SegVisRL: development of a robot's neural visuomotor and planning system for lunar exploration

Abstract

In this paper, we present an approach for the learning of a visuomotor system for a robotic rover using reinforcement learning (RL) within a simulation that combines both proprioceptive and exteroceptive information. With a rising interest in private lunar exploration, compact, power-efficient and cost-efficient rover concepts are becoming numerous. Many of these still use LIDARs for obstacle avoidance, however, using a camera could help achieve a more optimal system. Visual data contains structured information that can be used for understanding texture, localization and object recognition. That said, it is hard to use RL with raw high-resolution visual data due to the dimensional size, with over 6 million inputs in an image alone. By preprocessing the images to be segmented and down sampling to reduce size, we are able to achieve stable learning, and to teach a robot to understand footage to avoid obstacles and reach its goals. This method combines exteroceptive data from the camera and the goal location information along with proprioceptive information, such as the robot”s angular rotation. Moreover, whereas most approaches feed the RL output to a motion controller or trajectory generator, our RL agent is able to directly control the actuator outputs of the rover.