Deep Reinforcement Learning Control of an Autonomous Wheeled Robot in a Challenge Task: Combined Visual and Dynamics Sensoring

This paper presents a Deep Reinforcement Learning agent for a 4-wheeled rover in a multi-goal competition task, under the influence of noisy GPS measurements. A previous related work has implemented a similar agent to the same task using only the raw dynamics measurements as observations. The Proximal Policy Optimization algorithm combined to Universal Value Function Approximators resulted in a system able to successfully overcome very noisy GPS observations and complete the challenge task. This work introduced a frontal camera to add visual input to the rover observations during the task execution. The main change on the algorithm is on the neural networks' architectures, in which a second input layer was added to deal with the image observations. In a few alternate versions of the networks, Long Short-Term Memory (LSTM) cells were included in the architecture as well. The addition of the camera did not present a significant increase in stability or performance of the network, and the computation time require increased.