The Camera-Only Fight Is Back, and Robots Are Picking Sides Too

Sterling Anderson ran Tesla Autopilot back when it still had radar. He went on to co-found Aurora and is today General Motors' chief product officer, the one who makes the call on what sensors are put in the firm's first eyes-off vehicle. This week he had some words for the camera-only method his former company is so wedded to, calling it "a weird religious stance that ignores physics." Silicon Valley spent a decade wrangling over which sensors to use, and the fight did not so much end as go quiet while the robotaxis rolled out. And for once the ones with the most to lose are not carmakers, they are the builders of robots.

Physics, radar and a ten-minute countdown

Anderson made his case, without much subtlety, in an interview with The Verge published October 27, a week after GM's technology event in New York. Cameras are no match for dense fog and they do not gauge an object's instantaneous velocity as radar does by way of Doppler. Radar, he added, "is so cheap that dropping it doesn't make sense."

He made a quieter point about lidar. Since it works at a wavelength near visible light, it can co-train the camera systems and get them up to speed quicker than they could on their own. In that light, lidar is a training tool as much as a backup.

All of this is in service of a highway-only, eyes-off variant of Super Cruise. GM intends to put it in the 2028 Cadillac Escalade IQ with its complement of lidar, radar and cameras. A driver who has fallen asleep is never asked to take over in a rush; the system issues a ten-minute countdown and, absent any response, pulls off at the next exit and stops.

Tesla's bet, and what NHTSA found

Tesla has gone the other way. The company removed radar from the Model 3 and Model Y in May 2021, did away with ultrasonic sensors in 2022 and has insisted for years that eight cameras and a neural network are all a car needs. Elon Musk dismissed lidar as a fool's errand in 2019 and now the company's camera-only Cybercabs are making their way into driverless service in Austin.

Regulators have not been so lenient. In March 2026 the National Highway Traffic Safety Administration made its investigation of Full Self-Driving in poor visibility an Engineering Analysis, the final move before any recall is decided on some 3,203,754 vehicles. The probe has turned up nine crashes, one being a pedestrian fatality in November 2023. The agency was precise in its finding: Tesla's means of detecting camera degradation failed to pick up on such everyday impediments to visibility as sun glare, fog or airborne dust, and the driver was put on notice only moments from a collision. Those are, almost word for word, the conditions Anderson says physics will not let a camera handle.

Waymo says weak sensing flattens the safety curve

Waymo's data set is the other major part of this argument. The company puts lidar, radar and cameras on all its vehicles and has put in more than 200 million miles with only a rider on board. Co-CEO Dmitri Dolgov set out the company's position at Y Combinator's Startup School this summer, and he did not pretend the question is new.

His point is what happens after the demo works. Cameras can match a human driver and underpin a fine driver-assist product, he argued, but reliability has to climb an exponential ladder of nines, and a quick early result can mask a ceiling.

Dolgov also took on the cost objection. Waymo is on its sixth generation of hardware, each considerably cheaper than the last. The latest safety report from Waymo, built on 220 million rider-only miles, claims 94 percent fewer serious-injury crashes than human drivers would have caused over the same distance.

The third camp: it depends what you are building

Some are not inclined to take a side. Jiajun Zhu, co-founder and co-CEO of Nuro, put it in writing in March: the vision-only faction has its case to make with lower cost, quicker iteration and greater scalability, while the redundancy camp is no less well argued.

It may be what most of the industry thinks in private. A robotaxi that puts revenue in the bank every hour it is running has a different cost ceiling than a consumer car sold by the million; a walking pace delivery robot does not present the same risks as a 65 mile per hour truck. The question is seldom whether lidar is any good. It is whether the product can afford it.

Robots are running the same experiment with a smaller budget

That is where the debate is heading, and it is why the next round will not be settled in Austin or Phoenix. Humanoid robot makers have mostly gone camera-first, partly for cost and partly on the premise that a robot with human proportions ought to see as a human does. Stereo camera modules with a 60 millimeter baseline, akin to the space between two eyes, are now standard equipment. The outdoor machine is another matter. A sidewalk delivery robot, a yard tractor or a mining haul truck contends with dust, ice, vibration and direct sun, the very conditions NHTSA flagged, often in settings where lidar is too delicate for the environment or too expensive for the unit economics.

So the dispute in the robot world is less about adding a sensor than about how far a camera can be pushed. The camera module itself has become the battleground: it must have a 120 decibel high dynamic range to make it through a change from sun to shadow, global shutter synchronization for the two lenses to record the same moment, IP69K sealing to stand up to pressure washing and the ability to function at anywhere from minus 40 to 85 degrees Celsius.

There is as much of a Taiwan story here as a Silicon Valley one. oToBrite, the Vision-AI firm out of Hsinchu Science Park, has put over 1.2 million automotive cameras in the hands of Daimler Truck, Toyota and Xpeng among others. Today the company is selling GMSL2 camera modules and 3D stereo vision modules built to automotive grade for robots and unmanned vehicles. Its long-range stereo module has a 120 millimeter baseline for depth out to 20 meters, while the GMSL2 repeater announced in July doubles a camera link's reach to 30 meters for boom arms, harvesting heads and trailer-rear positions on big autonomous machines.

From January 6 to 9 oToBrite can be found at booth 4577 in the Taiwan Mobility Pavilion in the LVCC West Hall at CES 2027. For teams deciding how far cameras alone can carry their own trucks, robots or humanoids, the company is taking meetings, and engineers and sourcing leads can request time with the oToBrite team ahead of the show. On its second day the Tech Theater in the pavilion, a Sparknify production, will address the issue of where Vision AI's next market lies and why the answer will not be passenger cars.

Will physics or price decide the next five years?

The near-term outlook is easy to sketch because the dates are public. In 2028 GM will have its eyes-off Escalade on the market, fitted with three kinds of sensor. Tesla's camera-only robotaxis are already running and will either scale or stall on the very visibility issues NHTSA is looking into. Before the decade is out there should be enough driverless data from both sides to settle the matter for cars with crash rates per mile rather than sound bites. For Silicon Valley, the interesting part is that the answer will likely come from outside the car business. Robot fleets rack up miles in ports, mines, farms and warehouses where the operator, not a regulator, decides what is safe enough, and where a lapse in perception costs a shift, not a life. Those operators are already voting with purchase orders.

The risks run both ways. If Anderson is right, a camera-only fleet carries a tail risk that shows up in fog, glare and dust, rarely and then all at once. One bad week of incidents could turn public opinion on driverless cars in a way no safety report would fix. If the vision camp is right, the firms that spent a decade on lidar will have built the autonomy equivalent of a mainframe, superb and unaffordable, while cheaper fleets take the market. Either way the cost curve is doing more work than either side admits: Dolgov's own point that lidar gets cheaper every generation cuts against the idea that anyone has to choose now.

The controversy is also personal. Anderson, a former head of Tesla Autopilot, is publicly calling his old program's strategy unphysical. Musk has spent years calling lidar users fools. Both sides have a product to sell, so every quote in this article is a stance and not a conclusion. Critics of the multi-sensor camp say fusing three sensors is a harder software problem than reading one, and that redundancy adds failure modes of its own. Camera advocates point to the obvious: humans do it. Lidar partisans retort that humans also have accidents.

The signals to watch are concrete. Whether NHTSA closes its engineering analysis with a recall or without one. Whether Tesla's Austin fleet publishes per-mile crash figures that hold up at scale. Whether GM is able to deliver the 2028 Escalade as scheduled with lidar intact or if it pares down the sensors to make the numbers work. In the realm of robots, whether the makers of humanoids with only cameras put some depth sensors on them as they leave the demo floor for the outdoors. The CES booths in January will give an early read, since the camera suppliers, the lidar makers and the robot companies all sit within a few hundred feet of each other.

The sensor war never ended; it just moved to a bigger map.

Subscribe to Sparknify for Silicon Valley tech news and the people behind it.