From Understanding to Deployment: A Multi-Perspective Study of Sidewalk Delivery Robot Systems

Abstract

Local delivery demand often involves small parcels transported over short distances, yet these deliveries are frequently made by road vehicles designed for heavier loads and higher-speed travel. This mismatch can occupy valuable road and curb space while requiring vehicle and energy resources that are disproportionate to the payload. Compact, low-speed electric sidewalk delivery robots (SDRs) offer a pedestrian-scale alternative for appropriate short-distance deliveries. However, shifting delivery activity to sidewalk networks raises new social, strategic, and operational questions.SDR deployment is not a single optimization problem but a set of decisions made at different levels and times. Public-Facing Deployment addresses public acceptance of SDRs and concerns about their use of shared sidewalks. Strategic Resource Planning determines how many robots and swappable batteries are needed before demand is known. Dynamic Decision Diagnosis examines assignments made as new requests arrive while robots carry out earlier plans. Bounded Online Intervention asks whether adaptive decision support can improve those assignments while deterministic software checks feasibility and controls execution. These four layers organize the studies that follow.The first study examines public attitudes and concerns through unsolicited online discussion. Keyword searches identify SDR-related YouTube videos, and manual screening retains videos that show SDRs in public spaces or explicitly discuss them. Screening the comments for relevance yields 4,971 comments. After manual sentiment labeling, supervised classifiers characterize negative, positive, and neutral reactions, while Latent Dirichlet Allocation identifies ten topics, including employment displacement, food and cargo security, vandalism, pedestrian interaction, and safety. The findings describe attitudes and concerns expressed by this self-selected group of YouTube users.The second study plans fleet size and battery inventory using a two-stage stochastic model. The first stage chooses the number of robots and additional batteries before demand is known. The second solves an SDR pickup-and-delivery problem for each demand scenario, subject to precedence, time-window, capacity, battery, and swapping constraints. A continuous approximation estimates expected system cost across resource plans, while a customized adaptive large neighborhood search constructs scenario routes. In the Purdue campus case, higher demand generally calls for a larger fleet, while battery range affects the additional battery inventory needed. Demand varies across scenarios, but each routing problem uses fixed resources and a complete request set.The third study moves from static planning to dynamic operation. It develops an event-driven pickup-and-delivery simulator in which requests arrive while robots execute earlier assignments and parts of their routes remain committed. Matched branch replay reproduces the state before a selected assignment and tests alternative request-to-vehicle mappings. Each mapping uses the same deterministic route constructor, subsequent requests, and later routing heuristic. Across 200 cases, 17,512 reruns of 798 selected decisions show that another tested mapping achieves a better terminal outcome than the original heuristic's reconstructed assignment at 49.5–66.9% of selected decisions. Another tested mapping also outperforms the candidate with the lowest current-state objective at 51.0–60.8%. These comparisons reveal later differences in service lateness, operating time, travel, and the distribution of remaining work.The fourth study evaluates a bounded language-model assigner for the same dynamic problem. With n requests whose assignments can still change and K robots, there are up to K^n mappings before feasibility constraints are applied. The model proposes complete assignments using information available at the decision point. Deterministic software constructs routes, checks feasibility, rechecks the updated state, applies fallback when needed, and controls execution. Across 200 dynamic cases, the bounded assigner reduces service lateness by 88.03% relative to the strongest action-matched rule and increases operating time by 2.86%, while the aggregate objective differs by only -0.089%. The results show a service–efficiency tradeoff under the tested conditions, not overall objective superiority.Three conclusions emerge across the studies. First, deployment quality is multidimensional: public response, resource adequacy, service, efficiency, and operational control require different evidence and should remain visible in evaluation. Second, strategic capacity and dynamic adaptability are complementary because resources are committed before service while assignments respond to information revealed during service. Third, adaptive support can be evaluated more transparently when its information, action, timing, verification, fallback, and execution boundaries are explicit. The dissertation therefore contributes a layered evidence-and-decision framework for understanding, planning, and operating SDR systems as they move from static design assumptions toward dynamic service.