AI & Autonomy

The central bet in robot learning right now is that manipulation will follow the trajectory language did: train a large model on enough demonstration data and general capability emerges. Vision-language-action models, which map camera input and an instruction straight to motor commands, are the clearest expression of that bet.

The obstacle is data. There is no internet-scale corpus of robots doing physical work, and the substitutes all have costs — teleoperated demonstrations are slow and expensive to collect, and simulation is cheap but transfers imperfectly to real hardware. Much of the field's current effort goes into narrowing that gap.

Autonomous driving is worth watching as the closest precedent: a decade of well-funded work showed that the last few percent of reliability is far harder than the first ninety, and that supervised demonstrations tend to look much better than unsupervised deployment.