Boston Dynamics

Contribution to AI

Balance can be treated as a fall that is continuously interrupted rather than a posture to be held, and the machines built on that premise looked nothing like the careful, statically stable walkers that preceded them. Marc Raibert had shown with hopping machines at the MIT Leg Laboratory that the problem decomposed into hop height, forward speed and body attitude; the company he spun out in 1992 spent a decade on human-simulation software before turning that control theory into field hardware. BigDog, funded by DARPA from 2002 and running by 2005, was the first of these robots to work outside a laboratory — over rubble, mud and ice, righting itself after a shove. It also failed usefully: its petrol engine was too loud for the infantry role it was built for. The videos became a second contribution, setting public expectations of what machines could physically do a decade before language models did the same for text. In 2013 Atlas gave the DARPA Robotics Challenge teams a humanoid body they had not had to build, so software groups could work on whole-body control directly. Spot's sale in June 2020 at a published price put a number on a category that had only ever existed as demonstrations. The recent turn matters most for AI: hardware reliable enough to survive thousands of trials has become the substrate for learned policies, including language-conditioned diffusion models trained with Toyota Research Institute.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).