When multiple vehicles converge on an intersection, the traditional approach is to handle them in the order they arrive—First-In, First-Out (FIFO)—or via rigid time-slicing mechanisms like traffic lights. While intuitive, these methods create massive inefficiencies. The core bottleneck at an intersection is not the physical space itself, but rather the right-of-way transition cost. Every time an intersection switches from serving one direction to another, vehicles must decelerate, stop, or clear the conflict zone, destroying momentum and causing cascading delays.
To solve this, a growing body of research focuses on non-FIFO vehicle sequencing and flexible order coordination. By allowing vehicles to deviate from their strict arrival order, intersections can group vehicles traveling in the same direction, minimizing right-of-way transitions and maximizing throughput.
The Cost of Right-of-Way Transitions
To understand why flexible sequencing works, consider a scenario where cars arrive alternatingly from the North and East. Under a strict FIFO policy, every single car forces a right-of-way switch. The car behind must wait for the intersection to clear entirely before it can proceed.
As Arnaud de La Fortelle demonstrates in Analysis of optimal solutions to robot coordination problems to improve autonomous intersection management policies, transitioning away from these rigid first-come, first-served policies eliminates complete vehicle stops. By globally optimizing the scheduling order, average relative time loss under heavy traffic drops from forty percent down to roughly ten percent.
The underlying mechanism is explained microscopically by Liu et al. in Analytically Modeling Unmanaged Intersections with Microscopic Vehicle Interactions. By encouraging vehicles from the same direction to pass through in groups, the system minimizes the frequency of right-of-way transitions. This intentional deviation from FIFO sequencing allows platoons of vehicles to clear the conflict zone together, drastically reducing total intersection delay. Yan et al. reinforce this insight in Multi-agent Path Finding for Cooperative Autonomous Driving, showing that optimizing the crossing order can cut average traffic delays nearly in half compared to standard FIFO heuristics, particularly in high-density scenarios where uncoordinated arrivals otherwise trigger gridlock.
Decoupling Sequence from Trajectory
Finding the mathematically optimal crossing order for dozens of fast-moving autonomous vehicles is computationally daunting. To make this problem tractable in real-time, researchers decouple the challenge into two distinct layers: an upper-level discrete sequencing problem and a lower-level continuous trajectory planning problem.
For example, Xu et al. propose a bi-level cooperative driving strategy in A bi-level cooperative driving strategy allowing lane changes. The upper level optimizes the passing order, while the lower level plans the precise trajectories and lane changes required to execute that order safely. Similarly, Yan et al. introduce an algorithm called OBS-KATS, adapting Multi-agent Path Finding (MAPF) techniques from warehouse robotics to find the optimal crossing sequence while strictly respecting realistic vehicle physics.
Predictive Horizons and Constraint-Guided Safety
Reordering vehicles requires looking into the future. If a coordination system only considers vehicles currently inside the intersection, it cannot group upcoming vehicles effectively.
To provide these look-ahead capabilities, Black et al. introduce future-focused control barrier functions (ff-CBFs) in Future-Focused Control Barrier Functions for Autonomous Vehicle Control. These functions predict vehicle behavior over a future time horizon, allowing the system to proactively adjust trajectories and improve throughput by up to 43% while maintaining strict safety guarantees. Taking interactivity a step further, Samavi et al. develop a framework in SICNav-Diffusion: Safe and Interactive Crowd Navigation with Diffusion Trajectory Predictions that combines diffusion models for joint trajectory prediction with a bilevel Model Predictive Control (MPC) framework, simultaneously planning robot actions and refining human predictions for collision avoidance.
Even when applied to signalized intersections, incorporating predictive dynamics yields massive benefits. Hu et al. demonstrate in Incorporating Queueing Dynamics into Schedule-Driven Traffic Control that modeling how queues dynamically evolve as different signal timing plans are evaluated results in a 10–15% reduction in average wait times compared to myopic adaptive controls.
Distributed Optimization and Learning
Centralizing the coordination of every vehicle in a city creates massive communication overhead and single points of failure. As a result, recent breakthroughs leverage decentralized optimization and multi-agent reinforcement learning (MARL) to distribute the computational load.
In A Decentralised Approach to Intersection Traffic Management, Vu et al. model intersection control as a Distributed Constraint Optimization Problem (DCOP). By solving it locally via the max-sum algorithm, they achieve a 60% reduction in rush-hour waiting times and demonstrate rapid stabilization after dynamic disruptions, such as emergency vehicle arrivals. For centralized or edge-computed coordination of connected and automated vehicles (CAVs), Zhao et al. introduce CoBT-SRL in Constraint-Guided Behavior Transformer for Centralized Coordination of Connected and Automated Vehicles at Intersections, utilizing a safety-aware reinforcement learning framework paired with transformers to maintain a 0% collision rate while minimizing travel times.
When scaling up from isolated intersections to entire urban networks, coordination becomes a hierarchical challenge:
- Hierarchical Control: Ma et al. introduce a feudal multi-agent deep reinforcement learning framework (FMA2C) in Feudal Multi-Agent Deep Reinforcement Learning for Traffic Signal Control. Regional "manager" agents set high-level goals, while local "worker" agents execute signal adjustments, consistently reducing queue lengths across the network.
- Network Communication: Chu et al. address the chaotic nature of multi-agent environments in Multi-agent Reinforcement Learning for Networked System Control by implementing a localized communication protocol that allows neighboring intersections to coordinate, enabling the network to recover gracefully from peak congestion.
- Digital Twins: Kumarasamy et al. integrate decentralized graph-based MARL with a Digital Twin in Integration of Decentralized Graph-Based Multi-Agent Reinforcement Learning with Digital Twin for Traffic Signal Optimization, providing real-time data synchronization that yields a 38.94% average reduction in congestion during peak hours.
Hybrid and Vision-Based Extensions
The transition to fully autonomous, flexible intersections does not have to happen overnight; the same principles are being adapted to modern infrastructure using advanced sensing and hybrid control.
For instance, Tchuitcheu et al. utilize an internet of smart cameras in Internet of smart-cameras for traffic lights optimization in smart cities to dynamically prioritize queues and emergency vehicles without relying on in-car automation. To handle severe congestion at isolated intersections, Nie et al. develop the FMAA method in FMAA: A Flexible Signal Timing Method for An Isolated Intersection with Conflicting Traffic Flows, which pairs vehicle-to-infrastructure (V2I) data with adaptive all-red control to flush out conflicting movements efficiently.
Furthermore, Dong et al. combine dynamic speed guidance with bidirectional green wave coordination in A Multi-Level Speed Guidance Cooperative Approach Based on Bidirectional Periodic Green Wave Coordination Under Intelligent and Connected Environment, reducing vehicle stops by over 44% by shaping the arrival patterns of vehicles before they even reach the intersection. Finally, to prevent congestion from forming in the first place, Gol et al. leverage Signal Temporal Logic (STL) in Cause Mining and Controller Synthesis with STL to mine the "controllable causes" of traffic jams, synthesizing feedback controllers that proactively guide traffic away from bottleneck states.
By shifting the paradigm from rigid time-slicing to flexible, non-FIFO sequence optimization, these diverse approaches collectively demonstrate that the future of traffic management lies in minimizing right-of-way transitions through intelligent, group-based coordination.
Key Foundational Work
Analytically Modeling Unmanaged Intersections with Microscopic Vehicle Interactions (2018) — contributed 2 concept elements to this direction.
Analysis of optimal solutions to robot coordination problems to improve autonomous intersection management policies (2016) — contributed 1 concept elements to this direction.
A Decentralised Approach to Intersection Traffic Management (2018) — contributed 1 concept elements to this direction.