This project is based on two theses: (1) autonomous robotic manipulation requires a large repertoire of actions; and (2) autonomous manipulation does not decouple into separate arm and hand functions. The project's goal is to establish the principles and techniques to endow a robot with a large repertoire of manipulative actions, many of them involving an intimate coordination of arm and hand. The approach is to develop these principles and techniques using physics-based models as well as machine learning of empirical models. The project is developing and testing these principles and techniques by attacking several challenge tasks. The work is organized to address three primary challenges: (1) identifying actions; (2) modeling actions; and (3) orchestrating actions. For the first challenge, identifying actions, the project is adapting previous physics-based manipulation research, along with ordinary robotic engineering of behaviors inspired by humans and existing robotic systems. For the second challenge, modeling actions, the project is augmenting physics-based models work with empirical stochastic modeling drawn from previous applications of machine learning. For the third challenge, orchestrating actions, the project is adapting previous work on sensor-based planning and control. The project's expected broader impacts includes more capable robots, which will simplify deployment, enable new applications and improve existing applications, which ultimately serves to improve productivity, services, and national economic competitiveness.
