A single reroute decision, made seconds before congestion hits, can ripple through an entire city’s traffic network—either smoothing flow or shifting bottlenecks. In a world where 74% of Americans report traffic as a “major problem” (Pew, 2023), even a 1.8% reduction in travel delays isn’t trivial. It’s the difference between arriving on time and missing a flight; between a delivery making its window or adding to urban congestion. That’s the scale of impact reported in a new simulation study by Tony Kinchen, Ting Bai, and Andreas A. Malikopoulos (2026), which introduces a predictive rerouting system for connected and automated vehicles (CAVs) that cuts through traffic chaos with surgical precision—by rerouting fewer vehicles, not more.
Their framework, tested in a detailed model of New York City’s streets, uses a diffusion convolutional recurrent neural network (DCRNN) to forecast traffic and electric vehicle (EV) charging demand up to two minutes ahead. It then evaluates alternative routes not just for speed, but for CO₂ emissions, distance, and exposure to charging hotspots. The result? A routing policy that reduces the average travel-time index by 1.8% compared to no rerouting—outperforming four other active methods—while accepting only 41 route changes across all scenarios. By contrast, the next most aggressive methods triggered 146 and 153 reroutes. Less intervention. Better outcomes. That’s the paradox at the heart of this work.
This isn’t just about faster commutes. It’s about rethinking how we manage mixed traffic—where only a fraction of vehicles are controllable—as electrification adds new layers of complexity. Every EV heading to a charger is a data point in a larger system: its battery level, destination, and route choice affect congestion, emissions, and grid load. The DCRNN framework treats these not as noise, but as signals. And in doing so, it reveals a counterintuitive truth: smarter routing doesn’t mean constant redirection. It means waiting for the right moment to act.
The Science
The researchers built their experiment inside SUMO (Simulation of Urban MObility), a high-fidelity microscopic traffic simulator that models individual vehicles, lane changes, emissions, and EV energy use. They used a real-world-inspired network of Manhattan—638 road segments, 254 junctions—where 5% to 45% of vehicles were CAVs eligible for rerouting. All vehicles, human-driven or not, followed the same origin-destination demand patterns and encountered identical disruptions: scheduled intersection blockages and rear-end slowdowns designed to mimic real-world incidents.
At the core of their system is the DCRNN, a deep learning model that combines graph-based spatial reasoning with temporal forecasting. Unlike traditional models that treat roads as isolated links, DCRNN sees the city as a dynamic graph, where traffic on one street influences neighboring segments through diffusion-like patterns. It ingests historical data—traffic speed, congestion levels, CO₂ emissions, EV counts, and charging urgency—and predicts conditions 120 seconds into the future. The model was trained on 80% of chronological data, validated on the remaining 20%, and never exposed to future information during live simulation—a safeguard against unrealistic performance.
When a CAV approaches a decision point, the system generates up to three alternative routes using K-shortest-path logic, excluding the current path. Each route is scored using a weighted cost function:
where:
- $T'_e(t)$ is the predicted travel-time index (TTI), normalized to a maximum of 5.0;
- $C'_e(t)$ is current or estimated CO₂ emissions;
- $l'_e$ is normalized edge length;
- $Q_e(t)$ is a dimensionless “charging risk” score based on nearby EVs needing charge, station proximity, and temporal urgency.
The weights—$a=0.95$, $b=3.1$, $c=0.90$, $d=2.5$—prioritize emissions and charging risk over pure travel time. A route change is accepted only if it improves the total cost, preserves connectivity to the destination, and passes cooldown and duplication checks (e.g., no more than three reroutes per vehicle, 4-second cooldown).
This setup was compared against four benchmarks:
- No-reroute reference: Vehicles follow original routes regardless of congestion.
- K-shortest-path (KSP): Chooses among three shortest routes based on current congestion, emissions, and length.
- Predictive-density routing (PD): Anticipates when traffic density will hit critical thresholds (80 vehicles/km) and reroutes proactively.
- V2X proactive routing: Uses perfect knowledge of upcoming disruptions (180-second warning) to avoid affected zones.
All methods ran under identical conditions, with results averaged across five CAV penetration levels.
What They Found
The headline result is modest but meaningful: DCRNN reduced average travel-time index by 1.8% compared to no rerouting—the only method to consistently outperform the baseline across penetration levels. More strikingly, it achieved this with far fewer interventions. While KSP and V2X rerouted 146 and 153 vehicles respectively, DCRNN made just 41 changes—fewer than one-third as many.
Route Changes by Method
Total number of accepted route changes across all CAV penetration levels (5% to 45%) for each routing method.
| Label | Value |
|---|---|
| DCRNN | 41 |
| K-shortest-path | 146 |
| V2X proactive | 153 |
| Predictive-density | 5 |
| No-reroute | 0 |
This selectivity wasn’t accidental. DCRNN’s algorithm waits for high-leverage moments: when a small reroute can prevent cascading delays or avoid a charging bottleneck. In contrast, KSP and V2X act more frequently, often redirecting vehicles onto routes that become congested due to the rerouting itself—a phenomenon known as induced demand in transportation planning.
In terms of performance, DCRNN led all active methods in three key metrics:
- Lowest average TTI: 1.21 vs. 1.23–1.26 for others
- Highest average speed: 34.7 km/h vs. 33.1–34.1 km/h
- Lowest aggregate CO₂ emissions: 1,892 kg vs. 1,910–1,960 kg
Aggregate CO₂ Emissions by Method
Aggregate CO₂ emissions (kg) across all scenarios and penetration levels.
| Label | Value |
|---|---|
| DCRNN | 1,892 |
| K-shortest-path | 1,910 |
| Predictive-density | 1,960 |
| V2X proactive | 1,945 |
| No-reroute | 1,870 |
Yet here’s the twist: the no-reroute scenario had even lower emissions and total distance traveled. That’s because rerouting, even when beneficial for congestion, often means longer trips. Every detour adds kilometers—and thus emissions—even if it avoids idling in traffic. The tradeoff is real: reducing congestion can increase total vehicle kilometers traveled (VKT), a well-known challenge in traffic management.
The key insight is that DCRNN optimizes within the active rerouting space—it doesn’t claim to beat doing nothing. Instead, it shows that among systems that do intervene, less can be more. As
illustrates, DCRNN’s CO₂ emissions are the lowest among rerouting methods, but still higher than the no-reroute baseline. The same pattern holds for distance (
).
What explains DCRNN’s efficiency? The answer lies in its forecasting horizon and multi-objective scoring. While V2X knows exactly when and where disruptions occur, it lacks predictive traffic modeling—it reroutes based on a fixed rule (“avoid this zone”) rather than dynamic conditions. DCRNN, by contrast, forecasts how congestion will evolve and weighs the net benefit of rerouting. It might keep a vehicle on a slightly slower route if the alternative would overload another corridor or pass a charging station with long queues.
And it’s not just about congestion. The inclusion of $Q_e(t)$—the charging risk term—means the system avoids routes where EVs are likely to cluster at chargers, which can create localized bottlenecks. In one simulated case, a CAV was rerouted away from a station edge where DCRNN predicted a surge in EVs needing charge, even though the alternate route was 12% longer. The tradeoff paid off: the vehicle avoided a 4-minute delay caused by queueing vehicles.
Why This Changes Things
At first glance, a 1.8% improvement might seem underwhelming. But in transportation systems, marginal gains compound. The U.S. Department of Transportation estimates that traffic congestion cost $88 billion in 2021 in lost productivity and fuel (FHWA, 2022). A 1.8% reduction across major cities could save $1.6 billion annually—enough to fund 1,000 miles of new bike lanes or 200 electric bus fleets.
More importantly, this study challenges a deeply held assumption in smart mobility: that more data and more control lead to better outcomes. The V2X method, with perfect foresight of disruptions, rerouted the most vehicles but performed worse than DCRNN. Why? Because it treated rerouting as a binary switch—“avoid the disruption”—without modeling the downstream effects. It’s like evacuating a building without considering whether the exit stairs can handle the crowd.
DCRNN, by contrast, embodies a systems-thinking approach. It doesn’t just react; it anticipates feedback loops. This is critical in mixed traffic, where only a minority of vehicles are controllable. You can’t optimize the whole network by commanding a few nodes—you have to nudge them strategically.
The implications extend beyond traffic. As cities electrify, transportation and energy systems are merging. A 2023 NREL study found that uncoordinated EV charging could increase peak grid demand by 25% in dense urban areas. DCRNN’s charging-risk module is a prototype for mobility-aware grid management—routing vehicles not just to minimize travel time, but to smooth charging demand.
Consider a future where your EV receives a reroute suggestion not because of traffic, but because the charger at your destination is expected to be at 90% utilization in 15 minutes. The detour adds two minutes to your trip but saves 18 minutes of waiting—and prevents a local transformer from overloading. That’s the world this research points toward.
And it does so without requiring full automation. The framework works with partial CAV penetration—starting at just 5%. That’s crucial, because widespread CAV adoption remains distant. A 2024 McKinsey report estimates that even by 2030, only 15–20% of new vehicles sold will be highly automated. DCRNN’s effectiveness at low penetration means cities can start reaping benefits long before fleets are fully connected.
What’s Next
The study has limitations. It’s a single deterministic simulation—no Monte Carlo runs or confidence intervals. Real-world traffic is noisier, human drivers less predictable, and communication networks prone to latency. The DCRNN model assumes reliable vehicle-to-infrastructure (V2I) connectivity, which remains patchy in most cities.
Perhaps the biggest open question is scalability. The current model forecasts 120 seconds ahead—a useful window, but short for long-distance trips. Extending the horizon requires more complex models and greater computational load. And while the 1.8% TTI improvement is promising, it’s unclear how it scales with larger networks or higher disruption frequency.
Future work could integrate real-time grid data, allowing the system to respond to electricity prices or renewable generation fluctuations. Imagine a CAV being routed to a charger during a solar surplus, when electricity is cheap and clean—turning vehicles into mobile grid assets.
Another frontier is equity. The current model optimizes aggregate performance, but what about individual fairness? Could rerouting disproportionately benefit CAV owners while leaving human-driven vehicles in congestion? The paper doesn’t address this, but it’s a critical concern as smart mobility risks becoming a two-tier system.
Finally, there’s the question of incentives. Why would a CAV owner accept a longer route? Without rewards—toll discounts, priority charging, or carbon credits—compliance may be low. The technology works, but the ecosystem around it must evolve.
Still, the core insight stands: in complex systems, precision beats volume. We don’t need to reroute every vehicle to improve traffic. We just need to reroute the right ones, at the right time, for the right reasons. That’s not just smarter routing. It’s a new philosophy for managing urban mobility—one that sees the city not as a collection of roads, but as a living network of interdependent flows.
As climate pressures mount and cities densify, we can’t build our way out of congestion. But we might be able to think our way out. This study suggests that the most powerful tool in traffic management isn’t more infrastructure, more vehicles, or even more data—it’s better judgment.