Evaluating Navigation Policies Under Controlled Variations in Perpendicular Stream Crossing

Navigation through moving traffic requires policies to identify safe crossing opportunities while remaining robust to changes in environmental conditions. We study this problem using a controlled two-dimensional benchmark for perpendicular stream crossing, in which an agent must traverse streams of moving vehicles through a common binary Stop/Move action space. We compare four decision mechanisms: an Always Move baseline, a task-specific Predictive Gap controller, Proximal Policy Optimization (PPO), and an adapted Socially Attentive Reinforcement Learning (SARL) policy. Learned policies are trained in a nominal single-stream environment and evaluated on disjoint scenarios under controlled changes in temporal headway, vehicle-motion dynamics, and the number of sequential traffic streams. We additionally evaluate variation across independently trained model seeds to distinguish sensitivity to environmental shifts from variability introduced during training. This benchmark provides a reproducible setting for comparing learned and task-specific navigation methods and for assessing robustness under controlled environmental shifts.