Towards Optimization of Deep Reinforcement Learning for Autonomous Exploration in Single and Multi-Robot Environments

Autonomous exploration of unknown environments stands as a critical challenge in contemporary robotics applications. The application of Deep Reinforcement Learning (DRL) in robotics exploration has been developed to design efficient exploration strategies suitable for complex real world scenarios. In the past there have been reviews about the application of DRL in exploration, but the optimization of DRL has not been extensively reviewed. This survey focuses on exploration for both single robot and multi-robot applications, specifically addressing path planning aspects such as navigation, coverage, and trajectory. The survey emphasizes key elements of DRL derived from previous work, encompassing state and observation, action, and reward. Finally, the main contribution of this survey is to review how past work has optimized the DRL implementation to overcome inherent limitations in the approach, including challenges related to inefficient exploration in the training process, which collectively limit the adaptability and robustness of these algorithms in truly unknown environments due to their resource-extensiveness. From our survey, current research is lacking on exploration with reasoning entities for the training, relying instead on prior environmental knowledge. This underscores the need for exploration strategies with built-in reasoning and less dependence on pre-existing data.