Unsignalized Intersection Management

Updated 21 November 2025

Unsignalized intersection management is defined by methods that coordinate vehicle movement without traffic lights using decentralized and centralized scheduling strategies.
Techniques including combinatorial optimization, robust model-predictive control, and reinforcement learning improve throughput, reduce delays, and enhance collision avoidance.
Empirical studies show up to 65% travel time reduction and significant safety gains in mixed-autonomy scenarios through advanced real-time conflict resolution.

Unsignalized Intersection Management Strategy

Unsignalized intersections, where vehicular right-of-way is not governed by traffic lights or physical barriers, present acute challenges for coordinated, efficient, and safe maneuvering—especially in environments mixing human-driven vehicles (HVs), connected automated vehicles (CAVs), and robotic agents. Strategies for managing these intersections span low-overhead decentralized algorithms, combinatorial scheduling optimizations enabled by vehicle connectivity, robust model-predictive control, and reinforcement learning architectures. Below, key paradigms are organized by principle, problem formulation, algorithmic structure, implementation considerations, and empirical results.

1. Scheduling-Based and Combinatorial Approaches

Scheduling-based strategies treat intersection management as a constrained resource allocation or sequencing problem, formulating the core challenge as combinatorial optimization or graph-theoretic scheduling.

Neural Monte Carlo Tree Search for Multi-Intersection Scheduling

A representative recent method transforms the unsignalized intersection problem into a board-game abstraction, where each platoon's sequence of conflict-zone reservations forms a row on a grid, and the entry/exit of platoons through zones is scheduled via discrete delays $d_j$ (Shi et al., 24 Oct 2024). The objective is to minimize intersection makespan and cumulative delays, subject to Boolean collision-avoidance in each conflict zone: $\forall\,i\in\mathcal{C},\;j\neq k:\ (t^{j,i}_{entry}+d_j+\Delta^{j,i})\leq t^{k,i}_{entry}+d_k\ \lor\ (j\leftrightarrow k)$ This is solved with Neural Monte Carlo Tree Search (NMCTS), where upper confidence bounds are computed using learned policy and value networks. To enhance generalization and stabler training, Parallel NMCTS (PNMCTS) applies prioritized resampling and curriculum learning, showing robust reductions in crossing times—43% in light and 52% in heavy traffic compared with FIFO—while maintaining a 95% success rate on unseen scenarios (Shi et al., 24 Oct 2024).

MCTS with Heuristic Rollouts

Alternative tree-search strategies encode the passing order as leaves in a search tree, with MCTS performing iterative selection, expansion, simulation, and backpropagation (Xu et al., 2019). Performance is further improved with domain-aware rollout policies enforcing lane FIFO and earliest-conflict resolution, reducing the branching factor and improving convergence speed. In simulations, MCTS achieved near-optimal delay and throughput with millisecond runtimes, vastly outperforming FIFO at higher arrival rates.

Graph-Theoretic and Clique-Based Schedulers

Decentralized methods for low-to-medium density scenarios employ compatibility or harmony matrices encoding pairwise non-conflict among maneuvers, and solve the maximal clique problem on small graphs (typically $M\leq 5$ nodes) to maximize concurrent crossings (Katole et al., 2023). These methods are constant-time in practical deployments and, in simulations up to 350 PCU/hr/lane, demonstrated waiting times near those of V2I reservation-based controllers.

2. Priority-Assignment and Mixed-Autonomy Coordination

In mixed traffic streams, right-of-way assignment and priority queueing are employed to balance efficiency and safety in the presence of non-cooperative or unpredictable vehicles.

Heuristic Priority Queues (HPQ) with Per-Vehicle MPC

For intersections admitting mixed CAV/CHV/HV flows, a heuristic priority queue structures lane-wise head-of-line candidate selection and right-of-way allocation (Zhou et al., 2022). Rules explicitly differentiate CAV and CHV: CHVs are only allowed right-of-way if they conflict with no granted vehicles, while CAVs can be granted in the presence of at most one conflict but must switch to conflict-resolving MPC. Four per-vehicle control modes (car-following, cruise, waiting, conflict-solving) are native outputs of the integrated HPQ and local MPC. This joint structure reduced travel time up to 65% over adaptive signals or hybrid AIM benchmarks, achieving robust near-signalized throughput and smooth speed profiles (Zhou et al., 2022).

Automated Right-of-Way (AROW) Protocols

Distributed implementations of stop-controlled logic (SC-I) for CAVs embed FIFO and explicit turn-assignment into message-based protocols (AROW1–AROW5) with strict safety and liveness guarantees (Shah et al., 2023). Each approaching CAV, upon crossing a “detection threshold,” joins a leader/follower set, and right-of-way assignment is formalized as a coloring of a conflict graph with scheduling constrained to enforce FIFO among conflicts. The protocol is fully-distributed (O( $k$ ) messages, $k$ competing vehicles), ensures collision-freeness even under up to 25% CAV non-compliance, and shows 0–3% ambiguity events under simulation, compared to >14% for all-way stops or delay-based CIM.

3. Robust Model-Predictive Control, Uncertainty, and Covariance Steering

Hierarchical Robust Control with Tube-based RMPC

Decentralized frameworks resolve crossing order negotiation and receding horizon trajectory planning in the spatial domain (Pan et al., 2022). Optimal control is posed as a SOC-program following convexification of velocity and energy dynamics, and constraints for rear-end and lateral collisions are enforced via time differences along the vehicles’ spatial arc-length. Robustness to model errors and estimation noise is achieved with tube-based RMPC; tightened state/input sets guarantee recursive feasibility. Empirical results show up to 16% energy savings over FIFO sequencing for fixed average travel time, and a Pareto front showing that moderate time penalties yield large energy reductions under higher densities.

Chance-Constrained Covariance Control under Bandwidth Limits

Recent methods address state estimation and wireless resource constraints by jointly planning mean trajectories and covariance evolution, enforcing probabilistic collision constraints as deterministic SOC constraints (Bai et al., 5 Aug 2025). A context-aware status update scheduler allocates communication slots to vehicles according to urgency indices that incorporate age-of-information, tracking error, and local risk, mitigating the effects of stale state information under bandwidth restrictions. The method achieves collision probabilities down to 5% with 20 vehicles and outperforms sequential and receding-horizon baseline strategies in both collision risk and throughput.

Spatial-Domain MPC with Mixed Traffic and Motion Uncertainty

For intersections with HDV and CAV coexistence, spatial-domain change-of-variables (lethargy and cumulative time) enable a unified linear collision-avoidance formulation for crossing, following, merging, and diverging conflicts (Zhao et al., 5 Dec 2024). HDV path/speed uncertainties are decoupled and inner-approximated in the temporal domain, with the receding horizon MPC implementation employing a real-time iteration (RTI) scheme solving a single QP per control step. This achieves orders-of-magnitude speedup compared to full SQP, with negligible optimality loss ( $<$ 2.3%), and maintains required time gaps even under adversarial human driver behaviors.

4. Learning-Based and Game-Theoretic Decision Strategies

Reinforcement Learning for Coordination and Priority Ordering

Model-free RL policies (TD3, DDPG, DQN, PPO) are used to select crossing orders, lane assignments, or yielding behaviors. When fused with sequential real-time trajectory planners, they provide linear-scaling throughput and maintain safety guarantees (G. et al., 2023, Luo et al., 2022). RL-aided sequencers outperform five traditional heuristics—including FCFS—achieving up to 150% improvements in objective value and 30–50% reductions in average crossing time, with real-world robotic deployments confirming robust, collision-free execution.

Game-Theoretic and Prospect-Theory Models

At the behavioral level, unsignalized intersection crossing can be modeled as repeated static games, with payoffs mapping collision risk and efficiency through Prospect Theory value functions (Li et al., 2021). The Nash equilibrium of the 2x2 acceleration/deceleration choice matrix is computed at each decision epoch, and driver style (safety vs. speed emphasis) is tunable via a weight $\omega_j$ . In simulation, the prospect-theoretic strategy guaranteed approximately 98% safety in two-vehicle interactions, exceeded 85% safety even under extreme bandwidths, and achieved up to a 34% efficiency gain over leader-follower rules.

Safe RL with Action Masking and Robust Perception

For urban intersections populated with vulnerable road users, deep Q-networks enhanced with safe action-masking and observer/predictor LSTMs (for handling noisy, partially observed pedestrian states) are shown to maintain zero-collision performance while achieving competitive traversal times (Mokhtari et al., 2021). This approach employs explicit forward simulation of ego and pedestrian trajectories over a short horizon to filter out unsafe DQN actions, integrating LSTM-based belief updates and future trajectory predictions for robust decision making in dense and ambiguous scenarios.

5. Lane Assignment, Formation Control, and Lane-Change Planning

Flexible Lane Direction and Formation Control

In multi-lane settings, scheduling and lane assignments are joined in a two-dimensional (layer × lane) optimization. Lane-direction is treated as a decision variable to adaptively maximize simultaneous conflict-free crossings, with lane-change and longitudinal placement executed via discrete relative-coordinate system and multi-agent conflict-based search (Cai et al., 2021, Chen et al., 2021). A graph-based minimum-clique-cover schedule ensures maximal coexisting groups, and formation reconfiguration algorithm regularizes lane-changes and inter-vehicle gaps. This approach achieves 10–40% reductions in average intersection delay versus fixed-lane unsignalized and signalized strategies, with up to seven vehicles traversing the conflict zone as a batch under dynamic demand and turn ratios.

6. Responsible Safety Frameworks and Spatio-Temporal Envelopes

Right-of-Way Assignment and RSS Integration

Augmenting Responsibility Sensitive Safety (RSS) policies, right-of-way assignment (RWA) divides the region ahead of higher priority vehicles into forbidden (FA), negotiation (NA), and free (FR) areas, explicitly enforcing minimum separation as a function of relative speeds, braking capacities, and reaction delays (Xing et al., 2019). Pairwise negotiation mechanisms allow dynamic RWA transfer via V2V communication—within RSS-mandated bounds—enabling concurrent non-conflicting crossings and reducing intersection delay compared to RSS alone (8.1s vs ~22s at λ = 200 veh/h), though still outperformed by global-cooperative (MCTS) solutions.

7. Sensing, Uncertainty, and Real-World Implementation

Stereo Vision, Adaptive Sampling, and Bézier-Convex Hull Planning

Managing unsignalized T-intersections with sensory uncertainty, an adaptive depth-sampling scheme tightens worst-case closing speed bounds from stereo vision by modulating sampling interval as a function of depth-dependent variance. Collision avoidance logic combines these speed bounds with a geometric convex-hull test on candidate Bézier exit paths, accepting yielding or proceeding only if guaranteed by currently observed bounds (Kumar et al., 30 Dec 2024). NGSIM-based validation demonstrates minimum separation >3.8m and sub-0.2m lateral error, with no collisions across dense, multi-neighbor scenarios.

In summary, unsignalized intersection management strategies comprise a spectrum from decentralized, intent-based clique scheduling and low-overhead rule-based mechanisms to centralized, globally-optimal coordination solutions leveraging combinatorial search, robust MPC, and deep RL. Practically, real-time feasibility and deployment considerations (communication delays, bandwidth, mixed autonomy, perception uncertainty) drive the selection and hybridization of approaches. Empirical studies consistently demonstrate that advanced conflict-aware scheduling (MCTS, clique cover, RL-aided sequencing, robust tube-based MPC) achieves significant performance improvements in throughput, safety, and delay under a variety of traffic compositions and arrival regimes (Shi et al., 24 Oct 2024, Bai et al., 5 Aug 2025, Pan et al., 2022, Xu et al., 2019, Katole et al., 2023, Zhou et al., 2022, Li et al., 2021, Xu et al., 2020).