Learning Mobile Robot Control for Obstacle Avoidance Based on Motion Energy Neurons

Image motion due to self motion is an important cue biological systems use for gathering information about the environment. The motion energy model is commonly used to model the responses of motion selective neurons in the mammalian primary visual cortex. Here, we investigate the hypothesis that these low level responses are directly useful for navigation. This avoids the need for estimating a model of the environment and the delay incurred in computing it. In order to discover the relationship between the neuron responses and the motor control required to avoid obstacles in the environment, we use reinforcement learning to train a robot equipped with infrared depth sensors to avoid objects using the outputs of simulated motion energy neurons by minimizing the long term average of the infrared signals it receives. Our experiments with a Koala robot indicate that the motion energy neuron outputs can effectively trigger obstacle avoidance motions in advance of those triggered by the infrared sensors.