Abstract
Reinforcement Learning (RL) has shown remarkable success in solvingrelatively complex tasks, yet the deployment of RL systems in real-worldscenarios poses significant challenges related to safety and robustness. Thispaper aims to identify and further understand those challenges thorough theexploration of the main dimensions of the safe and robust RL landscape,encompassing algorithmic, ethical, and practical considerations. We conduct acomprehensive review of methodologies and open problems that summarizes theefforts in recent years to address the inherent risks associated with RLapplications. After discussing and proposing definitions for both safe and robust RL, thepaper categorizes existing research works into different algorithmic approachesthat enhance the safety and robustness of RL agents. We examine techniques suchas uncertainty estimation, optimisation methodologies, exploration-exploitationtrade-offs, and adversarial training. Environmental factors, includingsim-to-real transfer and domain adaptation, are also scrutinized to understandhow RL systems can adapt to diverse and dynamic surroundings. Moreover, humaninvolvement is an integral ingredient of the analysis, acknowledging the broadset of roles that humans can take in this context. Importantly, to aid practitioners in navigating the complexities of safe androbust RL implementation, this paper introduces a practical checklist derivedfrom the synthesized literature. The checklist encompasses critical aspects ofalgorithm design, training environment considerations, and ethical guidelines.It will serve as a resource for developers and policymakers alike to ensure theresponsible deployment of RL systems in many application domains.