Prediction Driven Behavior: Learning Predictions that Drive Fixed Responses

摘要

We introduce a new method for robot control that combines prediction learning with a fixed, crafted response---the robot learns to make a temporally-extended prediction during its normal operation, and the prediction is used to select actions as part of a fixed behavioral response. Our method is inspired by Pavlovian conditioning experiments in which an animal's behavior adapts as it learns to predict an event. Surprisingly the animal's behavior changes even in the absence of any benefit to the animal (i.e. the animal is not modifying its behavior to maximize reward). Our method for robot control combines a fixed response with online prediction learning, thereby producing an adaptive behavior. This method is different from standard non-adaptive control methods and also from adaptive reward-maximizing control methods. We show that this method improves upon the performance of two reactive controls, with visible benefits within 2.5 minutes of real-time learning on the robot. In the first experiment, the robot turns off its motors when it predicts a future over-current condition, which reduces the time spent in unsafe over-current conditions and improves efficiency. In the second experiment, the robot starts to move when it predicts a human-issued request, which reduces the apparent latency of the human-robot interface.