Published in Autonomous Robots, with Perla Maiolino, Ed Bray, and Fumiya Iida.
When a robot palpates a soft body, the way it touches shapes what it feels. In this paper, we propose a framework to study how a robot’s physical interactions influence the tactile information it gathers, and how to choose the interactions that make that information most useful.
We fitted a capacitive tactile sensor to a 3D-printed end-effector on a robotic arm and used it to probe soft silicone phantoms. Hidden inside were hard spherical inclusions, 7 or 20 mm across, placed 5 or 15 mm below the surface. Each location was probed with different motion strategies. We then compressed the resulting tactile data with principal component analysis (PCA) and grouped it with k-means clustering, without using any labels.
The results show that the right probing strategy can compensate for data that is limited or of poor quality. In a final palpation test, we trained a support vector machine (SVM) to tell the types of inclusion apart. The unsupervised cluster analysis alone, with no labels, predicted which motion strategy would work best and how the classifier’s performance would compare across strategies.
The machine learning side
Another part of the project asks how a robot can choose its actions, and even the shape of its body, to make touch more informative. I explain it in my talk at IEEE RoboSoft 2020, below. The paper is covered in “Efficient Bayesian Exploration for Soft Morphology-Action Co-optimization”.