Learning complex tasks using a stepwise approach

This paper explores a stepwise learning approach based on a system's decomp osition into functional subsystems. Two case studies are examined: a visual ly guided robot that learns to track a maneuvering object, and a robot that learns to use the information from a force sensor in order to put a peg in to a hole. These two applications show the features and advantages of the p roposed approach: i) the subsystems naturally arise as functional component s of the hardware and software; ii) these subsystems are building blocks of the robot behavior and can be combined in several ways for performing vari ous tasks; iii) this decomposition makes it easier to check the performance s and detect the cause of a malfunction; iv) only those subsystems for whic h a satisfactory solution is not available need to be learned; v) the strat egy proposed for coordinating the optimization of all subsystems ensures an improvement at the task-level; vi) the overall system's behavior is signif icantly improved by the stepwise learning approach.