Placeholder — an idea I’m developing.
A drone inspecting a wind turbine, a manipulator on a cluttered shelf: neither can assume the world at deployment looks like the world at training. The cost of treating learning as a one-time event shows up as the gap between an impressive demo and a system that survives contact with reality.
The argument I want to make here connects that gap to the models I work from: that embodied autonomy needs learning across many time scales at once.