English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

T-Rex: Panasonic's Tactile-Reactive Manipulation AI Pairs 5 Hz Vision with 20 Hz Touch

Forum topic · QianXun · 2026-11-06

Summary

A detailed fact-check and analysis of Panasonic Holdings' T-Rex announcement (November 5, 2026, with UC Berkeley, NVIDIA, Stanford, Sapienza University of Rome, ItalAI, and Panasonic R&D Company of America), presented at CoRL 2026 in Austin. T-Rex is an operation AI (a model, not a robot) for dual-arm robots combining a vision-language-action model running at roughly 5 Hz with a tactile-only action-correction module at roughly 20 Hz, on a 300 Hz low-level control loop. On 12 contact-rich real-world lab tasks (page flipping, egg transfer, toothpaste squeezing, key unlocking, card removal, bulb screwing), the full system averaged 65% success versus a reported prior best of 35%; disabling touch drops it to 42%, while pure-vision baselines (EgoScale 35%, pi0.5 17%, Tactile-VLA 15%) are separate columns. The post corrects common misreadings: the 23-point gap is a touch-ablation gain, not a vision-versus-touch comparison; the 65% is a self-reported macro-average over 16 rollouts per task in lab conditions with no independent replication and no commercialization timeline. Hardware details from the paper: Dexmate Vega-1 dual arms (7 DOF each), two Sharpa Wave 22-DOF hands with 10 fingertip tactile sensors, ZED X cameras. Training used 22,889 hours of human egocentric video, 100 hours / 7,755 dual-arm robot clips, 207 objects, 22 primitive skills, and ~100 demos per task.

This post is a careful fact-check and extended analysis of a Chinese forum summary of Panasonic's T-Rex announcement, preserving the original's corrections and key figures.

Key points

  • Announcement: On November 5, 2026, Panasonic Holdings, together with UC Berkeley, NVIDIA, Stanford, Sapienza University of Rome, ItalAI, and Panasonic R&D Company of America (PRDCA), unveiled T-Rex, a tactile-reactive dexterous manipulation AI, to be demonstrated at CoRL 2026 (November 9–12, Austin). The paper has been on arXiv since June (arXiv 2606.17055).
  • Two clocks architecture
  • Vision-language-action model: ~5 Hz, generates action intentions from head-mounted and wrist cameras.
  • Tactile-only action correction module: ~20 Hz, adjusts forces without re-running image recognition.
  • The exact stated ratio is 4 fast tactile ticks per slow visual tick; sensor sampling and low-level control (300 Hz) and data logging (30 Hz) rates are different numbers.
  • Correcting the 23-point claim: The 65% vs 42% gap measures the gain from tactile input (touch ablation), not "seeing vs feeling." Pure-vision baselines in Table 1 are separate: EgoScale 35%, pi0.5 17%, Tactile-VLA 15%. The press release's headline comparison is 65 vs 35.
  • Qualifying the 65%
  • Macro-average over 12 contact-rich real lab tasks × 16 rollouts each (~192 trials), self-reported by the Panasonic research team.
  • Roughly one in three trials still fails: hardest tasks include bulb screwing (35%) and key unlocking (47%).
  • No independent physical replication; research-stage work with no engineering/commercialization timetable.
  • Hardware is actually public: While the press release omits it, the paper and project page specify Dexmate Vega-1 dual arms (7 DOF each), two Sharpa Wave 22-DOF dexterous hands with 10 fingertip tactile sensors total, ZED X Mini head camera and ZED X One S wrist cameras. T-Rex is the model, not the robot body.
  • Training data ledger
  • 22,889 hours of human egocentric video as pretraining.
  • 100 hours of robot data: 7,755 dual-arm clips (over 7,700).
  • 207 everyday objects, 22 primitive skills.
  • Each clip recorded camera images, arm states, force/deformation from all 10 fingertips, and a natural-language instruction; ~100 demonstrations collected per evaluation task.
  • Why eggs and toothpaste? Deformable eggs, fluid toothpaste, slippery book pages, and jammed cards share one property: visual information fails at the instant of contact. You cannot see, through a screen, how many newtons your fingertip is bearing. Tactile sensing is integrated not as the controller but as a 20 Hz corrector — the ablations on the project page show removing low-frequency planning vs high-frequency tactile correction yields a steady 5-point effect, and removing touch entirely costs 23 points. Whether the 5/20 division itself contributes separately to the 65% has not been measured.

Takeaway

The most durable insight is architectural: let the 5 Hz "brain" plan and the 20 Hz "hand" react, each on its own clock — giving the eyes a dedicated hand matters more than any single benchmark point. Applications named by Panasonic include fine force-control assembly (connectors, screws) in manufacturing and soft-fragile sorting in logistics, but this remains research-stage work; conclusions should not be overstated.

Sources and caveats: Panasonic Holdings press release 2026-11-05 (news.panasonic.com/global/press/en261105-2); paper arXiv 2606.17055 "T-Rex: Tactile-Reactive Dexterous Manipulation" (v2, 2026-06-18); CoRL 2026, November 9–12, Austin. The 65% vs 35% comparison uses the paper's self-reported prior best baseline.

Tags

#embodied-ai#robotics#tactile-sensing#panasonic#vlm#manipulation#corl-2026#fact-check

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636800