Physics-, social-, and capability- based reasoning for robotic manipulation

Abstract

Robots that can function in human-centric domains have the potential to help humans with the chores of everyday life. Moreover, dexterous robots with the ability to reason about the maneuvers they execute for manipulation tasks can function more autonomously and intelligently. This thesis outlines the development of a reasoning architecture that uses physics-, social-, and agent capability-based knowledge to generate manipulation strategies that a dexterous robot can implement in the physical world. The reasoning system learns object affordances through a combination of observations from human interactions, explicit rules and constraints imposed on the system, and hardcoded physics-based logic. Observations from humans performing manipulation tasks are also used to develop a unique manipulation repertoire suitable for the robot. The system then uses Bayesian Networks to probabilistically determine the best manipulation strategies for the robot to execute on new objects. The robot leverages this knowledge during experimental trials where manipulation strategies suggested by the reasoning architecture are shown to perform well in new manipulation environments.