Panasonic T-Rex Tactile Robot AI Hits 65% in Lab Tests

Panasonic Holdings and Panasonic R&D Company of America, with collaborators at UC Berkeley, NVIDIA, Stanford, Sapienza University of Rome and ItalAI, said on 5 November 2026 that they have developed T-Rex (Tactile-Reactive Dexterous Manipulation), a robot manipulation system that corrects actions from fingertip touch in real time, according to a Panasonic HD press release, the arXiv paper and the project page. The work is a laboratory evaluation presented for CoRL in Austin from 9 to 12 November 2026 — not a commercial robot product.

On 12 contact-rich real-world tasks, T-Rex reached an average success rate of 65%, against 35% for the best baseline cited. Without tactile information, average success fell to 42%, Panasonic says.

How the tactile loop is supposed to work

T-Rex couples a slower vision-language action generator with a faster fingertip correction module. Motion from vision and language runs at about 5 Hz; tactile correction from ten fingertip sensors runs at about 20 Hz, so the robot can react to slip or excess force without re-running full image recognition each cycle. Training combined more than 200 everyday objects with 22 basic actions into over 7,700 bimanual episodes (100 hours), after pre-training on 22,889 hours of human egocentric video and fine-tuning with about 100 demonstrations per task.

Those numbers are evaluation metrics from the paper and Panasonic’s write-up, not field-robot warranties. Tasks in the suite include page turning, transferring eggs, applying toothpaste, unlocking with a key, pulling a card and screwing in a light bulb — contact-rich chores chosen to stress force and soft-object handling.

Dexterity research is racing factory roadmaps

Panasonic already told the market it wants humanoid robots by 2029; T-Rex is an earlier research brick for contact-rich hands, not that product timeline. Elsewhere, Boston Dynamics is training Atlas for factory work while touch-focused startups such as Zurich’s Enact raise money to give robots a sense of touch. The competitive question is whether high-frequency tactile correction can leave the lab without exotic sensor calibration every shift.

For UK manufacturers watching labour shortages in assembly and logistics — the use cases Panasonic names — the relevant test is still transfer: 65% on twelve scripted tasks is a research headline, not a line-ready SLA. UK health-and-safety and machinery rules would also apply long before any T-Rex-derived cell sat beside people.

Lab average is not a ship date

Watch next for the CoRL talk, any open release of the 100-hour dataset beyond the project page, and whether Panasonic folds the correction stack into a named industrial pilot. Until a product SKU appears, tactile robot AI here means T-Rex’s published lab scores — 65% with touch, 42% without — and a research collaboration, not a robot you can buy.

Why 20 Hz touch correction changes the VLA bet

Most vision-language-action stacks update too slowly for slip and soft contact. T-Rex’s design bet is that a dedicated tactile expert can refine actions an order of magnitude faster than the vision tower without discarding large-scale video pre-training. The ablation that drops average success from 65% to 42% without touch is Panasonic’s evidence that the fast path earns its complexity. Still, lab trials with calibrated fingertip sensors are kinder than oily factory floors; sensor drift and missing palm sensing are limitations the project page itself flags.

Panasonic’s stated aim — easing manufacturing and logistics labour shortages through finer insert, remove and soft-object handling — only becomes news for plant managers once a named cell, cycle time and safety case exist. Until then, T-Rex is a CoRL paper with strong contact-rich scores, not a shift roster replacement.

CoRL is the near-term milestone

The Austin CoRL window of 9–12 November 2026 is when peer researchers will pressure-test the 65% macro-average and the tactile ablations. Open questions for that audience include how well the temporal VQ-VAE encoder transfers across sensor hardware and whether the ~100-demo fine-tune recipe survives distribution shift on messier objects. Those are research questions; they are also the right ones for any OEM that might later licence the stack.

Sources

• Panasonic HD Co-Develops a Robot Manipulation AI… (5 November 2026)
• T-Rex: Tactile-Reactive Dexterous Manipulation (arXiv:2606.17055)
• Tactile-Reactive Dexterous project page

Related stories

• Panasonic Wants to Build Humanoid Robots by 2029…
• Boston Dynamics Trains Atlas for Factory Work as Spot 5.2 Goes Agentic
• Zurich’s Enact Says It Raised $30m to Give Robots a Sense of Touch