NRI: Collaborative Research: Accelerating Robotic Manipulation with Data-Enhanced Contact Mechanics

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

Robotic manipulation depends upon mechanical contact between robot and object. A better understanding of mechanical contact enables a wider range of more flexible manipulation techniques, which in turn enables the applications of greatest societal benefit such as eldercare, disaster response, or surgery. This project is developing a broader and more accurate understanding of frictional contact, using a fusion of physics and data. The project combines recent advances in a physics-based understanding of frictional contact with new machine learning techniques applied to a large corpus of experimental data. One operation of great interest is manipulation of an object held in the robot gripper, even when the gripper is very simple. Other operations of interest are handling objects in clutter, and manipulation of flexible objects, such as clothing. The project is attacking several central challenges: modeling frictional contact, modeling deformation, measuring small motions and interaction forces, gathering large amounts of data, and developing techniques for learning in a closed-loop system. Parametric and semi-parametric models enable the project to apply engineering models enhanced with observation data, for both planning and control. New machine learning techniques such as predictive state representations (PSRs) enable identification and modeling of previously hidden state, as well as learning in closed-loop systems. New infrastructure enables gathering of relevant, precise data, on a large scale. The project is developing and employing a Robotic Manipulation Arena, with a unique combination of manipulation resources and instrumentation to provide high volumes of high quality experimental data. The primary outcomes are robust and practical contact models, so that robots can work more dexterously and opportunistically.