Marc Raibert on Boston Dynamics, Legged Robots, and the Future of Robotics

Marc Raibert has been building legged robots for over 40 years — from a one-legged pogo-stick hopper at Carnegie Mellon, through MIT’s legendary Leg Lab, to founding Boston Dynamics and, later, the Boston Dynamics AI Institute. He walks through the engineering history behind BigDog, Spot, and Atlas, the philosophy of dynamic balance over cautious static stability, and why he now believes the hardest open problem in robotics is not athletic but cognitive.

Key ideas

1. Legged robots should bounce and fly, not creep. Raibert’s founding insight, from watching a slow, always-stable six-legged robot at a 1980s biomechanics conference, was that real animals stay upright by predicting and correcting their motion in flight, recycling energy through muscle and tendon rather than always keeping enough feet on the ground to be statically stable. That bet — dynamic balance over cautious multi-point contact — shaped every robot he has since built.

2. Dynamic balance decomposes into three solvable calculations. The original pogo-stick hopper reduced staying upright to: how much energy to add on each bounce (measured at the top of the hop), where to place the foot relative to the body’s centre of mass on landing (a pole-vaulter’s problem), and how much corrective torque to apply to body attitude while the foot is grounded — attitude cannot be corrected mid-air. Every later robot, quadruped or humanoid, builds on this same three-part structure.

3. Athletic intelligence is largely solved; cognitive intelligence is the bottleneck. Raibert splits robot competence into an athletic part (mechanical design, real-time control, energetics) and a cognitive part (planning, perception, judgement under incomplete information). He judges Boston Dynamics to have set the standard for the former; the AI Institute’s flagship project, Watch Understand Do, targets the latter — robots that learn a task by observing a human once, rather than being explicitly programmed for it. See Athletic Intelligence.

4. A strong robotics team runs on four traits: technical fearlessness, diligence, intrepidness, and fun. Fearlessness means taking on problems with no known solution; diligence means building a solution robust enough to survive an engineer deliberately sabotaging it (tugging a rope, pushing a door); intrepidness means tolerating failure at scale — Atlas needed 109 attempts across six weeks to reliably climb three steps; and technical fun, in Raibert’s view, is close to the whole point of being an engineer.

5. Hardware still matters, and hydraulics is unfairly written off. Raibert resists the assumption that only software and learning drive robotics progress. BigDog’s leap was integrating engine power, computing, and hydraulic actuation onto one platform; Boston Dynamics later re-engineered hydraulic valve design — largely unchanged since 1950s aviation — into components compact enough to power Atlas. Spot’s shift to electric actuation came from a single request from Larry Page for something smaller and less intimidating in a home, not from hydraulics being obsolete.

Content

From a disassembled robot arm to a life in legged locomotion

Raibert dates his conversion to robotics to 1974: as a brain-and-cognitive-sciences graduate student at MIT, he followed Berthold Horn back to his lab and saw a robot arm taken apart into a thousand pieces. Neurophysiology, his formal field, had felt too disconnected from the conceptual questions he wanted to answer — how control and thought actually work. The turn specifically toward legged robots came later, at a biomechanics conference where the only robotics talk on the programme showed a six-legged robot that always kept at least three feet on the ground, moving, in Raibert’s phrase, ‘like a table or a chair’. He recognised this as nothing like how animals actually move — they bounce and fly, predicting their own trajectory to stay balanced rather than never risking instability — and resolved to build the opposite: a machine that stayed up by continuously correcting itself in flight, not by never leaving the ground.

The pogo-stick robot and the funding pitch that built it

The first hopping robot began with a characteristically improvised funding story. While at JPL, Raibert connected with Ivan Sutherland — sometimes called the father of computer graphics — who encouraged him toward a project at Caltech. Raibert deliberately proposed three ideas, burying the pogo-stick robot in the middle between two boring-sounding options; Sutherland spotted it immediately. The two walked into DARPA’s Washington offices unannounced, found programme manager Craig Fields, and showed him a disassembled hopping-robot skeleton in a Samsonite suitcase. Fields committed $250,000 on the spot in 1980 — substantial funding at the time — with, in Raibert’s telling, no larger vision yet of where legged robots might lead; he was, in his own account, simply an academic trying to make progress and impress his colleagues.

The machine itself decomposed dynamic balance into three coupled problems, solved with the tools available at the time rather than an optimal design: energy management (estimating hop height, since a pogo stick’s energy sits entirely in gravitational potential at the top of each bounce), foot placement (calculated relative to the body’s centre of mass on landing, the same problem a pole-vaulter or high-jumper solves at takeoff), and attitude correction (applying torque between leg and body, but only while the foot is on the ground — the physics rule out correcting tilt mid-air). Raibert is candid that this first solution was a rough sketch rather than a tuned one; later researchers, including his own students, extended it to greater speed and to obstacles like stairs.

BigDog, hydraulics, and bringing robots out of the lab

Boston Dynamics itself began as a physics-simulation company, not a robot maker — Raibert describes a multi-year detour building a force-feedback surgical simulator, killed once it became clear surgeons expected to be paid to help develop it rather than pay to use it. The company found its way back to robotics through work with Sony on a running version of the Aibo robot dog, and then decisively through BigDog, funded by a DARPA biodynamics programme (Alan Rudolph) and built with Martin Buehler, recruited from McGill to get the robot ‘out of the lab and into the mud’.

BigDog’s central engineering leap was integration: earlier Leg Lab quadrupeds needed an offboard hydraulic pump, a room-sized VAX computer, and a hose tether; BigDog carried its own gasoline engine, hydraulic actuation, computing, and sensing on one platform, able to walk trails at the Marine Corps base in Quantico. Raibert argues indoor operation is harder than outdoor trail-walking, since a robot in a house cannot leave scuff marks or bump walls, whereas woods impose only obstacle-avoidance. He defends hydraulics against a perception of obsolescence: Boston Dynamics substantially reinvented hydraulic valve design, largely static since 1950s aviation, producing a five-kilogram power unit delivering five kilowatts — small and efficient enough for Atlas. BigDog’s descendant LS3 could carry roughly 1,000 pounds (against a 400-pound design target) for 20 miles on a tank of gasoline; the shift toward the far lighter, all-electric Spot followed a single request from Larry Page for something around 60 pounds and less intimidating in a domestic setting.

Natural movement, knees, and passive dynamics

Asked why Boston Dynamics’ robots move with unusual naturalness, Raibert points to a forward-looking, rather than purely reactive, control style: Spot recalculates its motion on roughly a one-to-two-second horizon, continually predicting rather than only correcting for where it already is; more dramatic manoeuvres, like a somersault, require looking further ahead to ensure momentum and rotation are set correctly at launch. He credits Tad McGeer’s 1980s passive-dynamic walker — a legged mechanism that walks down a slight incline using only gravity, springs, and dampers, with no computer at all — as proof that a machine’s own mechanics can perform part of the ‘computation’ of correct motion; even computer-controlled robots move more efficiently when the controller works with, rather than against, what the hardware naturally wants to do. Knees entered Boston Dynamics’ designs with BigDog; Atlas, Raibert notes, still does not walk quite as gracefully as a human, though its running is closer — walking, counter-intuitively, has proven the harder motion to make natural.

Athletic intelligence, cognitive intelligence, and the AI Institute

Raibert frames his post-Boston-Dynamics work at the AI Institute around a single distinction: intelligence has an athletic part (mechanical design, real-time control, energetics — the domain Boston Dynamics has spent decades mastering) and a cognitive part (planning, perception, and judgement under incomplete specification — the everyday reasoning a person uses to estimate a walk time and set a departure time accordingly). He judges the cognitive side to be robotics’ real bottleneck: most robots remain, in his word, ‘dumb’, requiring extensive human programming for each task. The Institute’s flagship project, Watch Understand Do, aims at robots that observe a human performing a task — starting with simple bicycle repairs — segment the demonstration into discrete skills, and reproduce it, without an explicit model of the objects involved, closer to how a person can navigate an unfamiliar room without first building an explicit map of it. Progress is organised around ‘stepping stones to moonshots’: never going more than about a year without a visible, if partial, result on the way to a multi-year target. On machine learning specifically, Raibert acknowledges a post-ChatGPT surge of interest — the Institute’s Zurich office, led by reinforcement-learning researcher Marco Hutter, works squarely on this — but notes that, as of the interview, the most impressive athletic performances still shown publicly lean on ‘traditional’ model-predictive control rather than end-to-end learning, with the two expected to increasingly merge. See Athletic Intelligence for the concept in full.

Team culture: fearlessness, diligence, intrepidness, and fun

Raibert names four traits behind Boston Dynamics’ engineering culture. Technical fearlessness is taking on a problem with no known solution and finding a simplified entry point to learn from. Diligence is building something robust enough to survive deliberate sabotage in testing — the videos of engineers tugging ropes on a stair-climbing robot or pushing back on a door Spot is opening exist precisely to demonstrate this, not merely for spectacle. Intrepidness is tolerating repeated failure: a video of Atlas dynamically climbing three large steps represents the 109th attempt, six weeks of iteration during which the robot failed for reasons ranging from perception errors to outright mechanical breakage. Technical fun, in Raibert’s account, is close to the reward structure of the job itself — craft satisfaction, external impact, teamwork, and pay, ‘paid four times’ over. His video-editing philosophy — no captions, no explanation, ‘do something worth showing and then show it’ — reflects the same instinct: showing failure alongside success is what makes the success legible, a lesson learned from his very first hopping-robot footage.

Competitors, the humanoid market, and the road ahead

On Tesla’s Optimus, Raibert is admiring of Elon Musk as a technologist but judges Optimus behind Atlas; he speaks well of Figure and Apptronik among the ten-or-so humanoid companies now active, and welcomes competition in the quadruped market as validating the category rather than threatening Boston Dynamics specifically. He sees warehouse logistics as currently the only clearly money-making robotics use case, with home and social robotics — a Spot-like companion — still awaiting the combination of performance, safety, and cost that has repeatedly defeated earlier consumer-robot companies. On AI risk, Raibert argues that a smarter-than-human system is not inherently threatening — by analogy to already living among humans smarter than oneself, and to cars, a technology with large real harms that is nonetheless judged, on balance, a net benefit as those harms are engineered down over time. Asked for career advice, his repeated formula is to imagine what one would do with no resource, opportunity, or skill constraints, and then work back toward what is realistic — paired with his long-standing personal motto, borrowed from a DARPA-era Marine colleague’s response to any objection: ‘why not?’

Related

• Athletic Intelligence — the athletic/cognitive intelligence split Raibert introduces here
• Marc Raibert — speaker
• Lex Fridman — host