Automatic acquisition of task theories for robotic manipulation

A task theory is a collection of models that a robot uses to predict the effects of its actions. This thesis addresses how a robot can build a task theory from observed action effects, thus bypassing the usual human-programming method of defining "what actions do", and making robots considerably easier to train. The question that this thesis addresses is: "What are sufficient conditions whereby a robotic manipulation system can learn action models that allow automated planning and successful execution of manipulation strategies?" This thesis, by presenting an implemented learning robot system, shows that acquiring action models can be automated, and further shows that the required software mechanisms are simple and few. The important characteristics of the learning agent and its environment are: (1) the complexity of the task, (2) the sensory and effectory abilities of the agent, (3) the learning mechanism, (4) the planning mechanism, (5) the experimentation method, and (6) the prior knowledge of the task possessed by the agent. Sufficient conditions for success are demonstrated with empirical results in three manipulation tasks: parallel-jaw grasping, the peg-in-hole problem, and the tray-tilting problem. The successful learning robots for these tasks generate task theories consisting of a set of funnels. Each funnel is a kind of operator that maps a region of the task state-action space to a region of the state space. Funnels can be acquired for continuous tasks using very little prior knowledge of the task, but the learning mechanism must be robust in the face of noise and non-determinism. A simple search-based planner suffices for generating manipulation plans, although this planner must reason about the reliabilities associated with each funnel. After demonstrating sufficient conditions for successful acquisition of task theories, the thesis considers the necessity of the characteristics of the implemented learning robots. In addition to sensing, effecting, planning, and learning, the thesis develops requirements on the prior knowledge that must be possessed by the learning robot. Task-specific prior knowledge can be very weak, but for some complex tasks, prior knowledge is required to achieve acceptable convergence times for the learned task theory.