Modeling and Learning Robot Manipulation Strategies

This paper describes a general approach to learning and planning robot manipulation strategies. Here, the strategies are represented using a discrete-event dynamical systems model where each node corresponds to a state in the robot task environment that triggers certain action schemata and each arc corresponds to a plausible action that brings the task environment into a new state. With such a representation, a manipulation strategy plan can be derived by searching a connected state transition path that is the most reliable. Here, we define the notion of reliability in terms of the estimated chance of success in reaching a desirable state. In the paper, we first present the formalism of discrete-event dynamical system in the context of robot manipulation tasks. Throughout the paper, we provide both illustrative and experimental examples to demonstrate the proposed approach.