Why Radar Keeps Surviving the "Cameras Are Enough" Argument, and Where the Argument Has a Point

Every so often, someone with a large platform declares automotive radar obsolete, and every time the industry quietly keeps putting radar on cars. The camera-only argument is not stupid, it is made by serious people with real engineering reasons, and parts of it are correct. But after fifteen years of working on both sides of this debate: I started my career shipping stereo-camera FCW and AEB, spent years building radar signal processing chains at Tier-1 suppliers, owned radar sourcing on an EV program and a robotaxi platform, and today lead a camera-based Visual SLAM team: I keep watching radar survive, and I think the reasons it survives are structural, not sentimental.

The most instructive version of this story is public. In May 2021, Tesla removed radar from North American Model 3 and Model Y production and moved Autopilot to “Tesla Vision,” accepting temporary feature limitations like a capped Autosteer speed to do it (Electrek, 2021). Nine months later, NHTSA opened a preliminary evaluation into unexpected “phantom braking” covering roughly 416,000 of those radar-less 2021–2022 vehicles, after 354 complaints in nine months (Carscoops, 2022). And by 2023, an FCC filing and a subsequent teardown confirmed a new higher-resolution radar (known internally as “Phoenix”) shipping in Hardware 4 Model S and Model X vehicles (Teslarati, 2023; Edge AI and Vision Alliance, 2024). The company that made the strongest public case against radar designed a better one. That arc (remove, struggle, redesign) is the whole debate in miniature.

What radar actually buys you is physics, not features

The camera-only argument usually frames radar as a redundant object detector: cameras see objects, radar sees objects, so radar is a second, worse copy of information you already have. That framing misses what radar contributes, because radar’s value is not that it detects objects, it is what it measures about them and when it keeps measuring.

First, radar measures range and radial velocity directly. A Doppler return gives you closing speed in a single measurement cycle, with no dependence on texture, lighting, or a neural network’s depth estimate. Cameras infer range and velocity (from stereo disparity, from monocular depth networks, from tracking objects across frames) and inference has failure modes that direct measurement does not. Anyone who has tuned an ACC target selection algorithm, as I did early in my career on a stereo-vision system, knows the difference between a measured closing rate and an estimated one when a vehicle cuts in at 30 meters. The estimated one is late, noisy, or both, exactly when you need it most.

Second, radar’s performance is largely indifferent to the conditions that degrade cameras: darkness, low sun in the lens, fog, spray, heavy rain, snow. This is not marketing, it falls out of the wavelength. On a robotaxi program where I owned the radar systems, we deliberately benchmarked candidate imaging radars in rain, fog, and snow as part of acceptance, because those conditions are precisely where the radar has to carry the perception stack. The camera’s contribution collapses gracefully or ungracefully depending on the stack, but it collapses; the radar’s mostly doesn’t.

Third (and this is the point program managers should underline) radar’s failure modes are decorrelated from the camera’s. A safety case for a hands-off or eyes-off feature is fundamentally an argument about residual risk, and two sensors whose failures are independent multiply their reliabilities in the tails. Two cameras, or one camera and a very good neural network, share failure modes: they both go blind in the same fog, saturate in the same sun glare, and mislearn from the same training distribution. When I sit in a DFMEA for a perception system, the question is never “how many detectors do we have,” it is “what single environmental condition or systematic fault takes out how much of our sensing.” Camera-only architectures concentrate that risk; camera-plus-radar architectures split it.

The program-level implication: radar is cheap insurance against the tail of the distribution, and the tail is where safety cases live. A corner radar is one of the least expensive perception sensors on the vehicle, and it buys you an independent physical measurement channel. When I have run sensor-suite trade studies, radar rarely wins on any single average-case metric, and it almost never gets cut, because cutting it makes the worst-case rows of the analysis turn red.

The regulatory floor is rising toward radar’s strengths

Even if you find the first-principles argument unpersuasive, the test protocols are not optional. In April 2024, NHTSA finalized FMVSS No. 127, which requires all new light vehicles by September 2029 (a deadline NHTSA proposed in 2026 to extend by roughly two years, though the performance requirements themselves stand) to stop for lead vehicles at speeds up to 62 mph, apply AEB at up to 90 mph for imminent vehicle collisions and up to 45 mph for pedestrians, and to detect pedestrians in darkness, not just daylight (NHTSA, 2024). Euro NCAP has scored pedestrian AEB in night-time scenarios since it added them to its VRU protocols in 2018, and current protocols split pedestrian points between day and night tests (Euro NCAP protocols).

Notice what these tests demand: long-range detection at high closing speeds, and vulnerable road user detection in low light. Those are, respectively, radar’s home turf and the camera’s hardest case. A high-speed AEB intervention at 90 mph needs reliable range-rate on a target far beyond where monocular depth estimates are trustworthy. Night pedestrian detection is achievable with cameras (better imagers and better networks have moved this a long way) but every camera-based night solution is fighting its sensor’s physics, while radar simply does not care that the sun went down.

This is why I tell engineering leaders that the sensor suite is not sized by the average demo; it is sized by the hardest mandated test in the worst mandated condition. Regulations and NCAP protocols keep moving toward exactly the corners where radar is strong, which is a structural reason (independent of any one company’s philosophy) why radar keeps reappearing on BOMs that someone tried to shorten.

Where the camera-only argument has a point

Here is the part radar advocates skip, and shouldn’t. The camera-only camp is right about several things, and I have lived most of them.

Bad fusion is worse than no fusion. The strongest form of the camera-only argument was never “radar is useless”; it was “when radar and vision disagree, a mediocre arbitration layer produces the worst of both.” That is correct, and I say this as someone who has built fusion systems: a radar track with poor angular resolution, fused naively with a vision detection, generates exactly the intermittent false positives that drivers experience as phantom braking. Classic automotive radar genuinely struggles to separate a stationary vehicle from an overhead bridge or a roadside sign, elevation resolution on older sensors is poor to nonexistent, and multipath ghosts are real. If your fusion layer treats a low-quality radar as an equal witness, you have added a confabulating sensor to your jury. Removing it can genuinely improve the median experience. The honest conclusion is not “remove radar” but “a radar you won’t invest in fusing properly is a liability”, which is a much less quotable sentence.

Cameras alone can carry a real safety feature set. This claim is not hypothetical. Subaru’s EyeSight shipped for years as a stereo-camera system with no front radar, and HLDI found it cut pedestrian-related injury liability claim frequency by 35 percent, the second generation by 41 percent (IIHS/HLDI, 2018). I shipped stereo-vision FCW and AEB more than a decade ago; a well-engineered camera-only stack has been sufficient for strong L1/L2 driver assistance for a long time. If your product is supervised driver assistance with a human fallback, the marginal safety value of radar is real but bounded, and an honest engineer should admit it.

Every sensor you keep is a production burden, not just a BOM line. This is the part of the argument that resonates most with my systems-side experience. A radar is not just its piece price. It is a supplier relationship with an RFQ, an eSOR, and a change-management history. It is an environmental qualification campaign (thermal cycling from −40 °C to +85 °C, vibration, humidity, EMC) that can uncover problems eighteen months before SOP. It is a mounting location fought over with studio and body engineering, a connector, a harness branch, and above all a manufacturing-line calibration station with its own cycle time and scrap rate. On one EV program, redesigning the radar calibration process and the associated line procedures produced six-figure documented savings, which tells you how much money was sitting in that one station to begin with. When a chief engineer says “every sensor we delete simplifies the factory,” they are not being philistines. They are describing costs that never appear in a perception engineer’s slide deck.

So the camera-only argument, steelmanned, is this: for a supervised L2 product, a first-rate camera stack beats a camera stack plus a second-rate radar, and it beats it on safety-per-dollar once you count the factory. Stated that way, I agree with it. That is also precisely why it does not generalize.

Why the argument keeps losing anyway

The argument fails at the boundary where the driver stops being the fallback. The moment a feature’s operational design domain includes night, weather, and high speed (and the moment your safety case can no longer write “driver supervises” next to every hazard) decorrelated sensing stops being insurance and becomes load-bearing. That is why no eyes-off or driverless program I am aware of runs camera-only, why the regulatory floor keeps rising into radar’s strengths, and why the most prominent camera-only advocate re-engineered a radar rather than staying the course. The phantom-braking evaluation that NHTSA opened in 2022 (ultimately covering some 695,000 vehicles) was finally closed in July 2026 after reported incidents fell from roughly 300 a year to a handful (The BRAKE Report, 2026), vision-only can be made to work well for supervised driving, and it took one of the best-resourced AI teams in the industry years of iteration to get there. Most programs do not have that runway, and no program should assume it at the safety-case level above L2.

There is also a quieter engineering reason radar survives: it keeps getting better faster than the argument against it. The radar that Tesla removed in 2021 and the imaging radars I was qualifying on a robotaxi platform two years later are barely the same category of instrument, modern imaging radar delivers elevation, dense point clouds, and enough angular resolution to separate a pedestrian from a parked car. The “radar can’t tell a bridge from a truck” objection was true of the sensor class it was aimed at, and is expiring with it. Meanwhile the objection’s own premise (that cameras plus compute improve on a steep curve) applies to radar too, because the radar processing chain from DOA estimation through tracking is exactly as amenable to better algorithms and ML as the camera pipeline. I have built both chains. Neither one is standing still.

What I would actually decide, and how

If I owned the sensor architecture decision on a new program tomorrow, here is the position I would defend in the room.

For supervised L2 with a robust driver-monitoring system, camera-primary is a legitimate architecture, but I would still carry a front radar for the high-speed longitudinal cases and the FMVSS 127 darkness scenarios, because the cost of one well-integrated front radar is small against the validation burden of proving a camera-only stack in those corners. For anything hands-off, eyes-off, or driverless, decorrelated sensing is non-negotiable, and the debate should move from “whether radar” to “which radar and how fused.”

And whichever way a program goes, the real lessons of the camera-only era should stick. Do not carry a sensor you will not invest in: a radar fused as an afterthought is a false-positive generator with a supplier contract attached. Budget the factory cost of every sensor honestly, calibration stations, qualification campaigns, and supplier management are where sensor-count decisions actually bite, and they are routinely missing from the trade study. Size the suite by the hardest mandated test in the worst condition, not the sunny-day demo. And treat “cameras are enough” claims as a question about the safety case, not the sensor: enough for what feature, in what ODD, with what fallback? Asked that way, the argument and its rebuttal usually turn out to agree on the engineering, and disagree only about which product they think they are building.

Radar keeps surviving because physics is patient. The camera-only argument keeps returning because economics is persistent. A good perception architect respects both, and writes a safety case that doesn’t depend on either side of the debate being right.