Where should the last passenger be dropped off? Anticipatory walking in ridepooling

• In ride-pooling, walking is beneficial, but often drop-offs are ignored due to the lack of future information. • Anticipatory walking method uses artificial stops sampled from demand patterns to optimize last-stop drop-offs and position vehicles for future requests. • Hierarchy-aware search prioritizes well-connected nodes to maintain robustness when demand predictions are imperfect. • Reverting mechanism maintains dual schedules and ensures decisions never worsen passenger experience compared to no-walking alternatives. • Real-world experiments show consistent rejection rate reductions, capturing 29-73% of Dynamic Walking’s theoretical gains while remaining practically deployable. Ridepooling can gain significant efficiency by going beyond the door-to-door paradigm, i.e., including some short passenger walks. Current methods typically rely on Static Walking, where the drop-off point is fixed at the moment of assignment. This frequently results in walking only at pick-up, since drop-offs may occur at the end of the current partial schedule of the vehicle, when the next stop of the vehicle is still unknown. In contrast, Dynamic Walking, where the drop-off point can be updated while the passenger travels, is generally considered unacceptable to users due to the uncertainty of both ETA and actual alighting location caused by it. Therefore, we propose an Anticipatory Walking method that captures most of the improvements yielded by Dynamic Walking, while remaining as practical as Static Walking. First, we generate an Artificial Stop sampled from two different sources of demand, which is placed in the schedule immediately after the drop-off to be decided. We then combine this a method that computes nodes’ hierarchies, to optimize the exact drop-off location. To prevent cases in which the prediction would result harmful once the actual next stop is revealed, we maintain an alternative schedule where the passenger is dropped at their exact destination, and select optimally between the two options. Large-scale simulations in three real-world instances (Manhattan, Utrecht, and Canberra) show consistent improvements over the Static Walking baseline and performance close to a Dynamic Walking benchmark. Our method achieves 29% of the improvements of Dynamic Walking in terms of rejection rate in Manhattan, 70% of it in Utrecht, and 73% of it in Canberra, without introducing unacceptable uncertainty and unreliability for users. These patterns align with network structure and level of sharing: gains are greater in cities with more hierarchical networks and less-shared demand-supply patterns, such as Utrecht and Canberra. In all, our Anticipatory Walking captures much of the benefits of Dynamic Walking, without imposing additional uncertainty on the users, and without the need for extensive forecasting or continual re-optimization.