A Constraint-Aware Proximal Policy Optimization Method for Task Scheduling and Dynamic Reconfiguration of Maritime Unmanned Systems

To address slow dynamic-reconfiguration response and difficulty in satisfying hard constraints in the task scheduling of heterogeneous maritime unmanned platforms, this paper proposes a constraint-aware proximal policy optimization (CPPO). CPPO models task scheduling as a Markov decision process; applies a three-layer action mask filtering invalid actions by platform survival and capacity, communication-link reachability, and payload capability matching to guarantee hard constraints; and introduces a staged penalty adjustment strategy to balance exploration and constraint internalization. In eight scenarios spanning 14 to 100 platforms under static and platform-failure settings, CPPO is benchmarked against Greedy, improved NSGA-II, and LSTM-PPO. Ablations show the mask is the dominant factor: it raises the task completion rate from about 25% to 53–97% and the constraint satisfaction rate from about 0.1 to 0.65–1.0, while the staged penalty adds gains. Compared with Greedy, CPPO shows a statistically equivalent completion rate (p > 0.05) and a significantly higher constraint satisfaction rate in small-scale scenarios (p = 0.006), while its inference latency remains about 1 ms, 2–3 orders of magnitude faster. Compared with LSTM-PPO, recurrent memory brings no gain in quality or constraint satisfaction, while training time increases 2–7-fold, validating the lightweight MLP encoder. CPPO provides an “offline training–online inference” solution for maritime unmanned system reconfiguration.