Graph-Enhanced Proximal Policy Optimization for Simulation-Based Point Cloud Coverage Planning on Curved Surfaces

Vollständiger Abstract

Worum geht es in dieser Arbeit?

This study examines the effects of graph aggregation and candidate-level scoring on PPO-based coverage planning for simulated curved point clouds. In a static uniform-coverage benchmark, a parameter-matched global MLP reaches 22.13% overall coverage, GCN-PPO reaches 76.97%, and Graph-Enhanced PPO reaches 88.37%; DFS remains the strongest at 95.80%. A separate 50-decision task introduces nonuniform priorities and motion costs. In that setting, GCN-PPO and Graph-Enhanced PPO obtain nearly identical weighted coverage (0.812 and 0.811, respectively), and no statistically reliable difference is detected. On a separate 1000-node scan-derived Stanford Bunny surface with retained holes and mesh-geodesic connectivity, GCN-PPO and Graph-Enhanced PPO again show comparable weighted coverage (0.450 and 0.445, respectively) and task utility (0.321 and 0.315, respectively), with no statistically reliable difference. Across the physically normalized zero-shot density and area tests, the direction of the small mean differences varies by condition, and none remain significant after multiplicity correction; both policies lose performance as the sampling density or workspace size changes. Direct adaptation on the known fivefold-area target graphs raises Graph-Enhanced PPO coverage from 0.631 to 0.753 for held-out starts. Because adaptation and evaluation use the same target graph instances, unseen large-graph generalization is not tested. Across the experiments, graph aggregation provides the clearest learned improvement. The auxiliary scorer improves the static benchmark result, but the added parameters prevent assigning that gain solely to a separate scoring mechanism; no consistent benefit appears in the other tests. Systematic traversal remains preferable when exhaustive uniform coverage is feasible. The study remains simulation-based: scan-derived geometry is evaluated, but validation on physical hardware is still required.