Addressing Multi-Objective and domain adaptation challenges in Reinforcement Learning through case studies in multi-agent navigation and visual servoing

Reinforcement learning (RL) has gained a lot of attention in recent years due to its potential to solve complex control problems. However, RL faces various challenges in multi-agent settings and high-dimensional action and observation spaces. This thesis addresses some of these challenges through two case studies: Multi-Agent Reinforcement Learning and Robotic Arm Manipulation. The first case study focuses on multi-agent navigation and proposes a novel class of RL-based controllers called least-restrictive controllers for multi-agent collision avoidance problems. This study aims to implement a high-level safe RL policy that provides safe navigation for multi-agent navigation in a shared environment with or without static obstacles. The proposed policy works in different tasks containing a different number of agents and different task objectives. The second case study proposes a novel visual servoing (VS) algorithm using sequential stochastic latent actor-critic and reinforcement learning. This study aims to overcome domain adaptation and control challenges in high-dimensional action and observation spaces. The proposed algorithm can adapt to a real robot through only single-shot transfer learning in representation learning parts.