Rodney Brooks: Robotics | Lex Fridman Podcast #217

Brooks emphasizes Moravec’s paradox: evolution spent hundreds of millions of years on sensorimotor skills, so what looks ‘easy’ (vision, movement, everyday manipulation) is actually vastly harder than chess or Go. Current AI leans on static datasets and labels, but lacks the rich, grounded, continuous perception that even a 16‑month‑old shows opening a window for the first time.

Tracing from Napier and Kepler to Turing and Knuth, Brooks argues that what we call computation is a human-invented model—optimized to what people with paper could do and what silicon can implement—not a fundamental property of the universe. Treating brains or the cosmos as ‘just computation’ is a metaphor, not a demonstrated fact, and may be misleading our search for intelligence.

Brooks finds ImageNet gains, AlphaGo, and especially AlphaFold surprising and valuable, but stresses they work in tightly constrained domains with full information and short state descriptions. They generalize poorly outside training conditions (e.g., Go on different board sizes) and solve labeling/optimization problems, not the grounded common sense humans use to navigate open worlds.

He notes autonomous cars have driven highways since the 1980s and demos are not new; what’s unsolved is robust behavior in messy urban environments, rare scenarios, and socially acceptable failure rates. Society tolerates 35,000 human road deaths per year but may not tolerate even tens of deaths from machines, and meaningful deployment likely requires infrastructure changes (e.g., protected lanes, redesigned streets).

iRobot’s Roomba succeeded by hitting a $200 price point and solving a clear consumer problem; many other home robots failed despite technical sophistication. At Rethink Robotics, shifting from a $3k force-controlled arm aimed at new users to a $25–35k arm competing with incumbent industrial arms misaligned with customer expectations and contributed to commercial failure.