Abstract

We propose a method that estimates a location of the extrinsic contact line, a contact line between a grasped object and an external environment, then uses it as an input to a peg-in-hole insertion policy. A graphical model and an active tactile feedback controller work collaboratively to get an accurate estimate. The controller helps the graphical model to get a better estimation by regulating a consistent contact mode. The better estimation from the graphical model makes it easy for the controller to regulate the contact. With such an estimation module, input to the insertion policy reduces from a high-resolution tactile image to a single contact line. It enables the insertion policy to train in a simple simulation rather than in a real experiment. Lastly, we demonstrate and evaluate the active extrinsic contact line estimation and the trained insertion policy together in a real experiment. We show that the proposed method inserts various-shaped test objects with a higher success rate and fewer insertion attempts than previous work with end-to-end approaches.

Contact: Sangwoon Kim (sangwoon@mit.edu)