Meridia Insight Tech for Good Frontiers

When Self-Driving Cars Arrive, Traffic Models Won’t Survive

A new traffic model shows that self-driving cars could boost urban throughput by 22%—but only if traffic signals adapt in real time.

At 45% CAV penetration, a 0.5s reduction in time gaps boosts throughput by 22%—if signals adapt.

At 45% CAV penetration, a 0.5-second reduction in automated vehicle time gaps boosts intersection throughput by 22%—but only if traffic signals adapt. This isn’t a prediction from a decade away. It’s what the new MIXSim model reveals about the fragile balance between vehicle automation and urban infrastructure today (Haris & Roncoli, 2026). As cities inch toward mixed traffic—where human-driven and connected automated vehicles (CAVs) share roads—the old rules of traffic flow no longer apply. Capacity isn’t fixed. Congestion doesn’t spread uniformly. And the performance of an entire network hinges on microscopic decisions: how closely a CAV follows the car ahead, whether signals prioritize platoons, and how routing adapts to shifting bottlenecks. MIXSim, a novel macroscopic traffic model, captures these dynamics for the first time at scale, revealing that the future of urban mobility depends not just on how many self-driving cars we have, but on how well our models—and our signals—can keep up.

The Science

Traffic simulation has long operated on a spectrum: microscopic models like SUMO or Vissim track individual vehicles, capturing lane changes, acceleration patterns, and driver behavior in granular detail. These are powerful, but computationally expensive—too slow for testing citywide control strategies across hundreds of scenarios. On the other end sit macroscopic models like the Cell Transmission Model (CTM), which treat traffic as a continuous flow, dividing roads into cells and modeling vehicle movement through conservation laws. These are fast, scalable, and ideal for studying network-level phenomena like congestion propagation, spillback, and signal timing (Daganzo, 1994).

But traditional macroscopic models assume homogeneity: all vehicles behave the same. That assumption collapses in mixed-autonomy traffic, where CAVs can follow more closely, accelerate more smoothly, and coordinate at intersections in ways human drivers cannot. Prior multiclass CTM models have begun to address this by treating CAVs and human-driven vehicles (HDVs) as separate classes, allowing vehicle composition to influence link capacity and wave speed (Levin & Boyles, 2016). MIXSim goes further. Developed by Muhammad Haris and Claudio Roncoli, it is a first-order macroscopic model designed specifically for signalized urban networks with mixed autonomy, integrating three critical innovations:

  1. Commodity-specific routing: Each vehicle is tagged not just by class (HDV or CAV), but by origin, destination, and class—forming distinct “commodities.” This preserves route-level detail while maintaining computational efficiency.
  2. Composition-dependent fundamental diagrams: The triangular fundamental diagram—which relates traffic density to flow—is no longer fixed. Instead, its parameters (capacity, critical density, wave speed) adapt dynamically to the local mix of HDVs and CAVs in each cell.
  3. Signalized node dynamics with movement-specific saturation: At intersections, the saturation flow rate—the maximum discharge rate during green—is not a constant. It depends on the composition of vehicles requesting each turning movement, allowing CAV platoons to clear faster than HDV queues.

The model operates on a directed graph of links and nodes, with each link divided into cells. State variables track the number of vehicles per commodity per cell. Flow between cells is governed by sending and receiving functions derived from the fundamental diagram, while routing fractions determine how vehicles split at junctions. Crucially, the time gap—the temporal headway between successive vehicles—differs by class: CAVs can operate at 0.5–1.0 seconds, while HDVs typically require 2.0–3.0 seconds. This difference drives the entire model’s sensitivity.

What They Found

The authors tested MIXSim on a synthetic urban network under varying CAV penetration levels (0% to 100%) and CAV time gaps (0.5s to 3.0s), comparing fixed-time, actuated, and adaptive signal control strategies. The results reveal nonlinear, sometimes counterintuitive, responses to automation.

First, capacity gains are not linear with CAV penetration.

Network Throughput vs. CAV Penetration at Different Time Gaps

Throughput increases with CAV share but plateaus beyond 60% penetration due to signal bottlenecks.

Network Throughput vs. CAV Penetration at Different Time Gaps
LabelValue
CAV Time Gap: 0.9s1.18 x baseline
CAV Time Gap: 1.5s1.12 x baseline
CAV Time Gap: 3.0s1.05 x baseline

shows how network throughput increases with CAV share—but only up to a point. At a CAV time gap of 0.9 seconds, throughput rises steadily from 0% to 60% penetration, gaining 18% over all-HDV conditions. But beyond 60%, the curve flattens. Why? Because capacity is bottlenecked not by vehicle spacing, but by signal timing. If signals don’t adapt to shorter headways, the extra discharge potential goes unused.

Network Throughput vs. CAV Penetration at Different Time Gaps

Throughput increases with CAV share but plateaus beyond 60% penetration due to signal bottlenecks.

Network Throughput vs. CAV Penetration at Different Time Gaps
LabelValue
CAV Time Gap: 0.9s1.18 x baseline
CAV Time Gap: 1.5s1.12 x baseline
CAV Time Gap: 3.0s1.05 x baseline

Second, the value of tighter CAV spacing depends entirely on infrastructure responsiveness. When CAV time gaps shrink from 3.0s to 0.5s, the theoretical maximum saturation flow increases by 60%. But

Throughput Gain from Reducing CAV Time Gap (0.5s vs 3.0s)

Signal control strategy determines how much of the theoretical capacity gain is realized.

Throughput Gain from Reducing CAV Time Gap (0.5s vs 3.0s)
LabelValue
Fixed-Time Signals9 %
Actuated Signals14 %
Adaptive Signals22 %

shows that under fixed-time signals, realized throughput gains are only 9%. Under adaptive control—where green times adjust to queue lengths—the same change yields a 22% improvement. This gap underscores a central finding: automation benefits are not automatic. They require coordination between vehicle capabilities and signal logic.

Throughput Gain from Reducing CAV Time Gap (0.5s vs 3.0s)

Signal control strategy determines how much of the theoretical capacity gain is realized.

Throughput Gain from Reducing CAV Time Gap (0.5s vs 3.0s)
LabelValue
Fixed-Time Signals9 %
Actuated Signals14 %
Adaptive Signals22 %

Third, routing matters. In MIXSim, when CAVs can reroute in response to congestion, they avoid bottlenecks more effectively than HDVs. But this can shift congestion rather than eliminate it. In one scenario, CAVs rerouted around a saturated intersection, reducing their own delays by 30%—but increasing HDV delays by 15% due to spillover. The model captures this feedback: as CAVs clear faster, they free up upstream capacity, which in turn affects routing decisions downstream.

The model also reveals thresholds in system behavior. For instance, when CAV penetration exceeds 50%, backward wave speeds—the rate at which congestion propagates upstream—can exceed free-flow speeds under tight time gaps. This violates a standard assumption in first-order models and necessitates stricter numerical stability conditions (Haris & Roncoli, 2026, Eq. 16). In practical terms, it means congestion can spread faster than vehicles can travel, creating sudden, hard-to-manage gridlock if not properly modeled.

Why This Changes Things

For decades, traffic engineers have relied on static capacity values: a lane can handle 1,800–2,000 vehicles per hour, full stop. MIXSim shows this number is obsolete. In a mixed-autonomy world, lane capacity becomes a function—of vehicle composition, time gaps, signal control, and even routing behavior. A lane with 80% CAVs operating at 0.6s headways can support over 2,800 vehicles per hour. The same lane with 20% CAVs might manage only 2,100. Treating capacity as fixed risks both overbuilding infrastructure and underestimating resilience.

This has immediate implications for cities investing in smart corridors. Consider a pilot project deploying CAVs on a downtown route. If planners use traditional models, they might predict modest congestion relief. But MIXSim suggests the outcome depends critically on whether signals are upgraded. Without adaptive control, CAVs may simply queue more efficiently—reducing delays slightly, but not unlocking systemic gains. With it, the same vehicles could increase corridor throughput by 20%, delay the need for expansion, and improve reliability.

Moreover, MIXSim challenges the assumption that automation will uniformly benefit all road users. The rerouting experiment shows CAVs can exploit their agility to avoid congestion, potentially shifting burden onto HDVs. This isn’t just a technical detail—it’s a fairness issue. As cities deploy CAV-dedicated lanes or priority signaling, models like MIXSim can help quantify trade-offs and design equitable policies.

The model also reframes the timeline of automation’s impact. We don’t need 100% CAV penetration to see effects. At 30%, the local composition begins to alter wave speeds and discharge rates. At 50%, the network behaves fundamentally differently. MIXSim provides a tool to stress-test infrastructure against these transitions, identifying tipping points where small changes in penetration or behavior trigger large shifts in performance.

Compare this to existing tools. Microscopic simulators can capture heterogeneity, but struggle with city-scale, long-duration experiments. Aggregate models like the Network Macroscopic Fundamental Diagram (NMFD) offer speed but lose vehicle-class detail. MIXSim occupies a crucial middle ground: fast enough for policy exploration, detailed enough to capture composition-dependent dynamics. It’s not a replacement for microsimulation, but a complement—a “policy microscope” for probing systemic trade-offs.

What's Next

MIXSim is a framework, not a finished product. The authors acknowledge several limitations. First, it assumes perfect CAV compliance with time gaps and routing—unrealistic in early adoption phases. Second, it does not model lane-level interactions, platoon formation, or V2X communication protocols, which could further amplify capacity gains. Third, routing fractions are exogenous: the model uses them as inputs, rather than simulating route choice in response to travel times.

Future work could integrate dynamic traffic assignment, allowing CAVs and HDVs to adapt routes based on real-time conditions. This would enable studies of equilibrium behavior: Do CAVs converge to system-optimal routes, or do they create new inefficiencies through competitive rerouting? Another direction is calibration. The model’s parameters—especially CAV time gaps and signal response logic—must be grounded in real-world data from pilot deployments.

Perhaps most importantly, MIXSim opens the door to co-design: optimizing both vehicle behavior and infrastructure control in tandem. Imagine a city where CAVs don’t just react to signals, but communicate intent—“We’re a platoon of six, can you extend green by 3 seconds?” Signal controllers, informed by MIXSim-like models, could evaluate the trade-off: a small delay for cross-traffic versus a large gain in throughput. This isn’t science fiction. It’s the logical next step in a world where vehicles are no longer just users of infrastructure, but active participants in its operation.

The broader lesson is this: as autonomy reshapes mobility, our models must evolve too. We can’t manage what we can’t measure. MIXSim doesn’t just simulate traffic—it redefines what traffic is in the 21st century. And in doing so, it offers a roadmap for building cities that are not just smarter, but more adaptive, more equitable, and more resilient to the disruptions ahead.

[Uncaptioned image]
[Uncaptioned image] Source: Muhammad Haris, Claudio Roncoli
[Uncaptioned image]
[Uncaptioned image] Source: Muhammad Haris, Claudio Roncoli

Comments (0)

No comments yet. Be the first to share your thoughts.