Author: Preston Davis Culbertson

Publisher:

Published: 2022

Total Pages: 0

ISBN-13:

While humans and other social animals (such as ants) can easily form teams to move heavy or bulky objects, robots struggle in collaborative manipulation tasks, especially under partial information about the environment or the object being transported. This thesis looks to enable flexible, scalable coordination in robot teams, looking toward a future where robots not only move objects together, but also work together to perform autonomous assembly of structures and manufactured goods. In Part I of this thesis, we investigate methods for collaborative manipulation and grasp synthesis under considerable uncertainty about the object's size and physical properties. Using tools from nonlinear control, we present a novel decentralized adaptive controller for collaborative manipulation that allows a team of robots to asymptotically track a desired trajectory in SE$(3)$. We also study the problem of synthesizing robust grasps of objects using only RGB images; we present a method which leverages a novel learned object representation to generate risk-sensitive grasps which can reason about the ambiguity inherent in the object shape. In Part II of this thesis, we turn our attention to the problem of multi-robot assembly planning. We show this problem can be posed as a mixed-integer linear program, which can be solved to global optimality using commercial solvers, and present effective heuristic strategies which can be computed quickly. Further, we present a method that uses supervised learning to accelerate the online solution of general mixed-integer convex programs using offline data. We show our method provides significant speedups over commercial solvers in a variety of robotics problems, including grasp selection and task allocation.