Q-Learning in an Autonomous Rover

Robotics researchers often require inexpensive hardware that is freely distributable to the public domain in large numbers, yet reliable enough for use in different applications without fear of the hardware itself becoming a burden. In the past, researchers have moved towards robot simulations, in favor of the lack of hardware and ease of replication. In this paper we introduce an implementation of Q-Learning, a reinforcement learning technique, as a case study for a new open-source robotics platform, the Brookstone Rover 2.0. We utilize a Theano-based implementation of Google DeepMind’s Deep Q-Learning algorithm, as well as OpenCV for the purpose of state-reduction, and determine its effectiveness in our rovers with a color-seeking “rover-ina- box” task.