Autonomous Intersection Management

Updated 20 November 2025

Autonomous Intersection Management is a system that enables dynamic, vehicle-level scheduling at intersections through real-time time-space reservations.
It employs slot-based, optimization, and learning-based methods to maximize throughput, minimize delays, and ensure collision-free navigation.
Recent research in AIM tackles challenges such as multi-intersection coordination, resilience to sensor failures, and integration with mixed traffic scenarios.

Autonomous Intersection Management (AIM) is a class of approaches for coordinating the movement of connected and autonomous vehicles (CAVs) through intersections without reliance on traditional fixed-phase traffic signals. By exploiting fine-grained vehicle-level scheduling in space and time, AIM seeks to maximize throughput, minimize delay, and guarantee safety, even in heterogeneous, high-density, or mixed-traffic environments. This article surveys the key principles, canonical models, optimization and learning-based methodologies, robustness and safety frameworks, and current research frontier in AIM.

1. Core Principles and Operational Models

AIM fundamentally departs from signalized intersection management, which coordinates conflicting movements in coarse groups via time-phased "green" intervals. In AIM, each vehicle negotiates an individualized time–space crossing policy with an intersection manager (centralized, decentralized, or distributed). The canonical objective is to enforce, at all times,

$\|p_i(t) - p_j(t)\| \ge d_{\mathrm{safe}}$

for all pairs of vehicles $i, j$ whose trajectories could conflict, where $p_i(t)$ is the position of vehicle $i$ and $d_{\mathrm{safe}}$ is a prespecified minimum separation (Krishnan et al., 2018, Zhong et al., 2020).

AIM approaches decompose into several archetypes:

Slot-based / Reservation-based Schemes: Vehicles request discrete reservations (time, space, or combined space–time tiles) before entering the intersection. A central manager grants reservations if and only if proposed trajectories are non-overlapping, often formalized as a mixed-integer program (Krishnan et al., 2018, Zhong et al., 2020).
Optimization-based and MPC Frameworks: The intersection manager solves a receding-horizon optimal control or mathematical program, enforcing vehicle dynamics, collision constraints, and comfort/fuel objectives for all participants (Krishnan et al., 2018, Cederle et al., 14 May 2024).
Graph/Queue-based or Platoon-based Methods: The intersection is modeled as a network of nodes (lanes/platoons) and edges (conflict relationships), with scheduling or queuing rules for vehicle or platoon entry (Bashiri et al., 2018, Zhong et al., 2020).
Rule-based/Heuristic Approaches: Predetermined or context-aware right-of-way heuristics are applied. Example: "yield to the right" or prioritized merging (Krishnan et al., 2018).
Distributed and Learning-based Approaches: Each CAV learns or negotiates its own crossing schedule or policy, coordinating implicitly via local observations or explicit communication (Li et al., 2023, Cederle et al., 14 May 2024).

2. Trajectory Planning, Prioritization, and Scheduling

AIM must solve a constrained multi-agent scheduling problem over dynamic, usually non-convex, feasibility domains. Representative mathematical abstractions include:

Slot/Tilereservation (e.g., Dresner-Stone): For a vehicle $i$ , reserve a set of $(s, t)$ slots such that for any conflicting $j$ , $\delta_{i,(s,t)} + \delta_{j,(s,t)} \le 1$ (Krishnan et al., 2018).
Priority/Delay Minimization: Assign to each vehicle or platoon a reservation interval $[t_i^{enter},\,t_i^{exit}]$ so that for all conflicts

$t_i^{exit} + \Delta_s \le t_j^{enter} \;\;\text{or}\;\; t_j^{exit} + \Delta_s \le t_i^{enter}$

where $\Delta_s$ is a safety headway (Parks-Young et al., 2022, Bashiri et al., 2018).

Platoon-based Scheduling: Optimize a combined cost function over schedule permutations, trading off between total delay, delay variance, and fairness. Optimization is feasible in real time for reasonable platoon sets due to combinatorial pruning (Bashiri et al., 2018).
Learning-based Policies: Encode the intersection scene as a graph (vehicles as nodes, conflict relations as edges) and employ graph neural network-based actor-critic RL algorithms to predict per-vehicle accelerations or crossing choices (Klimke et al., 2022, Cederle et al., 12 Jul 2025). Pareto-optimal tradeoffs between throughput, emissions, and fairness can be systematically extracted (Cederle et al., 12 Jul 2025).

3. Safety, Robustness, and Real-Time Guarantees

Maintaining provable collision-avoidance is central to AIM. Methodologies span:

Hard Constraints and Formal Verification: Explicit disjointness of reserved space–time slots, mixture of integer-program feasibility, or formal temporal logic specifications (e.g., LTL/CTL) (Arfvidsson et al., 27 Aug 2024). Safety filters via reachability analysis produce time–state corridors that guarantee non-overlapping traversals under bounded uncertainty.
Chance Constraints and Uncertainty Quantification: Some approaches model cars with stochastic dynamics and compute bounding ellipses in position space to ensure that the probability of intersection overlap remains below a small threshold $\epsilon$ (Vitale et al., 2022). Event-triggered re-optimization further reduces computational/communication overhead compared to naive periodic execution.
Learning-based Safety Wrappers: Empirical methods inject conservative penalties for near-collision or missed time windows into the RL reward function, or use discriminators/adversaries to provide dense safety supervision during policy learning (Li et al., 2023, Li et al., 2022). Hybrid adversarial structures or hierarchical RL reduce the risk of "rush-in" collisions typical in flat (single-layer) RL (Li et al., 2022, Li et al., 2023).
Trust-aware and Mixed-Traffic Control: Real-world deployments must operate with partial penetration of trusted/connected vehicles. Embedding a trust authority quantifies agent reliability and scales buffer sizes accordingly to mitigate the risk from dishonest, faulty, or unpredictable agents. This augment traditional reservation logic with dynamic belief updating and subjective logic-based trust propagation (Cheng et al., 2021).
Fail-safe Integration with Legacy or Human-driven Vehicles: Hybrid architectures coordinate CAVs with reservation-based or local negotiation schemes, while legacy vehicles are managed by actuated signals. Mode-switching between centralized, decentralized, or fixed-signal regimes is triggered by thresholds in CAV penetration or congestion (Parks-Young et al., 2022, Yan et al., 2021).

4. Learning-Based, Multi-Agent, and Distributed AIM

Recent advances use deep reinforcement learning, distributed multi-agent coordination, and graph convolutional architectures to address the scalability and adaptation limits of classical AIM.

Observation and Coordination: Decentralized policies using local surround-view sensing can match or surpass the performance of centralized controllers in simulation. Parameter sharing across turning intentions and prioritized scenario replay help overcome sample inefficiency and focus training on the most challenging spatial configurations (Cederle et al., 14 May 2024).
Global vs Local Policy Generalization: Graph-based RL models leveraging rich node and edge features generalize much better across intersection layouts (including unseen ones) than traditional node-only or rule-based policies. Addressing upstream/downstream impact and management over arterial corridors remains an open line (Klimke et al., 2022, Zhong et al., 2020).
Multi-objective and Fairness-aware RL: Pareto-optimal tradeoffs between throughput, emissions (electric vs. petrol), and equity in scheduling are captured via multi-objective RL (e.g., TD3 with a parametric scalarization vector). Final policy deployment can be tuned via post-hoc fairness constraints on delay/throughput gaps between vehicle categories, accommodating both environmental and equity targets (Cederle et al., 12 Jul 2025).

5. Benchmarking, Performance Metrics, and Real-world Constraints

Formal benchmarking of AIM proposals requires scenario diversity and standardized metrics:

Metrics: Average delay, throughput, collision rate, space utilization, fuel/energy consumption, fairness (e.g., delay variance or group-wise wait times), and comfort (stopping rate, total acceleration). These are mathematically defined and routinely reported (Krishnan et al., 2018, Bashiri et al., 2018, Cederle et al., 12 Jul 2025).
Scalability and Real-time Feasibility: Classical mixed-integer optimizations become intractable for large $N$ , with real-time operation typically only feasible when $N\leq$ 10–20 in the manager's control zone. Learning-based and distributed approaches have shown promise in maintaining constant or sublinear per-agent computation, with empirical performance within 1–10 ms per decision (Gadginmath et al., 2020, Gunarathna et al., 2022, Cederle et al., 14 May 2024).
Simulation and Demonstrator Platforms: Urban-scale benchmarks use SUMO, SMARTS, FLOW, or in-house traffic simulators. Metrics are reported under both synthetic and real-world demand patterns, including Poisson arrivals, arrival surges, lane heterogeneity, and mixed-class (human/autonomous) fleets (Cederle et al., 14 May 2024, Yan et al., 2021, Klimke et al., 2022).

6. Open Challenges and Future Directions

Despite significant progress, AIM faces a range of unresolved technical and operational issues:

Multi-intersection and Corridor Coordination: The leap from isolated intersections to networked or corridor-level optimization remains limited. Synchronization over distributed AIM nodes and scalable reservation of platoon slots require further research (Zhong et al., 2020, Krishnan et al., 2018).
Resilience to Communication and Sensing Failures: Robust protocols that degrade gracefully in the presence of latency, dropouts, cyber attacks, or adversarial agents are critical (Cheng et al., 2021, Krishnan et al., 2018).
Mixed-autonomy Integration: Seamless coexistence and fair resource allocation between human-driven vehicles, CAVs, and vulnerable road users under uncertainty in intent remains a top priority (Yan et al., 2021, Parks-Young et al., 2022).
Formal Safety under Uncertainty: Real-time reachability analysis and temporal logic-based safety filtering are emerging as computationally viable for embedding safety contracts at deployment (Arfvidsson et al., 27 Aug 2024).
Energy, Emissions, and Social Objectives: Next-generation AIM controllers are increasingly incorporating environmental and fairness objectives, often via Pareto-optimal multi-objective reinforcement learning frameworks (Cederle et al., 12 Jul 2025).
Transparent, Standardized Evaluation: The field lacks widely adopted open benchmarks with mixed-traffic, multi-intersection, and adverse-event scenarios. Standardized data and simulator APIs are essential for robust comparative evaluation (Zhong et al., 2020, Cederle et al., 12 Jul 2025).

AIM represents a paradigm shift in intersection control—moving from phase-based external regulation to vehicle-level, personalized, and safety-guaranteed spatiotemporal scheduling. The combined progress in optimization, multi-agent/distributed learning, formal verification, and systems engineering continues to shape the trajectory of next-generation urban mobility.