Point Cloud-Based Reinforcement Learning for Autonomous Navigation of a Robotic Rover on Planetary Surfaces

This work proposes an autonomous navigation software stack to allow robotic rovers in open-field environments to reach given target coordinates in a safe, reliable, and efficient manner. The primary objective of this project is to train a Reinforcement Learning (RL) agent to acquire effective strategies for safely traversing its sorroundings until the desired destination is reached. The rover receives target coordinates from the user and point cloud data from dedicated sensors such as stereocameras or LIDARs, providing information about the surrounding environment morphology. Assuming the robotic rover has a reliable method to localize its position and orientation in the field and track its movements during operations, as well as a robust controller to reach a target pose in its immediate proximity within a few meters from its current position, we introduce DIANA-Gym, a simulation environment to train the agent on a virtual rover. Next, a state-of-the-art algorithm is selected to train and test an RL agent on the virtual environment. Although designed for planetary navigation and exploration purposes, we believe that the proposed framework could be adapted with minimal modifications to other similar open-field navigation tasks. By combining Reinforcement Learning with point cloud data, our proposed autonomous navigation software stack provides an efficient, reliable, and safe solution for autonomous exploration and navigation in challenging environments. The entire project is being developed within DIANA from Politecnico di Torino, a student team competing in the Rover Challenge Series, which challenges students from all the engineering areas to design and develop a prototype for an astronaut assistance rover platform.