The Economic Case Against Camera-Only Autonomy, From Waymo's Co-CEO
This week, Waymo co-CEO Dmitri Dolgov made his company's most forceful public case yet against camera-only autonomy. The argument blends engineering and economics: the roughly $60,000 multi-sensor package installed in every Waymo vehicle produces a safety record that, in Dolgov's view, justifies its cost against cheaper vision-only rivals.
Dolgov did not name Tesla, but his technical points pointed straight at it. A camera is a receiver. It captures light and has no way to measure motion directly, and its image quality drops in darkness, glare, rain, and snow. Lidar and radar generate their own signals instead of relying on outside light. Radar has a specific advantage: Doppler returns give a direct velocity reading that holds up in fog and storms. These are ordinary conditions that show up on regular roads. Low sun, darkness, and heavy precipitation are routine. Waymo's suite spreads the work: cameras add color and resolution, lidar supplies 3D structure, and radar cuts through weather and measures speed. The design also tolerates failures. A leaf or branch covering one lens disables a camera, but the vehicle still has other sensor types to navigate.
The argument's quantitative core is a reliability ladder Dolgov calls the “nines” problem. Vision-only perception, he says, can reach 90 to 99 percent reliability, a range suitable for driver assistance but not for full autonomy. Full autonomy requires the next nines, and each one is much harder to add than the last. Moving from 99 percent to 99.9 percent is a different engineering task. Dolgov's claim is that passive perception runs out of headroom before crossing that threshold.
The numbers show why this distinction matters in practice. At Waymo's current scale, roughly 500,000 paid trips per week, a 99.9 percent success rate still produces hundreds of failures weekly, and full autonomy must handle those without a human backup. Dolgov treats reliability as a statistical requirement rather than a feature checklist. The bar is a failure rate that survives daily volume. Demo rides cannot demonstrate such a rate.
The cost side of the argument is easy to miss but central to the strategy. A complete Waymo package, including vehicle, sensors, and driver software, costs about $60,000, and that price has dropped with each generation; the current system is the sixth. The investment is amortized across real revenue. Waymo runs about 500,000 paid trips per week in 15 U.S. cities and targets 1 million weekly rides. At that scale, an expensive sensor stack becomes a business rather than a research budget.
The Data Behind the Camera-Only Autonomy Debate
The empirical record draws the sharpest line between the two approaches. Waymo cites an IIHS study that puts its police-reportable crash rate 68 percent below the human-driver baseline. Tesla's own robotaxi data run the other way: a crash rate about three times worse than human drivers, measured with a safety monitor in the front seat, and 14 crashes logged over roughly 800,000 miles of robotaxi service. The IIHS metric is the one insurers and regulators already use, so Waymo's figure has a comparability that internal test data lacks. Its stated target of 1 million weekly rides would extend that record; each mile either reinforces the approach or reveals its limits.
That comparison flatters the vision-only stack. Tesla collected its crash data with a paid human fallback in the front seat; Waymo's 68 percent below-human figure comes from fully driverless operation. The raw numbers understate the gap between the two architectures.
| Metric | Waymo (cameras, lidar, radar) | Tesla robotaxi (cameras only) |
|---|---|---|
| Full package cost | About $60,000, falling each generation | Not disclosed |
| Operating scale | ~500,000 paid trips per week, 15 U.S. cities | ~800,000 miles of robotaxi service |
| Crash rate vs. human drivers | 68% lower (IIHS, police-reportable) | About 3x worse (own data, with safety monitor) |
| Reported crashes | Not disclosed in these figures | 14 in ~800,000 miles |
The leading counter-argument deserves a fair hearing. Tesla's wager is that neural networks trained on millions of consumer vehicles can use two cameras to get what Waymo gets from lidar. Humans drive with two eyes, the logic runs, so perception hardware cannot be the binding constraint; the software is. The cost logic is real as well. A vision-only stack is cheap enough to install in every car Tesla sells, generating training data at a scale Waymo cannot match with a fleet concentrated in 15 cities. The two-eye analogy is also weaker than it sounds: human drivers draw on a lifetime of learned context, while a vision system must generalize from a training distribution, a gap that no sensor count changes. It is the strongest version of the case, and it is why the debate will not be settled by sensor preference alone.
That argument breaks down on the only metric both companies claim to care about: the crash record. If camera-only autonomy were enough, public data would show it approaching parity with multi-sensor systems. The data do not. Tesla's own robotaxi figures show a crash rate roughly three times the human baseline, with a monitor in the front seat, while Waymo sits 68 percent below it. That gap defines the product, and it runs against the vision-only architecture.
Regulation turns the sensor debate into a market one. If agencies license operators on verifiable safety records, Waymo arrives with weekly volume figures, a third-party study, and accumulated miles. A vision-only operator would meet the same regulators with a crash rate three times the human baseline. That asymmetry, more than the lidar unit, makes the $60,000 package a defensible cost rather than a handicap.
Cost trends push the same direction. Each Waymo hardware generation has cut the bill, shrinking the price gap with vision-only systems, while the safety data tied to the expensive sensors keeps accumulating. The trend line matters more than today's price. If later generations keep lowering cost while the safety record lengthens, the economic case for camera-only erodes without anyone having to argue it. No challenger betting on cheap perception plus more data has published a dataset showing that convergence.
For decision-makers, the concrete takeaway is to watch crash-per-mile disclosures rather than demo videos; that is where the competition gets decided. Waymo's next milestone is the push toward 1 million weekly rides, roughly double its current volume, which would extend the statistical base of its safety claims. A challenger claiming full autonomy on cheaper sensors carries the burden of proof, and the first public data points from the vision-only camp cut against it.
Why This Matters
Whoever can document safety at scale will win the autonomy race, and today only one side has the numbers to do that. Dolgov has turned a technical critique into a regulatory and economic argument. If verifiable safety records become the licensing standard, camera-only autonomy enters the market carrying a burden its own data has not yet met. Crash-per-mile disclosures will show whether that burden lifts.
Related Articles
- Waymo Safety Study Proves Autonomous Vehicles Are Safer — With One Critical Caveat
- Waymo Launches Reference Driver Model to Benchmark Autonomous Safety Against Human Behavior
- Waymo Freeway Return Tests Safety Data Against New AV Rules
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.