Deep Reinforcement Learning for Robot Navigation in Unstructured Environments

This thesis is focused on deep reinforcement learning for mobile robot navigation in unstructured environments. Autonomous navigation has been mainly tackled through map-based algorithms, which are ineffective in many situations like exploration or rescue missions. For mapless scenarios, the simultaneous localization and planning (SLAM) represents a cornerstone on which a wide variety of algorithms are built. However the difficulty of maintaining a map from sensory inputs, typical of these methods, is leading the research community to look for alternatives. Deep reinforcement learning aims at solving the autonomous navigation problem end-to-end, by directly mapping high-dimensional inputs to actions, without the need for a model of the environment. In this thesis, a model-free reinforcement learning approach is adopted: a variant of the tabular Q-learning algorithm, called deep Q-learning, uses a deep neural network to approximate the action-value function and to map states into velocity commands, without the need of an expert or supervisor. The learning model is trained in simulation on TurtleBot3 and Curiosity mobile robots in two different environments. After that, the neural network trained on TurtleBot3 is transferred on Curiosity and then fine-tuned on new navigation environments. The results are then compared to those obtained by training the model from scratch, with random initialization of the parameters: this comparison shows how, thanks to the pre-training, the rover manages to reach on average a higher number of targets per episode throughout the entire simulation.