What this covers
Marc Raibert and Lex Fridman discuss the engineering of legged robots and the path toward machines that combine physical grace with reasoning. Raibert, who founded Boston Dynamics, centers the conversation on Dynamic Locomotion—the principle that natural robot movement emerges from predictive control working in concert with intelligent hardware design, rather than from reactive sensing alone. The talk traces how this philosophy shaped Boston Dynamics' most visible projects, from early experiments with hopping robots to the climbing feats of Atlas and the practical deployment of Spot, while examining why the cognitive side of robotics has lagged despite physical breakthroughs.
The range moves fluidly between technical specifics and broader strategic questions. Raibert explains passive dynamics, spring compliance, and model predictive control as foundations for efficient motion, then shifts to the build-break-fix testing regimen that turned spectacular failures into working behaviors—the Atlas stair-climbing video, for instance, required 109 attempts over six weeks. He addresses why warehouse automation is currently the only robotics segment making real money, and why indoor deployment proves harder than wilderness navigation. A recurring theme is Athletic vs Cognitive Intelligence: while Boston Dynamics solved the athletic side, most robots remain cognitively impoverished, needing extensive manual programming. Raibert also touches on YouTube's unlikely role in Boston Dynamics' rise, the economics of early DARPA funding, and why hydraulic power supply—often dismissed as antiquated—still offers advantages when redesigned for modern control. He argues that future robots must learn by observation, moving toward what he calls the fusion of physical capability with genuine understanding of novel situations.
Raibert argues that great robotics comes from a dynamic, athletic approach to legged locomotion—built on innovative hardware, real-time predictive control, and intrepid build-break-fix testing—and that the next frontier is fusing this 'athletic intelligence' with cognitive intelligence so robots can learn by watching humans.
- Natural, lifelike movement requires good hardware plus forward-looking dynamic prediction, not just reactive sensing
- Robots advance through intrepid testing where failure and repair are part of the paradigm
- The biggest current barrier to commercially viable robots is the cognitive side, not the physical side
Imperfect robots learning through failure are compelling because they mirror human vulnerability.
- A robot being terrible at a task and slowly becoming less terrible is extremely compelling and inspiring—we like flaws, and a robot learning, falling, being clumsy and confused acts as a mirror that reveals what makes humans special.
“I also do think sucking at a task is also compelling. The incremental improvement. A robot being really terrible at a task and then slowly becoming better.”
Quadruped robots like Wildcat achieve remarkable speed despite mechanical drawbacks.
- Wildcat, a quadruped along the way toward Spot, ran 19 miles per hour on flat terrain, possibly the fastest quadruped in the world, though it was also extremely loud due to its racing go-kart engine.
“Wildcat, which was one of along the way towards Spot, which is a quadruplet that went 19 miles an hour on flat terrain.”