Combining declarative, procedural, and predictive knowledge to generate, execute, and optimize robot plans

摘要

One of the main challenges in motor control is expressing high-level goals in terms of low-level actions. To do so effectively, motor control systems must reason about actions at different levels of abstraction. Grounding high-level plans in low-level actions is essential semantic knowledge for plan-based control of real robots. We present a robot control system that uses declarative, procedural and predictive knowledge to generate, execute and optimize plans. Declarative knowledge is represented in PDDL, durative actions constitute procedural knowledge, and predictive knowledge is learned by observing action executions. We demonstrate how learned predictive knowledge enables robots to autonomously optimize plan execution with respect to execution duration and robustness in real-time. The approach is evaluated in two different robotic domains.