Hai Yang

The software and data in this repository are a snapshot of the software and data that were used in the research reported on in the paper Urban Delivery with a Coordinated Fleet of Trucks and Drones by W. Zeng, H. Yang, H. Wang, and M. Qi.

Replication package for COMMTR-2026-0064.R2 ("Agent-Pilot: A Visual Hybrid Intelligence Closed-Loop Framework for Supervised Autonomous eVTOL Landing"). Completed: 160 landing trials (16 scenarios x 10), the 2,240-case vision-language safety benchmark with ground truth, 1,600 input images, two full prediction sets and metrics, 400 API-versus-local latency pairs, 65 human-pilot records, the raw OFAT PID sweep (13 configurations x 10 repetitions per axis), and twelve flights' telemetry with 221 board-corner annotations and 137 masks. All source code and the simulator-side MSFS 2024 developer kit are included. Run python repro/run_all.py to recompute 226 manuscript cells. The Replication Explanatory File is included. Raw videos (6.4 GiB) and the 1,600 lossless PNGs (3.6 GiB) exceed the 3 GB limit and are hosted at https://huggingface.co/datasets/yuxindu/agent-pilot-open with checksums in MEDIA_EXCLUDED.md. License: MIT; third-party components keep Apache-2.0.

The rapid growth in demand for unmanned aerial vehicle (UAV) operations requires effective allocation of limited low-altitude airspace under dynamically arriving requests. When operating at scale, thousands of UAVs simultaneously compete for access to shared airspace resources, and capacity allocation becomes a central challenge for maintaining efficiency and safety. We study a real-time route allocation problem in which heterogeneous UAV requests compete for capacity on a spatiotemporal low-altitude airspace network, with the objective of maximizing social welfare while ensuring conflict-free operations under link-time capacity constraints. We formulate the problem as an online combinatorial allocation model and develop a pricing-based primal-dual algorithm that assigns route access using dynamically updated link-time unit price. In contrast to conventional ground and airspace traffic management rules, such as first-come-first-served and greedy admission rules, the proposed method assigns link-time unit price based on capacity usage, so the route access is granted only when the request's reported valuation exceeds the total price of the route. The method is computationally efficient and admits a provable performance guarantee relative to the offline optimum. Numerical experiments show that the proposed method achieves near-optimal system performance at scale and significantly outperforms conventional benchmark policies. Using large-scale delivery data from China, we show that coordinated UAV deployment reduces delivery times by 77.7% on average, with reductions exceeding 90.0% in first-and last-mile operations, and improves accessibility to time-sensitive services in remote and mountainous regions. These results demonstrate that capacity-aware, market-based allocation provides a tractable and effective framework for managing UAV traffic in low-altitude airspace.

Crowd-sourced, randomly located vehicles in large-scale on-demand mobility systems result in uncertainty in customer waiting times, posing challenges for service reliability and system efficiency. We theoretically establish that, under nearest-eligible-vehicle assignment and the stated independence condition, any nondegenerate limit of suitably normalized customer waiting-time minima must belong to the minimum-GEV family. Under a lower-bounded spatial Poisson benchmark, we derive how the cumulative spatial intensity of eligible idle vehicles determines the minimum-GEV location, scale, and shape parameters. Numerical simulations show close finite-sample agreement between fitted minimum-GEV distributions and simulated waiting-time distributions under the specified designs, including expected eligible-set sizes as low as ten and random selection among the first three order statistics. Observations from e-hailing trajectory data provide descriptive distributional evidence that GEV distributions approximate customer waiting times. We also formulate a generalized-cost comparison across providers. A lower-bounded observed-quote benchmark links provider quotes to physical pickup-time distributions, while a behavioral benchmark uses independent minimum-Gumbel generalized-cost disturbances with a common scale and yields the Multinomial Logit (MNL) form. Moreover, we analytically derive a unique stationary throughput turning point associated with the Wild Goose Chase (WGC) phenomenon under an uncontrolled, lower-bounded benchmark. The resulting turning point provides a stress-test benchmark for examining how assignment, admission, dispatch limits, and repositioning alter stationary throughput.

Urban air mobility (UAM) can bypass congested ground corridors and provide substantial travel-time savings even before its supporting vertiport network is sufficiently dense to minimize ground access and egress. We study this mismatch between service attractiveness and infrastructure maturity in a setting where ride-sourcing provides both the direct ground alternative and UAM feeder service. Passengers differ in direct ride-sourcing travel time, so UAM selectively attracts travelers facing relatively time-consuming ground trips. We develop a two-stage model in which vertiport density is chosen before fares and passenger mode choice and compare decentralized air-ground provision with platform integration and a no-UAM benchmark. When pickup and repositioning are sufficiently efficient, UAM reduces ride-sourcing vehicle-hours traveled even with sparse vertiport coverage. Its effect on vehicle-kilometers traveled (VKT), however, changes sign at a unique vertiport-density threshold: below the threshold, long feeder movements increase VKT; above it, UAM reduces VKT. At a fixed network, integration lowers the door-to-door UAM fare and magnifies the prevailing VKT effect by shifting more passengers toward UAM. When vertiport investment is endogenous, integration also raises the private return to coverage and induces a denser network. For intermediate deployment costs, only the integrated network crosses the VKT threshold, reversing the sign of UAM's mileage effect relative to decentralized operation. A New York City application exhibits the same reversal: integration increases VKT on a fixed two-vertiport network but makes a third vertiport privately worthwhile, yielding lower VKT once infrastructure responds. Deployment support and a UAM fare instrument address the remaining coverage and passenger-allocation gaps. These results show that the transportation effect of platform integration cannot be assessed independently of the infrastructure investment incentives it creates.

Emerging mobility services such as Urban Air Mobility (UAM) complicate mode-choice prediction because travelers may position them differently relative to existing modes. Conventional econometric and machine-learning models mainly capture substitution indirectly through error distributions or population-level structures. They are not designed to flexibly capture traveler-specific pairwise substitution patterns across multiple modes. This study proposes a Persona-Conditioned Graph Neural Network (PC-GNN) that learns individual substitution structures with nonlinear preferences. Because such structures are difficult to identify from sparse stated-choice observations alone, large language models (LLMs) are introduced to elicit diverse priors from a subset of travelers and store them in a reusable belief bank. The priors contain persona vectors and substitutability matrices, representing the individual substitution structures of these travelers. For the remaining travelers, observed attributes are combined with retrieved LLM-elicited priors to construct multiple conditioned GNNs, and a retrieval-mixture module then produces the final prediction. We further relate PC-GNN to established discrete choice models and derive an approximation bound to characterize this retrieval-mixture mechanism. Using a stated-preference survey of nearly 1,600 respondents in the Guangdong-Hong Kong-Macao Greater Bay Area, the analysis reveals heterogeneity in UAM substitution. High-speed rail (HSR) emerges as its primary substitute, while car-oriented travelers exhibit a substitution pattern involving both HSR and private cars. PC-GNN also improves the mode-choice predictive accuracy by about 13% over second-best baselines. Ablation and scalability analyses further confirm the value of persona augmentation, individual-specific graphs, and behaviorally informed retrieval.

This paper studies a periodic permit-based driving restriction scheme in which commuters receive a finite number of non-tradable driving permits over a planning horizon. Driving consumes one permit, while taking public transit preserves the remaining permit balance. Commuters face uncertain daily travel needs and differ in the additional disutility of using transit when urgent needs arise. To capture the resulting intertemporal tradeoff, we formulate an individual dynamic logit model under a conjectured congestion path and embed the induced responses in a mean-field user equilibrium. The analysis identifies the dynamic opportunity cost of permit use as the key mechanism linking current driving decisions with future mobility flexibility. We establish structural properties of this opportunity cost and characterize how commuter heterogeneity affects driving probabilities through the interaction between urgent-state driving incentives and regular-state permit preservation. The analysis further shows that permit budgeting varies with the urgency environment and commuter heterogeneity, thereby clarifying who budgets and when in a periodic driving permit system. Numerical experiments then show that the timing, intensity, and persistence of urgent travel needs reshape equilibrium permit use and the resulting aggregate driving patterns. Compared with a rigid benchmark, the proposed scheme reallocates driving access toward more valuable days and reduces urgent-state transit burden. Permit endowment has a nonmonotone effect on system performance, highlighting the need to balance mobility flexibility and congestion control.

This paper studies a periodic permit-based driving restriction scheme in which commuters receive a finite number of non-tradable driving permits over a planning horizon. Driving consumes one permit, while taking public transit preserves the remaining permit balance. Commuters face uncertain daily travel needs and differ in the additional disutility of using transit when urgent needs arise. To capture the resulting intertemporal tradeoff, we formulate an individual dynamic logit model under a conjectured congestion path and embed the induced responses in a mean-field user equilibrium. The analysis identifies the dynamic opportunity cost of permit use as the key mechanism linking current driving decisions with future mobility flexibility. We establish structural properties of this opportunity cost and characterize how commuter heterogeneity affects driving probabilities through the interaction between urgent day driving incentives and regular day permit preservation. The analysis further shows that permit budgeting varies with the urgency environment and commuter heterogeneity, thereby clarifying who budgets and when in a periodic driving permit system. Numerical experiments then show that the timing, intensity, and persistence of urgent travel needs reshape equilibrium permit use and the resulting aggregate driving patterns. Compared with a rigid benchmark, the proposed scheme reallocates driving access toward more valuable days and reduces urgent-state transit burden. Permit endowment has a nonmonotone effect on system performance, highlighting the need to balance mobility flexibility and congestion control.