The paper presents a new approach combining Maximum Entropy Deep Reinforcement Learning with a model-based control architecture to create an adaptive controller for autonomous underwater vehicles (AUVs). The approach involves a Sim-to-Real transfer strategy, including a bio-inspired experience replay mechanism, improved domain randomisation, and an evaluation protocol executed on a physical platform. The method learns proficient policies from suboptimal simulated models of the AUV and shows a control performance three times higher when transferred to a real-world vehicle.
Publication date: 18 Oct 2023
Project Page: https://www.sagepub.com/
Paper: https://arxiv.org/pdf/2310.11075
