Touch2Touch and T2D2: Cross-Sensor Tactile Image Generation (CoRL 2025 Oral)

Forum topic · 二一 · 2026-05-13

Summary

Tactile sensors like GelSight, DIGIT, and OmniTact each have unique optical systems, resolutions, and imaging characteristics, so tactile models trained on one sensor cannot be directly used with another. A CoRL 2025 Oral paper introduces two cross-sensor tactile image generation methods: Touch2Touch, which performs direct end-to-end conversion when paired data is available, and T2D2 (Touch-to-Depth-to-Touch), which converts images through an intermediate depth representation without requiring paired data. This enables practical model transfer: a grasp-detection model trained on a DIGIT sensor can be reused on a GelSight sensor without retraining by converting new tactile images into DIGIT-style inputs via T2D2. The approach was validated on cup stacking and tool insertion tasks. The key insight is that different tactile sensors are simply different 'renderings' of the same physical contact, and depth serves as a universal intermediate language—sensors observe the same geometric deformation under different lighting and viewpoints.

Cross-Sensor Touch Generation (CoRL 2025 Oral)

Tactile sensors are among the most critical and diverse components in robotics — GelSight, DIGIT, OmniTact, and others each have their own optical systems, resolutions, and imaging characteristics. As a result, a tactile model trained for one sensor cannot be directly applied to another.

A CoRL 2025 Oral paper proposes two methods for cross-sensor tactile image generation:

Touch2Touch: direct end-to-end conversion when paired data is available.
T2D2 (Touch-to-Depth-to-Touch): converts tactile images through an intermediate depth representation, requiring no paired data.

Why it matters

Suppose you have a grasp-detection model trained on a DIGIT sensor, and you switch to a GelSight sensor — no retraining needed. Simply pass the new sensor's tactile images through T2D2 to convert them into DIGIT-style images, then feed them to your existing model. The approach was validated on cup stacking and tool insertion tasks.

Core insight

Different tactile sensors are just different "renderings" of the same physical contact. Depth serves as the universal intermediate language — different sensors are merely observing the same geometric deformation under different lighting and viewpoints.

*Source: [Cross-Sensor Touch Generation / CoRL 2025 Oral]*