The DARPA challenges offer a useful playbook for physical AI. They show how an impressive technical demonstration can help create an industry.The first DARPA Grand Challenge for autonomous vehicles took place in 2004. DARPA launched the Robotics Challenge eight years later, in 2012, with the Finals taking place in 2015. It is tempting to conclude that physical AI is therefore eight or eleven years behind autonomous driving.Physical AI can reuse a lot of what autonomous driving has already built: simulation, data engines, validation methods, safety engineering and fleet operations. But robots also face different problems. They have more varied bodies, much less real-world training data, and a far wider range of physical interactions.And in a factory (the lens I'm using), there is another constraint: the robot is part of a production system. If it stops for too long, the line may stop with it.So the lesson from autonomous driving is not when physical AI will arrive. It is how a difficult physical technology moves from an impressive demo to something reliable and economically useful.A year later, five vehicles completed the course. By 2007, DARPA had moved the challenge into an urban environment with intersections, traffic rules and other vehicles.Launched in 2012, it was motivated in part by the difficulty of sending people into disasters such as the Fukushima nuclear accident. Robots had to eventually perform tasks such as driving a vehicle, opening doors, turning valves, using tools, moving through debris and climbing stairs. At the 2015 Finals, 23 teams competed and only three completed all eight tasks.Physical AI does not start where autonomous driving started in 2004. Today we have better compute, neural networks, foundation models, GPU simulation and cloud infrastructure. We also have twenty years of experience building and operating autonomous machines.Physical AI is also much broader than factory automation. It will affect logistics, construction, healthcare, homes and many other environments.But I will often use the factory as the reference point in this article. It is the environment I know best, and it is also a useful reality check. In a factory, a robot eventually has to do more than demonstrate a skill. It has to meet a cycle time, work safely around people, recover from failures and create an economic return.Autonomous-driving companies learned early that making a vehicle drive once was not the hard part. The hard part was dealing with everything that happened after the first successful demonstration.They built massive simulation and testing systems. A failure found on the road could be replayed thousands of times. New software could be tested against libraries of known scenarios before being deployed again.