The Radio Model That Flips Connected-Vehicle Safety Conclusions
Dense traffic at a signalized intersection was supposed to be where connected-vehicle safety earns its keep: cars whispering warnings about hard braking, red-light runs, and blind spots to each other over 5G sidelink radio. Run the numbers with the channel model most researchers use, and the system looks like a triumph. Switch to a model that actually accounts for how vehicles compete for radio spectrum, and the picture inverts — simulated traffic slows down instead of speeding up, and predicted hard-braking events more than double (Bello et al., 2026). The safety case flips direction entirely.
That reversal is the quiet alarm at the heart of a new paper from South Carolina State, North Carolina A&T, Dakota State, and independent researchers. For years, the standard way to evaluate vehicle-to-everything (V2X) safety has been to couple a network simulator with a microscopic traffic simulator — OMNeT++, the Veins framework, and SUMO — so that radio traffic and road traffic influence each other inside one experiment. The trouble, the authors show, is that the channel models baked into that stack were built for an older radio technology (IEEE 802.11p, used in dedicated short-range communications) and they solve for signal propagation only. They ignore the brutal competition for radio resources that actually governs delivery in 5G New Radio sidelink Mode-2 — the mode where vehicles pick their own transmission slots with no base station to arbitrate. The result is that these models report near-perfect message delivery in exactly the dense conditions where real deployments lose most of their messages.
The paper's contribution is a fix that doesn't require rebuilding the simulation stack. The researchers distilled the reception behavior of a reference 5G implementation (ns-3 5G-LENA, calibrated on 3GPP scenarios) into a compact, closed-form model that drops straight into the existing Veins pipeline. They call it NS3Learn. It tracks the reference simulator's per-instant delivery decisions to within a mean absolute deviation of 0.06 — versus 0.44 and 0.55 for the two alternatives a practitioner would otherwise reach for. And because every one of its mathematical stages maps to a named physical mechanism — half-duplex loss, scheduling collision, receiver capture, decoding — the fitted coefficients are not a black box but a readable account of why messages get lost.
The Science
Connected-vehicle safety applications rest on the timely exchange of basic safety messages (BSMs) — small bursts announcing a vehicle's position, speed, and heading. Before any field deployment, researchers quantify the benefits of these systems almost entirely through simulation, and those simulated benefits inform deployment plans, infrastructure investment, and the safety cases regulators read. The dominant platform couples OMNeT++ (a discrete-event network simulator), the Veins vehicular framework, and SUMO (a microscopic traffic simulator), all first developed around IEEE 802.11p. A safety conclusion drawn on this stack — a change in time-to-collision, a change in mean speed — is only as trustworthy as the model that decides which messages arrive.
In 5G NR PC5 sidelink Mode-2, delivery is governed by three mechanisms that matter more than propagation, and all three sharpen as vehicles accumulate. First, collisions: two vehicles can independently pick the same transmission resource, since there's no base station allocating them. Second, half-duplex loss: a vehicle that is transmitting cannot simultaneously hear a neighbor transmitting in the same instant. Third, capture: a receiver can sometimes decode the stronger of two overlapping signals, which softens the blow of collisions. Every one of these is density-driven — worst exactly at the crowded intersections where cooperative safety is supposed to deliver its value, and where propagation conditions may be perfectly benign. A channel model that resolves only path loss, fading, and shadowing reports messages arriving that a real deployment would lose.
The one open-sourced OMNeT++ tool that models the sidelink, OpenCV2X, implements an older LTE standard (Mode-4) and hasn't tracked NR. Simu5G brings 5G NR to OMNeT++ but models the cellular Uu interface rather than the PC5 sidelink. The standards-compliant NR Mode-2 implementation lives in ns-3, inside the 3GPP-calibrated 5G-LENA simulator. So researchers were stuck: keep the SUMO-coupled pipeline and accept a model that omits the dominant loss mechanisms, or abandon the stack and go rebuild on ns-3.
The paper's third option is distillation. The researchers drove ns-3 5G-LENA with exported SUMO trajectories across vehicle densities, then labeled 10.5 million reception outcomes from its physical- and MAC-layer traces. Those labels trained a cascade of stages — each a logistic function of vehicle density or signal-to-interference-plus-noise ratio (SINR) — that correspond one-to-one to half-duplex loss, SB-SPS collision, capture, and decoding. Each stage is a logistic function so the composed model stays closed-form and cheap enough to evaluate for every single message in an event-driven simulation. The cost of fidelity moves offline, into one labeling-and-fitting campaign, and no protocol module remains to maintain at run time.
What They Found
The evaluation spanned two signalized urban networks, six penetration levels from 1 to 100 percent, and five random seeds per condition — three on a corridor and two at a structurally different intersection. The headline number: NS3Learn tracked the per-instant, in-range packet delivery ratio of ns-3 5G-LENA to a mean absolute deviation of 0.06. The two alternatives — the untouched stack base channel and a combined reference built from two published NR Mode-2 models — deviated by 0.44 and 0.55 against a per-seed variation of just 0.03 (Bello et al., 2026).
That comparison matters beyond accuracy for its own sake. A reception model matters where it changes a conclusion, and the paper pushes each representation all the way through to the driving behavior it induces. Under ordinary traffic, changing only the communication model reversed the direction of the simulated speed trend and more than doubled predicted hard-braking events. In other words, the same road network, the same demand, the same penetration — but swap which messages are allowed to arrive, and the safety answer flips.
The model also generalizes. Coefficients fitted at one signalized junction carried to a structurally distinct intersection with roughly 20 percent additional error — no re-estimation required. That transferability is what makes distillation practical: retargeting to a new radio configuration means repeating the offline fitting, not modifying simulation code.
How close each model tracks ns-3 5G-LENA delivery
Mean absolute deviation in per-instant packet delivery ratio compared to ns-3 5G-LENA. Lower is better; per-seed variation across conditions was 0.03.
| Label | Value |
|---|---|
| NS3Learn (distilled) | 0.06 |
| Stack base channel | 0.44 |
| Combined analytical reference (M3) | 0.55 |
Transfer error to a different junction
Mean absolute deviation at the fitted (source) intersection versus a structurally different junction with coefficients transferred without re-estimation, showing roughly 20% additional error.
| Label | Value |
|---|---|
| NS3Learn at source intersection | 0.06 |
| NS3Learn at distinct intersection (transfer) | 0.072 |
Hard-braking predictions flip with the reception model
Relative predicted hard-braking events under ordinary traffic. Changing only the communication model more than doubled hard-braking events and reversed the direction of the simulated speed trend.
| Label | Value |
|---|---|
| Reception model that omits resource competition | 1 |
| Reception model that includes resource competition | 2.1 |
Perhaps the most consequential capability is adversarial. The cascade includes a flooder term — a model of what happens when an adversary injects messages at high rates to starve legitimate vehicles of spectrum, a denial-of-service attack on the very mechanism the model describes. The authors found that an adversarial flood "registers" in their model at any flood rate, whereas neither reference represents the attack at all. The coefficient structure couples the flooder's rate with in-range vehicle density through a log-log interaction, so the model captures that flooding hurts more the more crowded the channel.
Why This Changes Things
The paper's deeper argument is about how realism gets transferred between simulators. Previous approaches to adding communication fidelity either derived analytical models (each valid within its own stochastic-geometry assumptions, each covering one slice of the mechanism) or ported full protocol implementations (expensive, hard to maintain, and prone to drift from the standard). Distillation sidesteps both. The authors' phrasing is apt: the cost of fidelity moves offline, into one labeling-and-fitting campaign.
There's a genuine intellectual honesty in how the cascade is structured. Every fitted number belongs to a named mechanism. The half-duplex stage is a logistic function of density; the baseline collision probability is logistic in density; capture gets two decode branches, one for the collision-free case and one for the collided case, and it's the collided branch that implicitly carries the interferer penalty — a near transmitter survives a collision that a far transmitter does not. "That asymmetry is the capture effect," they write. Because each stage maps to a physical mechanism, a researcher can sanity-check every coefficient against what it claims to encode, and can attribute degradation to contention versus propagation by abating single stages.
The practical implication for transportation agencies is direct. Most agencies and researchers evaluating connected-vehicle safety already operate a working SUMO/Veins/OMNeT++ pipeline. The paper shows they don't need to throw it away. They can keep the pipeline they've built and validated, bolt in NS3Learn, and suddenly represent the losses that dominate dense traffic at signalized intersections — and the denial-of-service impact that no current stack model captures at all. The comparison matters most where the stakes are highest: the dense intersection, the crowded approach, the moment when a missed basic safety message translates into a missed stop.
What's Next
The paper is honest about its limits. The in-sample accuracy of 0.06 deviation is exactly that — in-sample. The transfer experiment to a second junction adds roughly 20 percent error, which is reassuring but not a guarantee for arbitrary geometries. The reference model itself assumes an under-saturated channel and deviates near saturation, and the sparsest tested conditions extrapolate below the range where the collision model was validated. The authors also note that ns-3 5G-LENA replays a prescribed trajectory file, so its driving cannot respond to which messages arrive — the very feedback loop that Veins closes.
What the paper opens up is a question: which other expensive simulators could be distilled the same way? The methodology — label reception outcomes from a 3GPP-calibrated reference, fit a closed-form cascade whose stages map to named mechanisms, transfer across simulators and across geometries by re-fitting offline rather than rewriting code — is not specific to sidelink. The same logic could carry realism between any pair of simulators where fidelity lives in one and cost or maintainability in the other.
For connected-vehicle safety specifically, the message is uncomfortable and important. The field's conclusions about deployment benefits have been, in part, an artifact of the communication model — optimistic in precisely the conditions where those conclusions matter most. The fix here isn't a bigger simulator or a more elaborate protocol stack; it's a compact, readable, transferable account of how messages actually get lost in dense traffic. That's a model you can check, a model you can move, and a model that might make the numbers line up with the real world.