Summary
Tactile sensors such as GelSight, DIGIT, and OmniTact each have distinct optical systems, resolutions, and imaging characteristics, so tactile models trained on one sensor cannot be directly reused on another. A CoRL 2025 oral paper addresses this with two cross-sensor tactile image generation methods. Touch2Touch performs direct end-to-end translation when paired data between the two sensors is available. T2D2 (Touch-to-Depth-to-Touch) requires no paired data: it converts a tactile image into a shared intermediate depth representation, then renders it into the target sensor's style. The key insight is that different tactile sensors observe the same physical contact as different 'renderings,' with depth serving as a universal intermediate language—sensors simply view the same geometric deformation under different lighting and viewpoints. Practically, this means a grasp-detection model trained on DIGIT can be reused on GelSight without retraining: the new sensor's tactile images are converted into DIGIT-style images via T2D2 and fed to the existing model. The approach is validated on cup stacking and tool insertion tasks, enabling portable tactile perception across heterogeneous robot hand hardware.
Tactile sensing is one of the most critical yet most diverse components in robotics. Sensors like GelSight, DIGIT, and OmniTact each feature their own optical systems, resolutions, and imaging characteristics. The consequence: a tactile model trained on one sensor cannot be directly applied to another.
A CoRL 2025 Oral paper proposes two cross-sensor tactile image generation methods:
- Touch2Touch — direct end-to-end translation, used when paired data between the two sensors is available.
- T2D2 (Touch-to-Depth-to-Touch) — requires no paired data; the conversion goes through an intermediate depth representation.
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 — just 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 is validated on cup stacking and tool insertion tasks.
Core insight
Different tactile sensors see different *renderings* of the same physical contact. Depth serves as a universal intermediate language — sensors are merely observing the same geometric deformation under different lighting conditions and viewpoints.
*Reference: Cross-Sensor Touch Generation / CoRL 2025 Oral.*
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/177619964