Autonomous Berthing in Mapless Environments via Soft- and Hard-Constrained Reinforcement Learning

As pivotal mobile nodes within the Marine Internet of Things (MIoT), Unmanned Surface Vessels (USVs) rely on autonomous berthing as a critical service to ensure persistent data offloading and energy replenishment. However, this maneuver demands precise navigation within confined, high-risk environments, where traditional edge control methods often lack robust safety guarantees. To address these challenges, this paper presents a hybrid safety-aware reinforcement learning (RL) framework tailored for autonomous berthing. By directly processing raw LiDAR data, the proposed end-to-end policy eliminates the reliance on a priori global maps, thereby enhancing adaptability in unstructured environments. To overcome the inherent safety limitations of standard RL, we first introduce the Soft-Constrained Policy Optimization (SCPO) paradigm. Incor-porating a safety critic and a Lagrangian dual update mechanism, SCPO guides the agent to internalize safety constraints during the learning process. Furthermore, to bridge the gap between discrete decision-making and the continuous safety assurance required for docking, we propose a Hard-Constrained Safety Filter (HCSF). Grounded in Control Barrier Functions (CBFs), this module enforces safety constraints via real-time Quadratic Programming (QP) corrections on control inputs. Simulation results demonstrate that the proposed framework exhibits superior safety performance compared to existing baselines.