Source-linked AI summary
Simulating Cognitive Smart Freight Corridors with Agent-Based Models and Reinforcement Learning
Madelaine Martinez-Ferguson, Chun Wang, Mustafa Can Camur, Xueping Li
TL;DR
Smart freight corridors need evaluation methods that capture adaptive, system-level coordination beyond fixed control policies. This paper develops a three-layer ABM with V2X, RL, and MARL, finding differentiated gains across scenarios and stronger Cognitive resilience under most tested disruptions. The authors also identify limits in control scope, network topology, training, validation, and empirical calibration.
Problem
Existing smart-corridor approaches rely on predetermined policies, while physical pilots cannot cover the full space of stochastic demand, disruptions, infrastructure, and control configurations.
Method
The paper develops a three-layer agent-based model coupling physical infrastructure, V2X connectivity, RL corridor control, and MARL charging coordination.
Results
+27.0% throughput and −63.9% congestion for Cognitive versus Baseline, while Assisted and Cognitive achieve −7.5% and −7.9% energy per kilometer, respectively.
Takeaways & Limitations
RL/MARL control supports adaptive corridor management, with Cognitive showing superior overall disruption resilience except for station failure, where Assisted performs best.
Takeaways & Limitations
The study is limited by a unidirectional network, restricted control scope, incomplete training and validation, few replications, and lack of real freight-demand calibration.
Abstract
from arXiv · showhide
Smart freight corridors offer a practical pathway for connected and automated vehicle (CAV) deployment in freight transportation, but physical experimentation is expensive and existing approaches rely on predefined control policies that cannot capture adaptive behaviors. This paper presents an agent-based modeling (ABM) framework coupling a physical infrastructure layer, a connectivity layer (V2X), and a decision layer integrating reinforcement learning (RL) and multi-agent reinforcement learning (MARL) for platoon formation and charging coordination. We evaluate three scenarios (Baseline, Assisted, and Cognitive) using throughput, congestion, energy, emissions, and robustness metrics. Preliminary results indicate that the Cognitive scenario achieves higher throughput and lower congestion than the baseline, while the Assisted scenario delivers meaningful energy savings per kilometer through platooning. Sensitivity analysis shows that the throughput advantage of the smart corridor widens under conditions with high demand and that MARL coordination extracts greater utilization from fixed charging capacity than rule-based assignment.
1 INTRODUCTION
The paper frames smart freight corridors as a system-level pathway for deploying CAVs, combining infrastructure, V2X connectivity, and adaptive control to address congestion and coordination challenges. It proposes an ABM framework with RL and MARL to evaluate corridor performance under multiple operating conditions.
- Motivation: CAV benefits depend nonlinearly on penetration rates, infrastructure design, and corridor control policies, so vehicle-level technology alone is insufficient.Controlled studies report fuel or energy savings of up to 25% through cooperative platooning.
- Motivation: Smart freight corridors concentrate intelligent transport investments on high-impact freight links while enabling controlled operational policies.These corridors use infrastructure such as sensors and V2X communication to support connected freight operations.
- Research gap: System-level questions include which connectivity and automation capabilities, control policies, and disruption responses produce measurable corridor improvements.The paper highlights robustness under incidents and demand surges as an explicit concern.
- Research gap: Joint coordination across physical, cyber, and decision layers is analytically difficult under stochastic demand, heterogeneous vehicles, and dynamic V2I interactions.Real-world pilots also cannot cover the full combinatorial space of infrastructure configurations, disruptions, and control policies.
- Approach: The study develops a modular ABM integrating RL for platoon formation and lane management with MARL for distributed charging coordination.It compares Baseline, Assisted, and Cognitive scenarios using throughput, congestion, energy, emissions, and disruption-robustness metrics.
2 RELATED WORK
Prior work integrates smart-corridor infrastructure and connectivity with increasingly sophisticated vehicle and corridor controls, but adaptive coordination remains incomplete. This gap motivates a freight-focused framework combining learning-based control, charging coordination, and disruption evaluation.
- Smart freight corridors: Smart freight corridors combine physical infrastructure and cyber connectivity into cyber-physical systems supporting functions such as V2X connectivity and self-adaptability.These systems support intelligent transport applications including variable speed limits and cooperative adaptive cruise control.
- Smart freight corridors: Prior corridor studies optimize trajectories, vehicle grouping, or other coordination decisions across physical and cyber layers.The reviewed approaches include mixed-integer optimization and cyber-physical platooning frameworks.
- System-level assessment: Corridor management can improve capacity utilization, while charging availability affects electric-truck throughput and delays.These findings make traffic control and charging coordination consequential system-level decisions.
- Research gap: Existing decision layers largely rely on predetermined optimization models or fixed control rules that do not adapt to stochastic demand, disruptions, or evolving conditions.This limitation motivates adaptive learning methods for corridor control.
- Adaptive learning: Adaptive learning has been demonstrated for single-vehicle trajectories, but not yet for the paper’s combined freight-corridor coordination problem.The reviewed DRL framework focuses on energy use for one vehicle trajectory rather than multi-agent corridor coordination.
- Adaptive learning: MARL has coordinated multiple freeway control decisions, but prior work reviewed here targets passenger traffic without freight agents, charging coordination, or disruption scenarios.This identifies the missing freight-specific evaluation dimensions.
- Freight applications: Freight RL studies address route and charging planning, yet operate at network-routing level without treating corridor infrastructure as an active control layer.The reviewed work therefore differs from the proposed corridor-level infrastructure control perspective.
3 METHODOLOGY
The methodology models a smart freight corridor as a three-layer ABM linking physical infrastructure, V2X connectivity, and adaptive RL/MARL decisions. It simulates mixed-autonomy freight operations and compares progressively intelligent scenarios using multi-dimensional corridor states, rewards, and performance metrics.
- Corridor conceptual model: The corridor architecture integrates physical infrastructure, cyber connectivity, and adaptive decision making.Figure 1 represents the three-layer architecture and the associated agent types, information exchange, and policy outputs.
- Corridor conceptual model: The physical layer includes road segments, managed lanes, charging stations, and freight terminals with capacity, speed, queue, and disruption attributes.Charging stations have three ports and power levels of 150, 250, or 350 kW.
- Traffic model: The congestion model uses α = 0.15 and β = 2 for freight-dominated corridors rather than the standard β = 4.The paper states that β = 2 reflects more gradual congestion onset in freight streams.
- Connectivity layer: The connectivity layer provides configurable V2I and V2V messaging, modeled with a fixed 50 ms per-message latency and variable penetration rates.Baseline has no connectivity, Assisted adds V2X and supervised automation, and Cognitive further adds RL and MARL optimization.
- Decision layer: The decision layer uses one RL corridor controller for platoon formation and lane activation plus five MARL station agents for local pricing and priority actions.The decomposition assigns global corridor decisions to one controller and distributed charging decisions to communicating station agents.
- Simulation implementation: The modular Python ABM represents heterogeneous autonomy modes and individual state-of-charge trajectories directly while allowing mixed autonomy within each replication.Truck agents make route, platoon, charging, and advisory-compliance decisions.
- State representation: The corridor controller observes a 14-dimensional state containing congestion, incidents, station utilization, queues, platooning, battery, charging, scenario, time, and managed-lane variables.The observation vector is explicitly defined in equation (2) and its accompanying prose.
- Simulation workflow: Simulation steps last 0.08 hours and proceed through controller actions, traffic updates, truck decisions, charging and V2X updates, and metric recording.Each step advances the coupled corridor dynamics before recording evaluation metrics.
4 PRELIMINARY RESULTS AND DISCUSSION
Across matched corridor scenarios, Cognitive provides the strongest throughput and congestion outcomes, while Assisted and Cognitive reduce energy use through platooning. Sensitivity and disruption tests show that adaptive control is especially valuable under high demand and varied operating conditions.
- Evaluation Design: The evaluation compares Baseline, Assisted, and Cognitive scenarios on throughput, travel time, energy, emissions, congestion, and robustness using matched infrastructure.Each scenario uses a 10-hour horizon, a 1,500-truck fleet pool, and five replications; the performance metrics include completed trips per hour, average trip duration, kWh/km, CO2 proxy emissions, flow-to-capacity congestion, and disrupted-to-baseline robustness.
- Scenario Comparison: 27.0% higher throughput and 63.9% lower congestion index make Cognitive the strongest scenario relative to Baseline.Cognitive also reduces travel time by 7.2%, while energy consumption per kilometer improves under both Assisted and Cognitive.
- Charging Sensitivity: 98–101 trips/hr Cognitive throughput remains stable across charging-port counts, while queue wait and stranded-truck rates decline as ports increase.The steepest queue-wait improvement occurs between 5 and 15 ports, with diminishing returns beyond 15 ports; MARL sustains performance across more charging configurations than rule-based assignment.
- Scenario Comparison: −7.5% and −7.9% energy consumption per kilometer under Assisted and Cognitive, respectively, reflect platooning-induced aerodynamic drag reduction.Cognitive’s higher completed-trip volume increases its CO2 proxy emissions, although per-kilometer emissions remain comparable across scenarios.
- Demand Sensitivity: Cognitive’s congestion advantage widens at high demand, particularly when demand reaches at least 1.5× and rule-based policies saturate.Demand sensitivity varies origin-terminal demand from 0.25× to 2.0× across five replications.
- Disruption Robustness: Cognitive maintains robustness scores of 1.00 under accidents and weather and 1.31 under demand spikes, but Assisted reaches 1.02 under station failure.The robustness assessment injects disruptions during evaluation without retraining; Assisted collapses under accidents with a 0.00 score and a 155% travel-time spike.
5 CONCLUSIONS
The paper presents an integrated three-layer ABM in which learning-based control improves corridor performance and supports adaptive freight infrastructure decisions. Limitations include restricted control scope, a unidirectional network, and incomplete training and empirical validation.
- RL/MARL control improves throughput and energy efficiency across the evaluated corridor scenarios.The Cognitive scenario also demonstrates superior resilience under disruptions.
- The Cognitive scenario supports shifting corridor management from rule-based vehicle coordination toward adaptive intelligent infrastructure.The paper connects this shift to investment, operational policy, and governance decisions.
- The control framework is limited to platoon formation, managed lane control, and fixed station price adjustments.Future extensions include dynamic rerouting, variable speed limits, and competitive pricing.
- The current unidirectional network limits transferability to more realistic settings.The authors identify bidirectional, multipath topologies and policy evaluation across network configurations as next steps.
- Training and empirical validation require reward ablations, extended training, optimization benchmarks, additional replications, paired statistical tests, and real freight demand calibration.These extensions would strengthen evidence beyond the directional findings reported here.