The study addresses the Sim2Real gap in the field of vision-based tactile sensors for classifying object surfaces. They trained a Diffusion Model to bridge this gap using a small dataset of real-world images randomly collected from unlabeled everyday objects via the DIGIT sensor. The simulated images are then translated into the real domain using the Diffusion Model and automatically labeled to train a classifier. The evaluation was conducted on a dataset of tactile images obtained from a set of ten 3D printed YCB objects. The results show a total accuracy of 81.9%, a significant improvement compared to 34.7% achieved by the classifier trained solely on simulated images.
Publication date: 3 Nov 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2311.01380
