EAGER: Robust Data-Driven Robotic Manipulation via Bayesian Inference and Passivity-Based Control

Abstract This proposal will develop new knowledge regarding successful contact-rich motion generation for robotic manipulators when manipulating objects. Existing approaches to contact-rich motion generation depend on assumptions that help planning in simulation but make execution on real systems fail. The project involves collaboration between the UK PI and a Co-PI At Boise State University. The project involves a combination of computational work and experimental work. The computational work involves different tasks. These include 1) searching for feasible plans to manipulate an object even as it changes the number of contacts with the manipulator and environment, 2) searching for controllers that stabilize the plan, and 3) using Bayesian learning to learn controllers that are robust to model uncertainty, starting from the stabilizing controller. The experimental work consists of a robot arm manipulating a box-like object in a cluttered environment. The experimental work will be conducted by the UK PI.