TouchReason

Physical evidence for tactile-grounded reasoning

Paper · Project Page · GitHub · Touch-R1 Model

TouchReason is the dataset release hub for Touch-R1: Reinforcing Touch Reasoning in MLLMs. It connects multi-sensor contact observations with physical-property annotations and reasoning-oriented question answering, supporting models that use touch as physical evidence.

The research introduces TouchReason-1M for tactile-grounded learning and TouchReason-Bench for evaluating physical reasoning across sensors.

Release status: Dataset card available. Data files and loading instructions are in preparation. Collection figures below refer to the acquisition corpus reported in the paper.

Collection Overview

The collection combines heterogeneous optical tactile observations with synchronized contact-force measurements. It captures how objects respond to normal pressing and exploratory contact motions, connecting sensor observations with consistent physical judgments.

Paper-Reported Collection

Component	Scope
Synchronized tactile–force records	1,323,000
Valid contact sequences	14,700
Everyday objects	1,000+
Material categories	9
Optical tactile sensors	4

Sensors: DM-Tac X, GelSight Mini, Xense, and Tac3D.

A synchronized record couples tactile observations with deformation, shear, depth, force, and associated metadata. The research also provides ordinal physical-property labels and reasoning-oriented QA annotations.

Research Tasks

• Physical-property perception: interpreting hardness, roughness, protrusion, and material from contact.
• Ordinal and comparative reasoning: comparing physical attributes while respecting their ordered structure.
• Cross-sensor consistency: relating evidence from different optical tactile sensors.
• Visual–tactile conflict resolution: revising appearance-based judgments using contact evidence.

The accompanying TouchReason-Bench contains 4,800 QA pairs over 200 held-out objects, covering tactile perception and structured physical reasoning. See the paper for benchmark construction and evaluation protocols.

Access and Related Resources

Released archives, file organization, split statistics, and loading examples will be documented here when the data files are published.

• Model: Touch-R1
• Collection videos and object atlas: Project page
• Research code announcements: GitHub repository

Citation

@misc{lai2026touchr1reinforcingtouchreasoning,
  title         = {{Touch-R1}: Reinforcing Touch Reasoning in {MLLMs}},
  author        = {Yingxin Lai and Yafei Zhou and Fucai Zhu and Siyu Zhu and Weihao Yuan},
  year          = {2026},
  eprint        = {2605.27154},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.27154}
}