford, Sapienza University of Rome, and ItalAI. Its full name: Tactile-Reactive Dexterous Manipulation. It will be officially presented at CoRL 2026, the leading conference in robot learning, held in Austin from November 9 to 12.

The problem T-Rex tackles is simple to state, formidable to solve. The vision-language-action (VLA) large models powering today's robotic arms operate from images and text instructions. They plan a movement, execute it, then look at the result. In between, they are blind to contact. Yet human manipulation — turning a key, picking up an egg without crushing it, extracting a jammed sheet of paper — relies precisely on this real-time tactile feedback loop.

T-Rex therefore integrates 10-finger touch directly into the control loop. The result is measurable: a 65% average success rate on 12 real-world tasks, versus 35% for the best baseline model, and 42% for the same architecture stripped of tactile feedback. These figures, sourced from the Panasonic HD press release, place T-Rex among the highest-performing bimanual manipulation systems ever published to date.

And above all, the team opens everything: the 100-hour dataset, the architecture, the code. The official GitHub repository is public.

The Essentials

• T-Rex is a tactile VLA developed by Panasonic HD with UC Berkeley, NVIDIA, Stanford, Sapienza, and ItalAI, presented at CoRL 2026 (Austin, November 9–12).
• Asynchronous architecture: a vision-language motion generation module runs at ~5 Hz, while a tactile correction module refines gestures at ~20 Hz — without re-running image recognition.
• Results: 65% average success rate across 12 real-world tasks (turning a page, handling an egg, squeezing toothpaste, turning a key, picking up a card, screwing in a lightbulb), versus 35% for the best baseline and 42% without tactile feedback.
• Data: 7,700 bimanual episodes (100 hours), 200+ objects, 22 base actions, pre-training on 22,889 hours of human egocentric video, fine-tuning with ~100 demonstrations per task.
• Openness: a 100-hour tactile-reactive dataset made public (a ~50-hour subset in LeRobot v3.0 format), with the Mixture-of-Transformers architecture on a Qwen3-VL-2B backbone documented in the arXiv paper.
• Industrial targets: component insertion, removal, and screwing in factories, and handling of soft or fragile objects in logistics.

Recommended Tools

Why Haptics is the Missing Link in Dexterous Robotics

Because vision alone plateaus at around 35-42% success on precise contact tasks. This is the quantified lesson from T-Rex, and it is more important than the system itself.

Classic VLAs handle haptics in two failing ways, as summarized in the synthesis from EE Herald: either they ignore it completely, or they treat it as a static input—a snapshot captured at the beginning of the gesture, never updated. In either case, the robot cannot detect if it is gripping too hard, if an object is slipping, or if a key is not aligned with the lock.

Think about what you do when turning a page. You do not plan the gesture from a photo: your fingers constantly measure the resistance of the paper, adjust the pressure, and compensate for micro-slips. This is a high-frequency control loop, not a single decision. No 5 Hz vision-language pipeline can reproduce this by extracting images at each iteration—the computation would be prohibitive and the latency prohibitive.

This is exactly the architectural problem that T-Rex solves, and that is why this work deserves attention far beyond the Panasonic laboratory.

How T-Rex works: three experts, two frequencies

T-Rex is an asynchronous three-expert Mixture-of-Transformers (MoT) architecture, built on a Qwen3-VL-2B backbone. The full details are in the arXiv paper, but the logic boils down to three roles:

1. The latent expert predicts the visual evolution of the scene — it anticipates what the camera is about to see.
2. The action expert denoises motion trajectories at low frequency, around 5 Hz, combining vision and language. This is the "planning brain".
3. The tactile expert refines those trajectories at high frequency, around 20 Hz, by reusing the vision-language context already in cache. No new image recognition: the tactile module works in parallel, on contact signals from 10 fingers.

The coupling between the two frequencies is done via cascaded flow matching — a generation technique in which a coarse model produces a trajectory that a finer model refines. The tactile encoder is a temporal VQ-VAE, which compresses pressure sensor sequences into discrete representations that the transformer can work with.

The choice of the Qwen3-VL-2B backbone is significant

The team did not train a giant model from scratch. They started from a compact 2-billion-parameter VLM and added the specialized experts around it. This is a clear trend in 2026 robotics: performance comes from the control architecture and the data, not from raw model size. The result — 65% success — is achieved with a system that runs in real time on a physical robot, not in a simulation.

The data collection hardware

The data was collected on a Dexmate Vega-1 bimanual robot equipped with Sharpa Wave hands, with Manus gloves and VIVE trackers to capture human demonstration gestures. The recording pipeline stores tactile signals in HDF5 and video streams in MP4, with lossless compression — a technical detail that matters for anyone looking to reuse the dataset.

The numbers: 65%, 35%, 42% — what they really say

The +30-point success rate gain is the largest ever reported on this dexterous manipulation benchmark. According to the arXiv paper, T-Rex outperforms the reference dexterous-hand foundation models by +30% on average across the 12 tasks.

Let's break it down:

The most telling number isn't the 65%. It's the 42%. Remove the tactile from T-Rex's own system — same architecture, same data, same training — and you lose 23 points. This means that nearly half of the system's performance comes exclusively from the 20 Hz tactile correction loop. It's not a sensory gimmick: it's the component that carries half of the result.

The 12 tasks were chosen to be honest

Not choreographed demos. The benchmark covers four families of difficulty: insertion (key in a lock, card in a reader), deformation (turning a page, squeezing a toothpaste tube), force control (grasping an egg without breaking it, screwing in a light bulb without shattering it) and bimanual coordination (both hands working together on the same object). These tasks are notoriously difficult for robots precisely because they require continuous contact feedback.

An open dataset of 100 hours: the underestimated argument

The T-Rex dataset is probably the project's most lasting contribution. 7,700 bimanual episodes, 100 hours of tactile-reactive data, 200+ objects, 22 base actions — all public, with a subset of about 50 hours already converted to the LeRobot v3.0 format, the emerging standard for open robotics data.

Why does this matter? Because the bottleneck for VLAs is no longer compute, it's contact data. There are millions of hours of human video on the internet, but almost nothing that combines vision, language, and synchronized tactile signals for dexterous manipulation. T-Rex delivers exactly that combination.

The training recipe is reproducible

The pipeline is documented: pre-training on 22,889 hours of human egocentric video (to learn the structure of human gestures), then fine-tuning on the tactile-reactive data collected on the robot, with around 100 demonstrations per task to adapt the system to a new skill. One hundred demonstrations is a reasonable order of magnitude for an industrial team that wants to specialize a robot on a precise assembly operation.

This is no small thing. If 100 demos are enough for a new task, the cost of industrial deployment becomes calculable — and therefore fundable.

Industrial applications: where Panasonic wants to cash in

Panasonic is explicitly targeting the factory and logistics, not the domestic robot. The press release cites two families of use cases: insertion, removal, and screwdriving of components in manufacturing environments, and the handling of soft or fragile objects in logistics.

It's a coherent strategic choice. Insertion and screwdriving tasks have been the Achilles' heel of automation for thirty years: conventional industrial arms handle repetitive motions in structured environments, but fail as soon as they need to feel a bit engaging with a screw. A system that succeeds 65% of the time on tasks of this kind, with a clear path to improvement (more data, more tasks), becomes economically interesting well before reaching human-level rates.

The handling of fragile objects — foodstuffs, electronic components, pharmaceutical products — is the second market. Here again, touch is the answer: a robot that measures the pressure it exerts can handle a piece of fruit or a vial without crushing it, something no trajectory-programmed gripper can do reliably.

This trajectory toward industrial deployment fits into a broader trend I've covered elsewhere: Xiaomi puts its humanoids to the test with 8 hours of auto parts sorting in a factory, proof that major industrial players are now testing versatile robots in real production flows, not just for show.

What T-Rex Does Not Solve

T-Rex is an advance in manipulation, not an advance in cognition. We need to be precise about the limitations, because press coverage of robotic systems tends to inflate the promises.

First limitation: the 65% success rate remains far from a human operator, who approaches 99% on these tasks. A robot that fails one out of three times on extracting a sheet of paper is not deployable as-is on a production line without supervision.

Second limitation: the system is trained on 22 base actions and 200 objects. Generalization to completely new objects and gestures outside this distribution remains an open question — the paper reports results on its own benchmark, not on a universal evaluation.

Third limitation: the hardware. The setup relies on a Vega-1 robot with Sharpa Wave hands and instrumented gloves for data collection. Reproducing the results requires specific equipment and 10-finger tactile sensors, which are not yet standard industry components.

Finally, T-Rex says nothing about mobility or global environment perception. It is a fixed-station manipulation system. For a complete mobile bimanual robot, you need to look elsewhere — for example, NYU publishes Mabel, a mobile bimanual robot with 56 degrees of freedom for $8,056, which tackles the problem from the angle of a low-cost physical platform rather than the sensory loop.

T-Rex in the landscape: complementary, not competing

T-Rex occupies a specific niche: the sensory loop for dexterous manipulation. It does not directly compete with the other robotics efforts of 2026 — it stacks on top of them.

On one side, physical platforms are progressing: NYU's Mabel makes mobile bimanual hardware accessible, and Xiaomi's industrial trials show that humanoids can sustain factory work rates. On the other, control interfaces are progressing: real-time voice models like OpenAI's GPT-Realtime-2 and Google's Gemini 3.8 Live make voice instruction of a robot feel natural — background reasoning and tool execution without interrupting the conversation. And on the simulation and environment generation front, AlayaWorld is paving the way for real-time generated worlds that could serve as synthetic training grounds for manipulation policies.

T-Rex fills in the missing layer: high-frequency contact feedback, which is what separates a robot that watches its movement from a robot that feels it. The complete stack for a competent factory robot in 2027 will probably look like this: a tactile VLA like T-Rex for the hands, a mobile platform for locomotion, and a real-time voice interface for the human operator.

❌ Common Errors

Error 1: Believing that touch replaces vision

The figure of 42% without touch can make one believe that vision becomes secondary. This is false: without the pre-training on 22,889 hours of egocentric video and the latent expert that predicts visual evolution, the tactile expert would have nothing to correct. Both modalities are complementary—vision plans, touch refines. The solution: always evaluate multimodal systems by removing each modality, as done by the T-Rex benchmark.

Error 2: Extrapolating the 65% to all tasks

The 65% is an average across 12 specific tasks, with ~100 demonstrations per task. It does not translate to an object or gesture outside of the dataset. The solution: for any deployment, count the necessary demonstrations for your task and measure the success rate on your own benchmark, not the one from the paper.

Error 3: Underestimating the complexity of tactile hardware

10-finger contact sensors, Manus collection gloves, and Sharpa Wave hands are not components you plug in in an afternoon. Many teams fail to reproduce tactile-reactive results because they neglect the quality and synchronization of the capture. The solution: start with the recording pipeline documented in the repo officiel — HDF5, MP4, lossless compressed tactile — rather than improvising a custom format.

Error 4: Confusing "open dataset" with "ready-to-use"

The dataset is public, but only ~50 hours out of 100 are in LeRobot v3.0 format. Using the rest requires processing raw HDF5 files and aligning tactile streams with videos. The solution: start with the LeRobot subset, validate your training pipeline, and then expand.

❓ Frequently Asked Questions

What exactly is T-Rex?

T-Rex (Tactile-Reactive Dexterous Manipulation) is a bimanual vision-language-action model that integrates 10-finger tactile feedback into its control loop. Developed by Panasonic HD together with UC Berkeley, NVIDIA, Stanford, Sapienza, and ItalAI, it generates movements at ~5 Hz and corrects them at ~20 Hz via touch, achieving 65% success on 12 real-world tasks.

Why are the two frequencies (5 Hz and 20 Hz) important?

Because they reflect the actual structure of human movement: visuo-motor planning is slow, while contact correction is fast. By decoupling the two — the tactile expert reuses the cached vision-language context without re-running image recognition — T-Rex achieves contact reactivity without an explosion in compute cost.

Is the T-Rex dataset really accessible?

Yes. The 100 hours of bimanual data (7,700 episodes, 200+ objects, 22 actions) are published on the official GitHub repository, with a subset of about 50 hours in LeRobot v3.0 format. The files combine HDF5 for tactile signals, MP4 for video, and lossless compression.

Can T-Rex be used in industry today?

Not as it stands. The 65% success rate and the dependence on specific hardware (Vega-1, Sharpa Wave hands) make it a research result, not a product. But the recipe — 100 demonstrations per task, open architecture, public dataset — gives an industrial team a foundation to draw on for building its own tactile-reactive system.

Where will T-Rex be presented?

At CoRL 2026 (Conference on Robot Learning), in Austin, from November 9 to 12, 2026. The paper is available on arXiv (2606.17055), and the official Panasonic HD press release has been online since November 5, 2026.

✅ Conclusion

T-Rex doesn't make robots human, but it finally gives them the sense they were missing: touch — and with 65% versus 35% on 12 real tasks, the numerical proof that the 20 Hz tactile loop is worth 30 performance points. Start with the open dataset and the arXiv paper: it's the best entry point available today into tactile-reactive manipulation.