The Safety Debate Over Camera-Only Autonomous Driving
Waymo Co-CEO Dmitri Dolgov stated on August 4 during a Y Combinator Startup School talk that autonomous driving systems relying solely on cameras face fundamental safety limitations and struggle to achieve the ultra-high reliability required for full autonomy. He emphasized that cameras, as passive sensors, perform inconsistently in low-light conditions, glare, or adverse weather—scenarios unavoidable in real-world driving. While cameras may meet the basic requirements of driver-assistance systems, Dolgov argued that integrating active sensors is essential to break through the safety barriers for true self-driving capability.
Dolgov’s remarks directly challenged Tesla’s commitment to a camera-only approach. Tesla has progressively removed radar and ultrasonic sensors from its vehicles since 2021, fully adopting its "Tesla Vision" camera-only system. However, Waymo’s latest safety report revealed that its autonomous vehicles, across 15 U.S. cities and over 220 million miles of driverless testing, reduced serious-injury crash rates by 94% compared to human drivers—equivalent to a 17-fold improvement in safety. In contrast, Tesla’s robotaxis in Austin, Texas, demonstrated a crash rate three times worse than human drivers and still required human safety operators for supervision.
The Technical Advantages of Multi-Sensor Fusion
Waymo currently employs three types of sensors—cameras, lidar, and radar—using independent encoders to fuse data and enhance environmental perception. Dolgov illustrated the advantages of multi-sensor systems with real-world examples: during a dust storm in Phoenix, Arizona, cameras were nearly ineffective at recognizing road conditions, while lidar could still clearly detect pedestrians at the roadside. On unlit roads at night, lidar accurately captured sudden appearances of children or animals. Additionally, cameras are vulnerable to obstructions, such as leaves or mud covering lenses, which could disable the system. Waymo vehicles have relied on other sensors to safely return to service centers when cameras were obstructed by tree branches.
Dolgov stressed that improving the reliability of autonomous systems follows an "exponential ladder of nines" principle—moving from 90% to 99% reliability is relatively straightforward, but each additional "nine" (e.g., from 99.9% to 99.99%) requires ten times more effort. He noted that driver-assistance systems may only need one or two "nines" of reliability, but full autonomy demands stacking multiple "nines" to ensure safety over millions of miles. Tesla CEO Elon Musk has repeatedly claimed that Tesla’s Full Self-Driving (FSD) system is safer than human drivers, yet crash data from Tesla’s robotaxis in Austin shows performance far below human driving standards.
Scale and Cost Considerations
Waymo currently provides approximately 500,000 paid rides per week and is working toward a target of 1 million weekly rides. In comparison, Tesla’s robotaxi service operates only in low-speed, geo-fenced areas in cities like Austin, accumulating just 380,000 miles of driverless operation over the past year—far below Waymo’s scale. Dolgov stated that Waymo’s sixth-generation hardware has significantly reduced costs, with the total price of the vehicle and sensor suite expected to approach $60,000, challenging Tesla’s emphasis on the cost advantages of a camera-only system.
Tesla has previously dismissed lidar as a "fool’s errand," but Dolgov argued that as sensor prices continue to drop, the cost advantage of camera-only systems will diminish. He emphasized that the true deciding factor is whether a system can scale effectively. Analysis also highlighted that Tesla’s robotaxis have been involved in collisions in dense fog on California’s I-99 highway, underscoring the vulnerability of camera-only systems in low-visibility conditions. Waymo, through multi-sensor fusion, effectively handles adverse weather such as rain, snow, and dust storms.
Dolgov did not entirely dismiss the value of cameras but drew a clear distinction between the technical thresholds for "driver assistance" and "full autonomy." He argued that while camera-only systems may suffice for the former, they cannot bridge the safety and reliability gap required for the latter. As Waymo continues to expand its service, Tesla’s camera-only approach will face stricter market validation and safety scrutiny.
