Gabriele M. Caddeo;Andrea Maracani;Paolo D. Alfano;Nicola A. Piga;Lorenzo Rosasco;Lorenzo Natale
{"title":"Sim2Surf: A Sim2Real Surface Classifier for Vision-Based Tactile Sensors With a Bilevel Adaptation Pipeline","authors":"Gabriele M. Caddeo;Andrea Maracani;Paolo D. Alfano;Nicola A. Piga;Lorenzo Rosasco;Lorenzo Natale","doi":"10.1109/JSEN.2025.3530712","DOIUrl":null,"url":null,"abstract":"In this work, we tackle the simulated to real (Sim2Real) gap in vision-based tactile sensors for surface classification. Specifically, we target four surface types: flat, curved, edge, and corner. We first train a diffusion model (DM) with a small set of real-world, unlabeled images gathered from everyday objects using the DIGIT sensor. Next, we employ the TACTO simulator to generate images by uniformly sampling object surfaces from the YCB model set. The images are transformed into the real domain using the DM and automatically labeled, allowing us to train a surface classifier without the need of manual annotations. To further align features across the real and simulated domains, we use an adversarial approach during training. The evaluation on tactile images from 15 3D-printed YCB objects shows an accuracy of 83.2%, significantly higher than the 35.9% achieved by training only on simulated images, confirming the effectiveness of our method. Testing on data acquired with different DIGIT sensors yields 81.6% accuracy on average, outperforming training with labeled real data. Moreover, our method demonstrates robustness also when applied to a different vision-based tactile sensor, GelSight Mini, obtaining a remarkable 83.3% accuracy on a balanced dataset. Lastly, we validate our approach with a 6-D object pose estimation task using tactile data.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 5","pages":"8697-8709"},"PeriodicalIF":4.3000,"publicationDate":"2025-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10851808","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Sensors Journal","FirstCategoryId":"103","ListUrlMain":"https://ieeexplore.ieee.org/document/10851808/","RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, ELECTRICAL & ELECTRONIC","Score":null,"Total":0}
引用次数: 0
Abstract
In this work, we tackle the simulated to real (Sim2Real) gap in vision-based tactile sensors for surface classification. Specifically, we target four surface types: flat, curved, edge, and corner. We first train a diffusion model (DM) with a small set of real-world, unlabeled images gathered from everyday objects using the DIGIT sensor. Next, we employ the TACTO simulator to generate images by uniformly sampling object surfaces from the YCB model set. The images are transformed into the real domain using the DM and automatically labeled, allowing us to train a surface classifier without the need of manual annotations. To further align features across the real and simulated domains, we use an adversarial approach during training. The evaluation on tactile images from 15 3D-printed YCB objects shows an accuracy of 83.2%, significantly higher than the 35.9% achieved by training only on simulated images, confirming the effectiveness of our method. Testing on data acquired with different DIGIT sensors yields 81.6% accuracy on average, outperforming training with labeled real data. Moreover, our method demonstrates robustness also when applied to a different vision-based tactile sensor, GelSight Mini, obtaining a remarkable 83.3% accuracy on a balanced dataset. Lastly, we validate our approach with a 6-D object pose estimation task using tactile data.
期刊介绍:
The fields of interest of the IEEE Sensors Journal are the theory, design , fabrication, manufacturing and applications of devices for sensing and transducing physical, chemical and biological phenomena, with emphasis on the electronics and physics aspect of sensors and integrated sensors-actuators. IEEE Sensors Journal deals with the following:
-Sensor Phenomenology, Modelling, and Evaluation
-Sensor Materials, Processing, and Fabrication
-Chemical and Gas Sensors
-Microfluidics and Biosensors
-Optical Sensors
-Physical Sensors: Temperature, Mechanical, Magnetic, and others
-Acoustic and Ultrasonic Sensors
-Sensor Packaging
-Sensor Networks
-Sensor Applications
-Sensor Systems: Signals, Processing, and Interfaces
-Actuators and Sensor Power Systems
-Sensor Signal Processing for high precision and stability (amplification, filtering, linearization, modulation/demodulation) and under harsh conditions (EMC, radiation, humidity, temperature); energy consumption/harvesting
-Sensor Data Processing (soft computing with sensor data, e.g., pattern recognition, machine learning, evolutionary computation; sensor data fusion, processing of wave e.g., electromagnetic and acoustic; and non-wave, e.g., chemical, gravity, particle, thermal, radiative and non-radiative sensor data, detection, estimation and classification based on sensor data)
-Sensors in Industrial Practice