RoboAtlas
EN
Model family

T-Rex

Participating organizations

Participating organizationsNVIDIAParticipating organizationsPanasonicParticipating organizationsStanford UniversityParticipating organizationsUC Berkeley

Version scope and roles for these associations await verification.

View evidence

T-Rex is a jointly developed tactile-reactive bimanual manipulation model. It combines vision-language action generation with fast tactile refinement, trained with roughly 100 hours of tactile-rich robot data and evaluated on 12 contact-rich tasks.

Overview

Model type
Vision-language-action
Listed versions
T-Rex

Model history

November 2026

Panasonic details joint T-Rex research and CoRL presentation plans

New resourcesOfficial announcement

The November5 official statement confirms the joint research, reports the bounded tactile-task experiments and announces CoRL2026 presentation plans. It does not establish a new model version, first artifact opening or commercial deployment.

Previously reported

T-Rex already had a June research archive. That paper date did not establish when code or weights became accessible.

This announcement

The November5 official statement confirms the joint research, reports the bounded tactile-task experiments and announces CoRL2026 presentation plans. It does not establish a new model version, first artifact opening or commercial deployment.

Ordinary institutional update:same existing research and conference plans, not a new capability/access milestone. Preserve November5 as the statement date rather than a model release.

T-Rex official research statement
Verified November 9, 2026

June 2026

T-Rex first paper describes tactile-reactive manipulation

New resourcesOfficial announcement

The June15 archive introduces variable-rate MoT action/tactile experts and temporal tactile VQ-VAE, evaluated on12contact-rich tasks. The date identifies the paper, not first code, data or weight opening.

Previously reported

The paper identifies limited dynamic tactile feedback in vision-language-action policies and a shortage of tactile-rich training data. This is its research baseline, not earlier artifact-access evidence.

This announcement

The June15 archive introduces variable-rate MoT action/tactile experts and temporal tactile VQ-VAE, evaluated on12contact-rich tasks. The date identifies the paper, not first code, data or weight opening.

Ordinary research history:dated paper disclosure, without proven same-day access or major deployable-capability change. Model score96 does not automatically make it featured.

T-Rex official research archive and revision history
Verified November 9, 2026
View all important updates →

Versions

1

Papers & reports (1)

T-Rex: Tactile-Reactive Dexterous Manipulation

Paper available

arXiv: 2606.17055DOI: 10.48550/arxiv.2606.17055
Authors & affiliations (34)

Author order: Original order

  1. Dantong NiuEqual contribution

    UC Berkeley · NVIDIA

  2. Zhuoyang LiuEqual contribution

    UC Berkeley

  3. Zekai WangEqual contribution

    UC Berkeley

  4. Boning Shao

    UC Berkeley

  5. Zhao-Heng Yin

    UC Berkeley

  6. Anirudh Pai

    UC Berkeley

  7. Yuvan Sharma

    UC Berkeley

  8. Stefano Saravalle

    La Sapienza University

  9. Ruijie Zheng

    NVIDIA

First published
2026-06-15
Latest paper revision
v2 · 2026-06-18
Venue
Conference on Robot Learning (CoRL) 2026

Covered model versions

  • t-rex · T-Rex

    T-Rex tactile-reactive framework and task-post-trained developer experiments; released pretrain/midtrain initializations are not exact paper-task checkpoints.

Paper ↗PDF ↗

Official resources

Submit a source or correction