BMI-based framework for teaching and evaluating robot skills

Brain Machine Interface systems provide ways of communication and control of a variety of devices that range from domestic appliances to humanoid robots. Most BMI systems are designed exclusively to control devices using low-level commands, or high-level commands when devices with pre-programmed functionalities are available. In this paper, we build on our previous work on BMI-based Learning System in which we presented a different approach for designing BMI systems that incorporate learning capabilities that relieve the user from tedious low-level control. In this work, we extend the capabilities of our framework to allow a user to be able to teach and evaluate a robotic system by using a BMI. We provide general system architecture and demonstrate its applicability in new domains such as teaching a humanoid robot object manipulation skills and evaluating its performance. Our approach consists of 1) tele-operating robot´s actions while robot´s camera collects object´s visual properties, 2) learning manipulation skills (i.e. push-left, lift-up, etc.) by approximating a posterior probability of commonly performed actions when observing similar properties, and 3) evaluating robot´s performance by considering brain-based error perception of the human while he/she passively observes the robot performing the learned skill. This technique consists of monitoring EEG signals to detect a brain potential called error related negativity (ERN) that spontaneously occurs when the user perceives an error made by the robot. By using human error perception, we demonstrate that it is possible to evaluate robot actions and provide feedback to improve its learning performance. We present results from five human subjects who successfully used our framework to teach a humanoid robot how to manipulate diverse objects, and evaluate robot skills by visual observation.