Solving the Hardest Problem in Robotics | Fei-Fei Li with a16z

The gist

1. The main bottleneck in robotics is not hardware but data, and simulation can generate training data far faster than real-world collection.
2. World models need to stay geometrically consistent across viewpoints and interactions, because video-only models often make objects disappear or behave unrealistically.
3. Simulation lets teams test robots faster and more safely by covering lighting, friction, and other variables without physical risk or cost.
4. Robotics should move from fully structured factory settings to semi-structured warehouses before tackling homes, making humanoid robots premature for now.