EAGER: A Soft Contact Model for Simulation and Optimization

Roboti Llc, Redmond WA

Investigators

Abstract

Robots and humans interact with the world through physical contact. But since modern robots do not yet "understand" contact, they have very little capability to figure out how to do useful work. For example, robots today cannot figure out how to help a person with a sprained ankle climb a flight of stairs. This project will develop simulation models of contact that robots can use to "think" about useful work requiring contact. Through several mental experiments (simulations) a robot will be able to determine the best strategy to carry out whatever contact task it faces. These contact models will be computationally efficient, so the robot won't have to think too long, and realistic, so that the robot's plan has a high probability of success when executed. Since robots are relatively good at working with hard solid objects, this project focuses on soft bodies, such as people, food, and ropes. The resulting new algorithms will be incorporated in the publicly available physics simulator, MuJoCo, which can be embedded easily into the "brains" of future robots to expand the kinds of work robots will be able to do. The new soft contact model is derived from the Gauss principle of least constraint, applied to an augmented system that includes the rigid-body dynamics and the contact deformation dynamics. This yields a convex conic program in terms of the accelerations of the augmented system. The convex dual of this conic program is an optimization problem over contact forces. The project will develop this new theoretical framework, and explore its many properties including stability as well as computation of dynamics derivatives and their application to physically-consistent state estimation. Fast algorithms for computing both forward and inverse contact dynamics will also be developed. This will include a generalization to the projected Gauss-Seidel method that can handle conic constraints, as well as preconditioned conjugate gradient method exploiting the convexity and smoothness provided by the new framework.