Browse
Log in / Register
Share
Today, March 28, 2025, Waymo — the Alphabet-owned robotaxi company — revealed it has logged over 200 million miles of fully autonomous driving across major U.S. cities. In a blog post, vice president of onboard software Srikanth Thirumalai distilled the experience into 10 fundamental truths about building AI for self-driving vehicles. The lessons, drawn from real-world operations in Phoenix, San Francisco, Los Angeles, Nashville, and Austin, offer a rare look at how one of the most advanced autonomous fleets thinks about artificial intelligence.
While Tesla relies on cameras alone for its end-to-end neural network, Waymo’s 200 million miles have taught it that multi-modal sensor fusion is indispensable. Radar, lidar, and cameras each cover weaknesses the others have — for example, lidar handles low-light environments where cameras struggle, and radar cuts through fog. The blog emphasizes that no single sensor type can reliably handle all edge cases, and combining them creates a safety net that pure vision systems lack.
Early autonomous vehicle development relied on specialized modules: one for pedestrian detection, another for car tracking, another for traffic light recognition. Waymo discovered this “modular spaghetti” becomes unmaintainable at scale. Instead, the company moved toward fewer, high-capacity models that can handle multiple tasks. This approach is more agile and easier to update as the fleet grows — a lesson that directly challenges the industry’s historical tendency to pile on small, narrow AI components.
Replaying recorded data is useful for debugging, but Waymo found that closed-loop simulation — where the AI interacts with a dynamic virtual environment — creates essential feedback loops. The AI can try out new behaviors, fail safely, and learn from millions of synthetic miles without risking real traffic. This is a key reason why Waymo’s models improve faster than those relying solely on logged drives.
Waymo’s fleet of more than 3,000 vehicles has generated over 500,000 paid rides, but the 200 million miles represent a massive dataset of rare edge cases — from jaywalking pedestrians at night to construction zones with shifting lane lines. Thirumalai noted that safety is a direct result of rigorous, real-world experience, not just code reviews. Every mile exposes the system to new scenarios that can’t be predicted in a lab.
The Insurance Institute for Highway Safety (IIHS) conducted a study comparing Waymo’s driverless vehicles with human drivers in four cities. On a per-mile basis, Waymo’s robotaxis were involved in 68% fewer crashes. While critics point out that AVs can still cause accidents, this data suggests the technology significantly outperforms the average human — a claim that gives Waymo leverage in regulatory debates.
Tesla’s push for a single neural network that learns “from pixels to steering” captures the industry’s imagination, but Waymo’s approach remains more modular. By keeping separate models for perception, prediction, and planning — but with fewer, high-capacity models per category — the company maintains interpretability. Engineers can pinpoint why a decision was made, which is critical for safety certification and debugging.
Waymo’s blog highlighted that what works for 10,000 miles might break at 100 million. The company had to redesign its AI stack multiple times to handle the scale of its operations. For example, model updates that were safe for a small test fleet could introduce subtle regressions when deployed across 3,000 vehicles running 24/7. Continuous monitoring and rollback capabilities became essential.
Despite the autonomy, Waymo still relies on human operators for remote assistance in edge cases, as well as for labeling data used to train models. The blog noted that the best AI systems are built with a tight loop between human expertise and machine learning. This pragmatic view contrasts with the “full autonomy with no humans” rhetoric sometimes heard in the industry.
Thirumalai concluded the blog by stating that these 10 lessons are just the beginning. With plans to expand to new cities and possibly highway operations, Waymo expects each new challenge to refine its AI principles. The 200 million miles are a milestone, not a finish line. As the company pushes toward billions, it will continue to test assumptions about sensor fusion, model architecture, and simulation fidelity.
Waymo’s transparency — publishing specific AI lessons — is rare in a sector where companies often guard their secrets. The insights provide a blueprint for other AV developers, especially those trying to catch up with the leader. They also give regulators a clearer picture of what safety-critical AI should look like: multi-modal, scalable, and validated through real-world miles plus simulation. For the public, the data reinforces that robotaxis are already safer than human drivers — a key stepping stone to wider adoption.
