Robotics

T-Rex: Tactile-Reactive Dexterous Manipulation

Dantong Niu, Zhuoyang Liu, Zekai Wang, Boning Shao, Zhao-Heng Yin, Anirudh Pai, Yuvan Sharma, Stefano Saravalle, Ruijie Zheng, Jing Wang, Ryan Punamiya, Mengda Xu, Yuqi Xie, Yunfan Jiang, Letian Fu, Konstantinos Kallidromitis, Matteo Gioia, Junyi Zhang, Jiaxin Ge, Haiwen Feng, Fabio Galasso, Wei Zhan, David M. Chan, Yutong Bai, Roei Herzig, Jiahui Lei, Fei-Fei Li, Ken Goldberg, Jitendra Malik, Pieter Abbeel, Yuke Zhu, Danfei Xu, Jim (Linxi) Fan, Trevor Darrell

Featured June 17, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Robots can now do tricky tasks like peeling fruit or opening locks by combining what they see and hear with a new "sense of touch" system that reacts super fast, thanks to a smart AI architecture and a big new dataset of human-like interactions.

In depth

The paper introduces T-Rex, a framework for tactile-reactive dexterous manipulation that addresses limitations in existing Vision-Language-Action (VLA) models. It combines large-scale human egocentric pre-training with a novel variable-rate Mixture-of-Transformer-Experts (MoT) architecture. This architecture integrates low-frequency visuomotor planning with high-frequency tactile refinement, enabled by a spatial-temporal tactile VQ-VAE encoder and a new 100-hour tactile-rich dataset.

Key Takeaways

1
The T-Rex framework achieves superior dexterous manipulation by integrating high-frequency tactile feedback into VLA models.
2
A novel variable-rate Mixture-of-Transformer-Experts (MoT) architecture allows for asynchronous processing of slow visual-language context and fast tactile signals.
3
The T-Rex Dataset, a 100-hour tactile-synchronized teleoperation dataset, and a three-stage training recipe are crucial for acquiring tactile-reactive behaviors.

Conceptual Flow

HIGH LEVEL

How T-Rex Learns Dexterous Touch

The robot learns complex hand movements by first watching many human videos, then practicing with real touch data, and finally refining specific skills.

Human Videos

Robot Touch Data

Specific Task Examples

Learn & Refine

Skilled Robot Actions

T-Rex Outperforms Other Robot Hands

This new robot system can do many delicate tasks much better than older systems, especially those needing a precise sense of touch.

Old Robot Methods

T-Rex Method

Compare Success

Much Higher Success Rate