Abstract
Learning-enabled control requires training to produce competent policies consistently when success requires every robot to finish. This study evaluates masked entity-attention multi-agent proximal policy optimization (MAPPO) for simulated landmark coverage by five differential-drive robots. Configurations were compared under a fixed reward, optimizer, stopping rule, and 100 held-out scenarios, using trained seeds as statistical units; the actor observes the robot–landmark scene but not the externally computed assignment. Across ten policies per principal configuration, masked entity attention achieved 94.6±4.7% deterministic full-team success, against 9.3±12.4% and 6.8±5.1% for wider (1.21 times its parameters) and ordinary fixed-index controls, every masked-attention seed exceeding every seed of both fixed-index controls. Near-parameter-matched Deep Sets and unmasked-attention controls reached 73.7±14.0% and 72.1±13.1%, overlapping it. The same policies retained 91.4–96.3% mean success from three to eight robots without retraining, and masked attention remained separated from both fixed-index controls at every rate in a 45-policy learning-rate design. Held-out scenes come from the training distribution; under shift, success fell to 70.9% in an unseen larger workspace, while landmark, sensing, and actuation perturbations kept it above 93.3%. Masked entity-attention MAPPO gives the most consistent full-team completion among the tested designs; broader robustness and physical deployment remain unestablished.