Aug 2026· Frontiers in Built Environment· Vol 12· 0 citations· 43 references
TL;DR
PeopleFlow is introduced, a geometry-aware multi-agent framework that automatically converts architectural floor plans into simulation environments and applies force-driven pedestrian dynamics with congestion-aware routing, giving architects, safety engineers, and digital-twin practitioners a systematic basis for identifying and prioritizing safer design alternatives.
Abstract
Evaluating evacuation performance across building types requires methods that capture how spatial layout, pedestrian behavior, and operational disruptions interact under stress. Static compliance checks do not address this need, and most existing simulation tools require manual geometry preparation with limited support for reproducible cross-layout comparison. This paper introduces PeopleFlow, a geometry-aware multi-agent framework that automatically converts architectural floor plans into simulation environments and applies force-driven pedestrian dynamics with congestion-aware routing. A structured evaluation across 450 core simulations spanning nine layout types and five operational conditions reveals a 7.94-fold range in clearance time, driven primarily by geometric topology and routing policy rather than behavioral parameters alone. A further 210 supplementary runs confirm generalization to real-world floor plans; consistency checks against published evacuation drill ranges show full band overlap with a mean midpoint deviation of 0.83 s. The primary contribution is a reproducible comparative pipeline that couples automated floor-plan ingestion, scenario-controlled simulation, and artifact-backed reporting, giving architects, safety engineers, and digital-twin practitioners a systematic basis for identifying and prioritizing safer design alternatives.
Autonomous micromobility vehicles (AMVs) need local planning that balances task progress, safety, and pedestrian interaction quality in pedestrian-rich shared spaces. Evaluating such planners is difficult because studies vary scenarios, seeds, metrics, outputs, and aggregation rules, while single-score leaderboards hide which behaviors produce a ranking. This paper proposes a repeatable, auditable, multi-objective benchmark protocol for AMV local planning. Before comparison, it fixes the scenario set, repeated seeds, measured metrics, stored outputs, planner-interface records, and aggregation procedure. We demonstrate it with a frozen robot_sf_ll7 campaign: 47 shared-space scenarios, three evaluation seeds, and 141 scenario-seed episodes per planner in one differential-drive AMV configuration. The stress test surfaces a descriptive safety–performance separation: a Proximal Policy Optimization (PPO)-family profile reaches higher observed mean task success than a classical reciprocal-avoidance baseline, while that baseline keeps lower collision exposure. Absolute success stays low for both, with most scenarios unsolved by either. Because this learned policy was trained on a superset of the evaluation scenarios, its higher success reflects behavior on the benchmark set, not held-out generalization—an overlap the protocol records per planner rather than hiding in one score. The finding is bounded to these configured pipelines, not a universal planner-family ranking. The contribution is an auditable comparison framework tracing results from the scenario matrix and fixed seeds to episode records, aggregate reports, and manifests.
Lennart Luttkus, L. Mikelsons· Future Transportation· 0 citations
: High-density crowd accidents in complex urban walkway networks represent significant public safety risks characterized by non-linear dynamics and unpredictable emergence. While existing crowd models have primarily focused on evacuation within controlled facilities, they often lack the capacity to explain how behavioral heterogeneity interacts with real-world walkway networks to generate localized congestion. This paper proposes an Agent-Based Modeling (ABM) framework specifically designed to analyze the impact of diverse pedestrian behavior compositions on network-wide congestion dynamics. The proposed framework adopts a centralized time-stepped execution structure that integrates complex movement factors, including multi-destination trip chains and density-responsive navigation. To demonstrate its practical utility, the framework is applied to a realistic urban scenario in the Itaewon district, Seoul, utilizing empirical datasets on transportation demand and network topology. The study investigates how varying the distribution of behavior profiles—such as shortest-path, route-choice, and crowd-avoid strategies—reshapes emergent crowding patterns. The results highlight the framework’s strength as a robust scenario-based analysis tool for identifying systemic risks and providing actionable insights for crowd management in dense urban environments.
Gyu Ri Kim, Sang Hwa Lee, Jang Won Bae· Proceedings of the 16th Inte...· 0 citations
Path planning for power-line inspection requires asset-specific clearance from conductors, towers, and insulators, together with temporal separation from moving obstacles. Existing methods often optimize path length or a soft risk objective without showing that a single, explicit feasibility predicate is enforced during both search and postprocessing. We propose PL-RiskPlanner, a constraint-consistent framework that applies asset- and obstacle-specific checks before admitting grid edges, ranks the admitted edges using a power-line risk field, predicts moving-object positions at estimated UAV arrival times, and applies the same sampled verifier during smoothing. The contribution lies not in a new or faster A* search rule but in a power-line-specific feasible-set formulation that remains consistent across search, shortcutting, and final evaluation. A bounded margin accounts for uncertainty in velocity estimates. We evaluate eight executable planners across six synthetic scenario families and conduct controlled experiments on obstacle density, component ablation, risk weights, route crossings, prediction noise, and equal verifier budgets. Across 1800 nominal PL-RiskPlanner trials, no safety flag was recorded by the independent evaluator (Wilson 95% upper bound, 0.21%), and the mean verified-path length was 110.64 m. In controlled route crossings, the no-prediction variant produced 222 flags in 300 trials, whereas no safety flag was recorded for the complete planner. At velocity-noise level 0.4, the uncertainty margin changed the flag count from 203 to 47 and increased the mean simulator-reference dynamic clearance from −0.27 to 2.23 m. A 3600-trial equal-verifier-budget audit quantifies the resulting completion–computation trade-off without equating algorithm-specific iteration counters. Although the 50,400 trial records are specific to the stated simulator, they provide a reproducible protocol for safety-constrained inspection-path planning.
Cong Li, Yonggang Li, Lei Tan et al.· Electronics· 0 citations
In the critical domain of emergency response, rapid building evacuation and rescue are often hindered by complex layouts and dynamic hazards such as fire spread, toxic smoke, structural instability, and unpredictable occupant behaviour. To address the challenge of optimizing rescue sweep strategies, this paper establishes a comprehensive mathematical framework that transitions from traditional static planning to dynamic, adaptive decision-making, with the objectives of minimizing rescue completion time and maximizing occupant safety under uncertain conditions. First, a three-level evaluation system based on the AHP–Entropy Weight Method is constructed to quantify 32 critical influence factors, ranging from environmental conditions and building layout to occupant behaviour and technical support. The resulting weighted framework provides a data-driven foundation for subsequent optimization models. Three representative building configurations are then examined. For a 1D linear layout, a Divide-and-Conquer strategy reduces completion time by 42% compared with a single-entry approach. For a 2D branching layout, the rescue problem is formulated as a load-balanced Multi-Agent Traveling Salesman Problem (MA-TSP). For a 3D high-rise structure, a hierarchical Pincer Strategy is developed by nesting the MA-TSP within a multi-objective vehicle-routing framework and strategically leveraging vertical assets such as fire-service elevators. Finally, a Dynamic Adaptive Replanning Model is introduced using time-dependent hazard functions, dynamic edge costs, and occupant re-entry events. A Rolling-Horizon Replanning mechanism continuously updates rescue routes as the emergency evolves.
Qihao Chen· Journal of Frontiers in Tech...· 0 citations
Agentic Designer is proposed, a progressive, multi-agent framework that formulates structure-aware interior layout generation as an iterative and constraint-verified decision process and significantly outperforms state-of-the-art methods, demonstrating substantial improvements in strict structural adherence and functional design coherence.
Zhijing Yang, Hao-Cheng Lin, Zhihua Xu et al.· IEEE Transactions on Pattern...· 1 citation
Pedestrian safety at urban intersections is influenced by complex interactions among environmental conditions, social cues, and boundedly rational behavioral responses, making intervention planning under uncertainty particularly challenging. Existing studies have largely emphasized descriptive analysis or predictive modeling, with limited attention to prescriptive decision support under data scarcity. This study proposes a Bayesian-network-based robust optimization framework for pedestrian safety planning, complemented by a large language model (LLM)-assisted parameter-elicitation process. We first construct a three-layer Bayesian network based on the stimulus–organism–response paradigm to represent the propagation of risk from environmental cues through latent psychological mechanisms to behavioral violations and accident risk. To support model initialization when site-specific behavioral data are limited, we introduce an LLM-assisted elicitation protocol that maps literature-based qualitative evidence to intervention mechanisms, effect directions, and qualitative strength classes. Numerical parameter ranges are subsequently assigned through explicit mapping rules rather than generated directly by the LLM. We then formulate a bi-objective robust optimization model that distinguishes between physical interventions, which alter root-node distributions, and cognitive interventions, which modify conditional probability tables. Using the ϵ-constraint method, the framework generates Pareto-optimal intervention portfolios and evaluates cost–risk trade-offs under behavioral uncertainty and adverse operating scenarios.
Ming Liu, Yueyu Ding, Jinfeng Li· Applied Sciences· 0 citations