Raibert, Marc
▶ Watch on YouTubeCore ideas
Raibert traces his path into robotics to a 1974 moment at MIT, following Berthold Horn back to his lab and seeing a disassembled robot arm in pieces on a table — an image that redirected him away from single-neuron neurophysiology (which he felt was too disconnected from questions of control and thought) toward robotics and AI. His entire career, he argues, has rested on a contrarian philosophy: rather than build cautious, statically-stable robots (like the tripod-gaited six-legged robot he saw at a 1970s biomechanics conference, which he found unlike any real animal), he pursued "aggressive," dynamically-balanced, under-actuated machines — starting with a single-legged pogo-stick hopping robot funded by a chance DARPA hallway pitch (Ivan Sutherland walked him in to see program manager Craig Fields, who funded it for $250,000 in 1980) that had no practical goal beyond understanding the fundamentals of animal locomotion. He describes building successive generations of legged robots at Carnegie Mellon's and MIT's "Leg Labs" — including the first robots to perform 3D somersaults, with graduate student (and former champion gymnast) Ron Robert helping translate real athletic technique into control algorithms — while stressing that even elite human athletes and gymnasts typically can't articulate the physics of what their own bodies are doing.
At Boston Dynamics (founded 1992 as a physics-simulation spinout of the MIT Leg Lab, initially building tools unrelated to physical robots, including a force-feedback surgical training simulator that the team killed once they realized surgeons expected to be paid to help train the system rather than pay for it), Raibert describes rediscovering the company's true calling through a Sony collaboration (making a running version of the AIBO robot dog) before building BigDog under a DARPA-funded biomechanics program — the first robot to integrate power (a gasoline engine converted to hydraulic actuation), computing, and legs entirely onto one mobile platform, tested by hiking DARPA's "Guadalcanal Trail" at Quantico. He traces the lineage from BigDog to the load-carrying LS3 (rated for 400 lbs, once tested carrying nearly 1,000) to an all-electric prototype built at Larry Page's request (during Google's ownership of Boston Dynamics) for a lighter, less intimidating 60-lb robot — which became Spot. He credits deep hydraulics innovation (custom valve and power-supply designs far beyond 1950s aerospace-derived technology, achieving five kilowatts of power from a five-kilogram unit) as essential to performance, arguing hardware innovation is far from a solved problem despite conventional wisdom to the contrary, and explains that Atlas's naturalistic movement comes from forward-looking dynamic prediction (recalculating roughly 1-2 seconds ahead for walking/obstacle navigation, and calculating full rotational/momentum trajectories in advance for maneuvers like somersaults) rather than backward-looking correction — while noting that, ironically, robotic running looks more human than robotic walking, since walking depends on subtle passive-dynamic compliance (springiness in joints and structure) that remains harder to replicate than dynamic, "controlled falling" gaits.
Raibert now leads the Boston Dynamics AI Institute, which he frames around a distinction between "athletic intelligence" (mechanical design, real-time control, energetics — where he believes Boston Dynamics set the standard) and "cognitive intelligence" (planning, perception, and generalization — which he argues is the current bottleneck holding back even physically capable robots from broader commercial usefulness). Flagship institute projects include "Watch, Understand, Do" (robots learning tasks, starting with simple bicycle repair, by observing humans without an explicit pre-built model of the task or environment — analogous to how a person can navigate an unfamiliar room without first building an explicit map) and "Inspect, Diagnose, Fix" (extending Spot's existing data-collection capabilities, like thermal imaging and gauge-reading, into actual autonomous fault diagnosis and repair). He describes the institute's guiding methodology as "stepping stones to moonshots" — requiring visible incremental progress at least annually even while pursuing multi-year target goals — and is candid that traditional model-predictive control still outperforms pure machine-learning approaches for most of Boston Dynamics's most impressive athletic feats (like the celebrated three-step Atlas jump, achieved only after 109 recorded attempts over six weeks), while expecting eventual convergence between the two paradigms, citing the institute's new Zurich office led by reinforcement-learning researcher Marco Hutter.
He identifies four qualities essential to a strong robotics team — technical fearlessness (tackling problems with no known solution), diligence (building genuinely robust systems, illustrated by Boston Dynamics's famous videos of engineers shoving Spot or yanking ropes to test resilience rather than curated "perfect run" footage), intrepidness (tolerating repeated, filmed failure as a normal part of the process — Atlas is deliberately engineered to survive frequent real-world falls, and the team budgets for spare parts and repair crews rather than keeping fragile "gold-plated" prototypes on a shelf), and technical fun (engineering as a rare combination of scientific rigor, creative craft, teamwork, and pay). On Boston Dynamics's famous videos, he says he personally did final editing for years, following a strict "no explanation" philosophy — show only things worth showing, without titles or narration slowing down the raw demonstration — while acknowledging retrospective criticism (e.g., BigDog videos not disclosing a human operator) that shaped how the team communicates its work going forward.
On competition, Raibert expresses admiration for Elon Musk's technological ambition (Tesla, SpaceX) and suggests Atlas may have partly inspired Tesla's Optimus, while noting Optimus isn't yet at Atlas's level and other humanoid competitors (Figure, Apptronik) have strong teams; he argues competing quadruped and humanoid companies actually help Boston Dynamics by normalizing the category itself ("which quadruped do I want" rather than "do I even want one"), and predicts costs for legged robots have far to fall, with warehouse logistics currently the only clearly profitable robotics use case. On AGI and existential AI risk, he pushes back on doom-focused framing by analogy — noting he isn't threatened by people smarter than himself, and that computers being smarter needn't imply loss of human control, comparing the risk/benefit calculus to cars (which kill roughly 1.25 million people annually worldwide yet remain broadly beneficial and are becoming safer over time). He closes with practical career advice (imagine what you'd do with no resource, skill, or opportunity constraints, then work backward to what's realistic) and a simple personal philosophy: "you have to have fun while you're here... it would be a waste not to."
Excerpts
On the conference talk that first inspired him to pursue dynamic, "aggressive" legged locomotion instead of cautious statically-stable walking:
“One talk was by a robotics guy, and he showed a six legged robot that walked very slowly. It always had at least three feet on the ground... And I just looked at that and said, wow, that's wrong. That's not anything like how people and animals work because we bounce and fly... I definitely got inspired to say, 'Let's try the opposite.'”
On why Boston Dynamics deliberately filmed engineers disrupting Spot mid-task rather than only showing clean successes:
“If you just show a video of it hopping, it's a so what. If you show it falling over a couple of times, and you can see how easily and fast it falls over, then you appreciate what the robot's doing when it's doing its thing... the reaction you just gave to the robot getting interfered with or tested while it's going through the door, it's showing you the scope of the solution.”
On why the famous three-step Atlas jump took 109 recorded attempts, and what that reveals about the team's culture:
“It took 109 tries and we have video of every one of them, we shoot everything... it's not like we're cheating by just showing the best one, but we do show the evolved performance... intrepidness... might be the most important ingredient. Robotics is hard, it's not going to work right right away, so don't be discouraged, is all it really means.”