Correction of Robot Behavior based on Brain State Analysis

Abstract :

A challenge in adaptive systems is self-monitoring of their own performance to self-correct erroneous behavior. Learning models used for self-adaptation of system's behaviors can be improved by using external evaluation, e.g., using error related potentials (ErrPs) measured on a human evaluator. In the proposed approach, the robot's behaviors are adapted based on the combined use of reinforcement learning (RL) model and singletrial detection of ErrPs. Here, ErrPs are used as feedback for the RL model. In the previous study, we showed that single-trial detection of ErrPs is feasible in a realistic scenario (Kim and Kirchner, 2013). The goal of a planned study is to improve performance of the robot's behavior by using the proposed combined approach (i.e., the combined use of RL model and single-trial detection of ErrPs). To this end, the concept for the scenario was developed, which allows us to correct robot's wrong actions by using ErrPs as feedback for the RL model. The next step is to prove the concepts of the proposed approach by using the developed scenario concept.