Analysis
Waymo escalated its long-running technical dispute with Tesla this week, with co-CEO Dmitri Dolgov laying out the company's clearest public case yet for why camera-only self-driving cannot reach full autonomy. In an exclusive given to Axios, Dolgov said "weak sensing" -- his term for a camera-only stack -- hits a safety ceiling long before it reaches superhuman performance, drawing on more than 200 million fully autonomous miles Waymo has now logged. Electrek characterized the post as Waymo calling Tesla's approach a "false summit": a path that looks like progress but tops out well short of the goal.
The technical claim is specific. Waymo says that in its own internal testing, removing lidar or radar from its sensor stack measurably degraded the system's visibility and object-detection reliability, even though cameras alone still "worked" most of the time. The company's argument is that self-driving safety cases are won or lost in the rare, ambiguous edge cases -- glare, fog, a pedestrian partially occluded by a parked truck -- where redundant sensing catches what a single modality misses.
How we got here
Tesla pulled radar from its production vehicles in 2021 and dropped ultrasonic parking sensors in 2022, consolidating entirely around "Tesla Vision" -- the bet that cameras plus a sufficiently capable neural network can replicate and eventually exceed human driving, since humans navigate almost entirely on sight. Elon Musk has called lidar "a fool's errand" and said companies relying on it are "doomed," a position he has held publicly since at least 2019. Waymo, spun out of Google's self-driving project and part of Alphabet, has run the opposite bet from the start: cameras, lidar and radar together, with high-definition pre-mapped routes as a further layer of redundancy. Pulse has tracked Waymo's rise through its 2026 funding round and robotaxi expansion.
The competitive landscape
- Waymo -- Alphabet-owned, fully driverless robotaxi service live in multiple US metros, 500,000+ weekly trips, more than 14 million paid rides in 2025 alone (roughly 3x its 2024 volume). Sensor suite: cameras, lidar, radar, HD maps.
- Tesla -- camera-only Full Self-Driving stack, preparing a wider rollout of its purpose-built Cybercab; robotaxi pilot has operated with a paid human monitor in the front seat rather than fully driverless in most markets so far.
- Zoox (Amazon) -- lidar-and-camera robotaxi with a bidirectional, no-steering-wheel vehicle design, live in limited markets.
- Baidu Apollo Go and Pony.ai -- China's largest robotaxi operators, both running multi-sensor stacks similar to Waymo's, scaling aggressively in Wuhan, Beijing and Guangzhou.
Every well-funded robotaxi operator besides Tesla has converged on multi-sensor redundancy, which is precisely the point Waymo is making publicly: this is not a two-sided technical debate inside the industry so much as Tesla holding a minority position against a broad consensus.
The numbers in context
Waymo's 200 million autonomous miles and 500,000 weekly driverless trips are real, measured operating data from a service with no safety driver. Tesla's FSD has accumulated vastly more total miles across its consumer fleet, but the overwhelming majority of those miles involve a human driver supervising and ready to intervene -- a fundamentally different risk profile than Waymo's zero-occupant-intervention model. The comparison Waymo is implicitly inviting -- its fully driverless mileage against Tesla's supervised mileage -- is not apples to apples, but it is the comparison that matters once Tesla's Cybercab tries to operate without a human in the loop.
What it means for founders and investors
The practical read for anyone building or backing autonomy, robotics or physical-AI companies is that sensor-fusion redundancy remains the industry's dominant safety architecture, and a camera-only bet carries real technical risk that a pure scaling argument -- more data, bigger models -- may not fully resolve. Lidar costs have fallen sharply over the past five years, weakening Tesla's original cost argument for going camera-only. Any startup pitching a camera-only perception stack for a safety-critical application should expect this exact question in diligence: what does redundancy buy you, and can you prove your system's edge-case failure rate without it.
The test that actually settles this argument is not a blog post from either side -- it is whether Tesla's Cybercab can run fully driverless, with no human monitor, at Waymo-scale trip volumes without a materially worse safety record. That data does not exist yet.
