Waymo: Seven Lessons from Moving from Demo to Physical AI (Category Robotics)
I saw it. speech Dmitry Dolgov at Y Combinator Startup School 2026This is the difference between a spectacular demonstration of AI and a system that can be trusted in the physical world. Dolgov is one of the founders of the Google Self-Driving Car Project. 2009 year and became Waymo in 2016-M, and now co-CEO company. Prior to Google, he was an autonomous driver at Toyota and worked for the Stanford Racing Team on the Junior car for the DARPA Urban Challenge. 2007. He graduated from MIPT and received a PhD in computer science from the University of Michigan. For almost two decades, it has gone from planning algorithms to robotaxis.
The main frame of the report: for physical AI, the slogan “move fast and break things” does not work – the error does not have an undo button, and the price is not measured by tokens. There are three differences between digital AI:
- Solutions are needed in milliseconds
- There is no ready-made “internet of the physical world” with marked data
- Safety must be proven before mass launch
Dolgov calls the following points seven lessons. In fact, they are folded into an engineering sequence.
1Demo is the maximum. 1percentage In 2009–2010 Over the years, a team of about twelve engineers has passed. 100 Thousands of autonomous miles and ten routes 100 miles without human intervention. The demo took over 18 months away 500 thousands of trips a week - about 15 years. Each successive order of reliability required new architectural solutions, redundancy, and verification of the long tail of rare events.
2You need to choose not the fastest start, but a technological curve that will reach the desired reliability. Waymo combines cameras, LiDAR and radar: in the LiDAR report, the camera earlier notices a pedestrian in a dust storm and children in the dark, and redundancy allows the car to return to the depot when the branch closes part of the sensors. And don’t tie the architecture to today’s iron price: Waymo is already in its sixth generation.
3️ You need to be able to ride not one wave of AI, but several in a row. Waymo rebuilt Driver around CNN (wind-network)Transformers, now VLM and world models. The difficulty is not in the research demo, but in the transfer of technology to products critical to safety, without regression and at the same time simplification of the stack. The Waymo Foundation Model combines perception, scene understanding, forecasting and planning.
4️⃣ "Bitter LessonRichard Sutton works, but the structure doesn't necessarily contradict it. Waymo uses structure-augmented end-to-end: learned representations are complemented by explicit objects, semantic features, and a road graph. The structure doesn’t dictate the machine’s response, but it does provide real-time security checks and strong signals for reinforcement learning and evals. Something like a story with deterministic Sonar tests and ontologies
5Physical AI simulator is not an auxiliary tool, but a second big AI product Open cycle (open-loop) The test asks "What would you do?", a closed cycle. (closed-loop) It shows how action changes the world. The Waymo World Model simulates both movement participants and sensor readings. So you can check synthetic rarities: a car stopped on the highway, an airplane on the road, snow on the Golden Gate Bridge or an elephant at an intersection.
6Around the model you need a flywheel of three AI systems: Driver works, Simulator reproduces the world, Critic evaluates the result Real trips make the simulation more accurate, it generates difficult edge cases, the critic scores, the Driver learns and gets back on the road. Data becomes an advantage not as a warehouse of records, but as a closed production loop.
7The main moat is not a model, but evals and metrics If the team can’t quantify good enough, it keeps improving the demo. For a robot, the checks should cover the model, sensors, computing platform, behavior, external services, and operations. In Waymo, this role is played by the Safety and Readiness Framework. According to company statistics, 220 Millions of fully autonomous miles by the end of March 2026 The Waymo Driver has been involved in serious or fatal accidents. 94Less than expected for drivers in the same areas. This is data from Waymo, but the principle is important: trust is built on verifiable operation, not architecture from the slide.
Theses that I took for myself from history The best moment of physical AI often looks like nothing happened: the system worked out a dangerous situation, and passengers did not notice it. The number of “nine” required is a product choice: it predetermines sensors, redundancy, simulation, budget, and speed to market. The “end-to-end” or “modular” argument is too harsh. Waymo’s practical answer is a learnable general model plus just as much structure as is needed to scale and validate. Safety and evals are part of the product and a cumulative asset. The model can be replicated faster than fifteen years of data, operations, and proof of reliability. A flywheel without metrics can rotate anywhere. Data gives an advantage only when the critic distinguishes progress from beautiful movement on the spot.
#Robotics #AI #Engineering #Architecture #Product #Bigtech