VR simulation of mobile robots for social behaviour learning

In robotics, one of the main areas of interest for the research community is the creation of environments where machines and humans can safely coexist. This brings to two main issues: training and testing of robots in a controlled but realistic setting, and their correct navigation in social spaces, while causing no harm to people. Virtual reality allows the simulation of such environments, ensuring safety and allowing the training of robots in areas where errors are controlled and confined. This way, robot behaviour can be supervised and demonstrated by the operator, able to enter the training environment with the use of a Virtual Reality headset. This thesis explores this type of simulation by training a ROS-based four-wheel robot model tasked with navigating a laboratory-like area to reach a specified target, respecting social constraints. This process is carried on using human behaviour as a reference, implementing Imitation learning using Behavioural Cloning (BC) and Generative Adversarial Imitation Learning (GAIL), and confronting it with Reinforcement learning approaches, in particular Proximal Policy Optimization (PPO). Additionally, a hybrid learning pipeline that combines demonstrations with reinforcement learning is considered. Unity is used as the virtual development platform, alongside ML-Agents toolkit for Neural Network training. The objective is to assess which learning strategy achieves safer and more efficient navigation, and to evaluate the potential benefits of combining human demonstrations with autonomous exploration. The VR framework realized for this thesis can be adapted for future applications and different tasks, that can integrate human robot collaboration or more complex environments.