You Have to Run Before You Can Walk: Marc Raibert on Making Metal Look Alive

Lex Fridman introduces Marc Raibert as the founder and longtime CEO of Boston Dynamics, now executive director of the Boston Dynamics AI Institute, with more than forty years of legged machines behind him at CMU, the MIT Leg Lab, and the company that made BigDog, Atlas, Spot, and Handle. The origin story is quieter than the robots. Raibert was always a builder. His father had wanted to be an aerospace engineer; a grandmother from the old country thought that sounded like a grease monkey, so he became an accountant and filled the basement with tools. The son watched him assemble electronics kits and still has a couple of the old ones.

The conversion happened in graduate school, in January 1974, during MIT's Independent Activities Period. Raibert followed Berthold Horn back from class to the AI lab. On the table was a robot arm taken apart into maybe a thousand pieces. From that day on, he says, he was a roboticist. He had been an electrical engineer at Northeastern and a neurophysiology student in Brain and Cognitive Sciences, and single neurons were not conceptual enough. He could not see how they would get him to control systems or thought. The arm, still in pieces, spurred his imagination toward what it might become.

That tension between studying brains and building bodies has flickered on and off at MIT ever since — David Marr visiting the AI lab, later the Center for Brains, Minds and Machines. As a young professor Raibert went through a platonic phase in which function was everything: mobility, dexterity, perception, intelligence. The ideal robot, sitting still, should look like a pile of junk, which a lot of his machines did. Then it would start to move and you would feel life in it. They designed without worrying about the aesthetics of the structure. Later he admitted that the look of the thing combines with the lifelike motion, but the heart of the work is still making them do something interesting.

Fridman notes that Raibert's robots are not cautious. A lot of machines creep through the world as if afraid of it. His, especially early on, were aggressive and underactuated — hopping, wild, fast. The philosophy starts at a biological locomotion conference in Philadelphia while he was still a graduate student. Almost every talk was muscle, neurons, respiration. One robotics talk showed a six-legged robot that always kept at least three feet on the ground, tripod-stable like a table or a chair, and moved very slowly. Raibert looked at it and thought, that is wrong. People and animals bounce and fly. They have to predict the next running step. They use springiness in muscle and tendon; energy circulates instead of being thrown away each time.

He did not yet know how to make a hopping robot balance in 3D. The first idea was just bouncing: a spring in the leg, an actuator, an energy regime. Balance came later. Then came the one-legged pogo-stick machines. He thinks manipulation now needs the same break. The community has worked on robot hands for fifty years and is nowhere near human skill, he says, because the work is too safe: geometry of the part, careful positioning, a static grasp. People and animals juggle, hug multiple objects, sort them in the hand. Being more aggressive will mean things fail for a while. It is a longer-term bet, and he is trying to push the new institute that way. His old graduate-school friend Matt Mason, later director of CMU's Robotics Institute, once analyzed a Julia Child cooking reel and counted something like forty ways she handled food. None of them was grasping. She nudged, rolled, flattened with a knife.

The Leg Lab started at Carnegie Mellon, not MIT. Raibert was a professor there from about 1986; the first hopping machines were already working around 1982, a 3D version in 1983, a quadruped sketched in 1984 or 1985 that only really got going around 1986. Before CMU he spent three years at JPL. There he met Ivan Sutherland, often called the father of computer graphics, who encouraged a Caltech project and had an endowed chair. Raibert made a list of three possible projects, deliberately boring on the top and bottom, pogo-stick robot in the middle. Sutherland spotted the one worth doing and gave him about $3,000. Raibert built a first model with his own hands. It did not work. It was a beginning.

Sutherland and Raibert took that skeleton to Washington in a Samsonite suitcase. In those days you could walk the DARPA hallway and look in offices. They found Craig Fields, then a program manager, later DARPA's director. Raibert says you could almost see the drool of excitement. Fields sent $250,000 — serious research money in 1980 — which arrived after Raibert landed at CMU. Did he see practical machines coming? No. The motivation was the fundamentals of how animals work. As an academic he was trying to do the next thing, make progress, impress colleagues if he could, and have fun.

On the technical side he hired Ben Brown, and together they sorted out a machine that could actually work. Once the hardware had enough control of its degrees of freedom, balance was not that hard at the performance they needed. They started on an inclined air table with about six feet of travel, then a planarizer that could run around the room, then the scary step: 3D, bouncing, while inverted-pendulum carts of the era only drove back and forth on a line. The simple control story has three parts. First, keep it bouncing: energy lives in the spring, the height, or the velocity, and at the top of the hop it is all in height, so you measure how high you went and add or bleed energy. Second, choose where to put the foot relative to the center of mass — the pole-vaulter problem of being too far forward or back when the pole plants. Third, keep the body upright by applying torque between leg and body only while the foot is on the ground; the physics do not work in the air.

Fridman asks how far it can tilt before recovery is impossible. Raibert says they did not optimize. They got a sketch of a solution and pushed it. Later students and other labs wrote better algorithms for speed and for stairs, where you cannot land on the edge of a step. Finding those other people, he says, was probably the smartest thing he ever did. Looking at Boston Dynamics now, at the engineers who really make stuff work, he calls himself only the dreamer. Colleagues did not so much doubt that hopping was possible as think it was a waste of time. Some later told him it inspired them to attack harder problems. Legged locomotion turned out to be useful after all.

He was not an enthusiast for humanoids at first. Form mattered less than function, and a lot of humanoid work felt like cosmetics. After they started, the connection with people — lay and technical — was special even when the machines were not that much like us. Spot gets the same treatment. Media and YouTube commenters call the robots terrifying; when he takes Spot on the road, crowds want selfies, want to operate it, pet it, put clothes on it.

Boston Dynamics was not founded in 1992 as a robot company. Raibert left MIT — with a few overlapping years — to sell a physics-based simulation tool and do robotics-related consulting. Working with Sony pulled them back. They programmed a running Aibo, built tools for QRIO, Sony's small dancing humanoid that he does not think ever reached market, and, looking back, got themselves back where they belonged. The Aibo work was just legs they had made, swapped onto the dog, shown to Sony over years of weekly calls. There was no Zoom. They installed six ISDN lines and conferenced at something like 10 Hz, straining to understand English across the Japan boundary. The lesson through the viewfinder was blunt: something that small does not look exciting when it runs. It scampers. You need slow motion. From the hopping-machine days he had always asked how a thing would look on video, and running Aibo did not look cool.

One early "robot" was not a robot. It was a surgical simulator with force feedback, tools on SensAble devices, 3D graphics on a Silicon Graphics machine that doubled as the office heater. Surgeons looked down into a mirror at a virtual body and practiced anastomosis, stitching vessels, while a scoring system dinged them for tearing tissue or touching the wrong place. At a trade show they posted scores like a video game and competitive surgeons lined up, then came back when someone else's number was higher. The business did not exist. Raibert's team thought surgeons would pay to train; the surgeons thought the company should pay them to teach. Selling to hospitals was a marketing campaign a bootstrapped twelve- or fifteen-person company could not run. They had no investors until Google bought them after twenty years. One day he and Rob Playter looked at that simulator and a knee-arthroscopy trainer, said this is not going to work, and killed it. Once they did, it felt right. It was the milestone where they understood who they were.

BigDog put them on the map. Alan Rudolph at DARPA opened a biodynamics program, 42 proposals came in, three were funded, and BigDog was one of them, alongside the climbing robot RiSE. They hired Martin Buehler from McGill, who was crucial for getting the machine out of the lab and into the mud — build it, break it, fix it, one of the company's mottos. The MIT Leg Lab quadruped could already do many of BigDog's tricks, but the hydraulic pump sat in the room on hoses, a VAX computer occupied the next room with its own air conditioning, and the robot itself was mostly actuators and gyros. BigDog integrated a gasoline engine, the conversion of rotation into hydraulic power, and all the computing onto one platform you could take into the woods.

They spent a lot of time at the Marine Corps base in Quantico on a trail DARPA had named as a milestone, the Guadalcanal Trail. Hiking with a robot were fun days. The woods were not actually the hardest environment. In a forest you can take any path uphill as long as you avoid obstacles; you do not scuff the trees. In a house you cannot leave marks or bang walls, and early robots were not comfortable colliding. On Guadalcanal a skilled operator still did all the visual perception. The robot balanced and placed its own feet, but if the operator was wrong, it would not go. Years later they went back with a precursor to Spot and an amateur could drive it up and down on the first try.

BigDog became LS3, the load carrier. It was designed for 400 pounds; they had it carrying about 1,000, including one LS3 carrying the other. Gasoline, twenty miles. Then an intermediate machine: battery, motor, pump, still hydraulic actuation. Larry Page asked, almost as simply as that, whether they could make something that weighed 60 pounds and would not be so intimidating in a house with people. That was the inspiration for Spot as it exists. They prototyped the same size, the first all-electric, non-hydraulic robot. Raibert still loves hydraulics. The world treats them as old-fashioned and icky, and they do drip, but the force you can get in a light package is hard to beat. Most hydraulic valves were designed in the 1950s for airplanes and then made safe enough to fly in, with little change to the basic idea. Boston Dynamics' engineers put new circuits in Atlas valves, moved some computing on board, and shrank mass. They built a football-sized, five-kilogram hydraulic power supply that makes five kilowatts — motor, pump, filters, heat exchanger, valves in one package. Hydraulics still have room. Spot went electric because a house cannot tolerate the drip.

People who think you no longer need to innovate hardware, Raibert says, are wrong. The other half of natural movement is a dynamic approach: predict how the motion will evolve and use that prediction when you send commands, instead of servoing, which is backward-looking. For Spot the horizon is only a couple of seconds — a second or a second and a half — recalculated a tenth of a second later, coordinating obstacles, foot placement, and balance. A somersault looks further out. At launch you have to coordinate momentum and rotation so a landing is within reach. The first somersault was planar, on a boom, around 1985 or 1986; the calculation is not that complicated once you have jump height for time and rotation rate. Robert Playter, then a graduate student and a champion gymnast, made a two-legged robot do a 3D somersault at MIT, tucking the legs to get enough spin. Athletes, Raibert notes, do not really know in physics terms how they do it. They are coached. Building the robot is partly how you find out.

Atlas still does not walk like a person. Running is closer. That inversion is the point of the old motto: you have to run before you can walk. Walking is a strange folding and unfolding of the knee, with compliance and springiness a human can get just right. Along the way they also built Wildcat, a quadruped that hit 19 miles an hour on flat terrain, probably the fastest they ever made and, with a racing go-kart engine, the loudest — complaints from three buildings away. Early Leg Lab machines often had no knees. Ben Brown's first sketch was a kangaroo-or-ostrich thicket of springs and joints; Raibert's job was to simplify. They filmed horses, and Raibert dissected an ostrich: most of the muscle is up in the feathers, the "backward knee" halfway down is an ankle, the thing on the ground is toes. The morphology is the same across these animals. Knees arrived with BigDog, a compromise. Negative work at a contracting joint is expensive if you cannot store it; BigDog had pogo springs below and no storage in the knees. Spot later took those springs out as the company moved toward more energy-driven control.

Passive dynamics is the other half of elegance. Tad McGeer in the 1980s built a walker that went down an incline with no computer — legs folding, unfolding, swinging, with dampers and springs, a mechanical computer. The body is a participant; the brain should not have to dictate every motion. Most robots today still let the computer govern everything. If you take into account what the machine wants to do, Raibert thinks you get something more efficient and more graceful. Asked the most beautiful animal motion, he goes to slow-motion cheetahs, the back, the tail whipping in turns, pronghorns pronking to confuse a chase — and then says humans, especially gymnasts, are pretty good too.

At the Boston Dynamics AI Institute Raibert likes to split intelligence in two. Athletic intelligence is mechanical design, real-time control, energetics — the standard he thinks Boston Dynamics set. Cognitive intelligence is the everyday kind: the meeting is at 9:30, Maps says twenty minutes, leave at nine. Robots are short of that. Most are pretty dumb, which is why skilled people spend a long time programming them. If robots are going to satisfy the dreams, they need to be smarter. The institute exists to combine the physicality with the cognitive side. One flagship is Watch, Understand, Do: a robot watches a human do a task, understands what it is seeing, and does the task — on-the-job training, currently science fiction on a longer timeframe. He demonstrates athletic intelligence at the table by rotating objects he has never exactly touched, fetching things from a pocket without looking. We do not have much of a clue, he says, how that works.

Open problems on the athletic side are less "can it parkour" than commercial ones. Will it increase productivity? He thinks they are close and not quite there; the missing cognitive layer is probably the biggest barrier even for physically successful robots. Reliability is another. Spot teams have worked the tail of the curve; about 1,500 units were out in practical daily use at the time of the conversation, and getting there takes a huge effort. Cost is still high, and he does not think it has to be. Hyundai, now the primary owner of Boston Dynamics, brings car-making skill that can push price and reliability. The institute's method, written into his original proposal, is stepping stones to moonshots. Go more than a year without a tangible status report — even a simplification of a multi-year goal — and you lose feedback and motivation. That was also how Boston Dynamics worked: show progress as you go, at least to yourselves.

Watch, Understand, Do has been broken into pieces. Segmenting a raw observation into acts is itself a creative, intelligent step: pick this up, put it here, stack that on it is not obvious in the pixels. They translate video or live observation into a description, then map it onto skills, starting with simple bicycle repair — a tube into another tube, a latch, which should be in range. He thinks you should not need an explicit bicycle model, the way you do not need a lecture about a room in order to walk through it. Another project is Inspect, Diagnose, Fix: the Maytag fantasy. Spot already collects thermal data, gauge readings, and sound; humans still interpret health, diagnose, and repair. Robots should close that chain. Since ChatGPT, optimism about machine learning has been huge. Physical robots are different from language; pixel values are not words. The institute opened a Zurich office under Marco Hutter, a leader in robot reinforcement learning, still half-time at ETH. Even so, Raibert is blunt: the most impressive performances in the world are still mostly "traditional" controls, especially model predictive control. Atlas videos you have seen are mostly MPC. They have started doing some learning that is really incredible, not all shown yet. He expects a mating of the two.

A great team, Raibert says, needs four things: technical fearlessness, diligence, intrepidness, and technical fun. Fearlessness is willingness to take a problem you do not know how to solve, find an entry point or a simplification, learn, and come back — Watch, Understand, Do is a mountain. Diligence is refusing a narrow solution that breaks at the first variation. Boston Dynamics videos of an engineer shoving a door Spot is opening, or tugging a rope to drag a stair-climbing robot backward, are that diligence made visible. They show the scope of the solution, not a demo on a clean staircase. He learned the same lesson on the first 3D hopper: a clip of it hopping is a so-what; a couple of falls, fast and easy, make you appreciate standing.

He was the final edit on most Boston Dynamics videos until three or four years before the interview. The theory is no explanation. If they cannot see it, it is not the right thing. Do something worth showing, then show it — no titles that slow you down, no distractions. It is hard to get people to buy in. Critics of early BigDog clips noted the human driver; at the time the company was telling a mobility story, not an autonomy story, and they have tried to stay honest as the audience's questions changed. Intrepidness may be the most important ingredient. Robotics will not work right away. He shows long strings of outtakes. Atlas climbing three big dynamic steps took 109 tries, every one filmed. Once it did it, the success rate was high; they are not cheating by showing only the best take. They show evolved performance. In those six weeks the robot collapsed on "go," missed a jump because perception failed, sprayed oil from a broken part. Humans were in the loop on data-driven learning. The accomplishment of Atlas is a machine robust enough to take that testing. It still breaks. Plan spare parts and repair people. Do not be the gold-plated lab that keeps the robot on a shelf. You cannot make progress that way.

Technical fun, in his accounting, pays an engineer four times: the craft of making something with your own hands or design; impact when other people admire it; the satisfaction of a team whose skills you do not personally have; and a salary. Engineering is science and art at once. Scientists study what is out there; engineers make what did not exist. He thinks that is the higher calling, even if the public ranks science first. The satisfying moment can be as small as a linkage that comes out half the weight and just as strong. The beautiful thing in robotics is motion evocative of life — or a high-performance act bigger and faster than other robots, in a few narrow dimensions starting to approach people. They have had a well-known ballerina programming machines and are building choreography tools in the dance world's language. Almost all of the dancing so far is on a fixed time base: press go and it plays out, not listening, not watching. He wants robots that dance with people. Brown University's Choreo Robotics class, Sydney Skybetter and Stephanie Tellex, already mixes computer scientists and dancers on two Spots. Performing next to a human with 200 degrees of freedom is still daunting. One reverse idea: robots doing popping and locking with oscillations faster than a person can, showing the humans up. It is on the plan. They have not quite gotten there.

Great engineers want an environment where interesting engineering is already happening — a chicken-and-egg that took Boston Dynamics years. Early on they paid a little less and only got people who loved it, including makers without professional degrees who had built bicycles and kayaks, people who repaired cars and motorcycles in garages as kids. In the JPL and Caltech days, if Raibert came home smelling like the oily machine shop, his wife would say he had had a good day. YouTube put the company on the map. He remembers a sales candidate from a firm with real products that nobody had heard of: everybody knew Boston Dynamics, and they did not have any products at all. They caught the YouTube wave, and it had a huge impact on the company and on what young people imagine robots can be.

Fridman asks about Optimus. Raibert admires Elon Musk as a technologist: Tesla turned a niche under one percent into a thing every car company tries to copy; SpaceX has more or less replaced NASA, with only a little exaggeration. He would not count him out for anything. He does not think Optimus today is where Atlas is. He has visited Figure and Apptronik and thinks both have good teams. Musk has resources and ambition. Reading between the lines, Raibert would like to take a little credit: it is hard not to think that seeing Atlas was a bit of an inspiration. He would love to host a robot meetup now that he is not officially inside Boston Dynamics — he is on the board — Atlas and Optimus hanging out. The institute bought everything it could on day one: a good-size Spot fleet, a couple of ANYmals from Hutter's company, Franka and Universal Robots arms, while they start building their own. It is not quite a playground where they all frolic together. Fridman is invited to look.

He is not a commercial guy and for years at Boston Dynamics barely thought about competition. There were no products to compete with. They fought for DARPA money, and in programs that might have been contests they became the robot provider: LittleDog, the DARPA Robotics Challenge, they supplied Atlas and did not compete. Cognitive work feels like competition for the first time — compute, talent, hiring. On the physical side, ten or twelve humanoid companies will compete; quadruped rivals are actually good for Boston Dynamics, because the customer's question becomes which quadruped, not whether a quadruped can do the job. That is how you know a category exists. Cost has a long way to fall. He is optimistic that the asymptote is nothing like today's prices. Money-making robotics is still mostly warehouses plus old-fashioned fixed arms in factories. Home social robots have already seen clusters of failures — one year, three companies went under — because performance, safety, and cost have to land together. Cameras in the home will probably go the way of smartphones, which already have two or three or four cameras and a microphone that almost nobody treats as suspicious, if the product and the boundaries are right.

Will robots reach human-level or superhuman intelligence? Raibert says intelligence is many things; computers are already smarter than people at some of them and not even close at others. The AGI conversation puzzles him. He asks Fridman whether anyone in the world is smarter than he is, and whether that feels like a threat. Even if computers were smarter, he does not see why that is automatically dangerous, especially if they remain available and under our control, which is how computers generally are. Fridman lays out the misalignment story: 10x, 100x, different morals, harm on the way to a goal, too late to see. Raibert compares it to the line in *Oppenheimer* about whether the first reaction ignites all matter on Earth. Possible, he guesses, but he personally does not think it is worth worrying about that way. Every technology is opportunity and risk. Cars pollute and kill about 1.25 million people a year worldwide, and he still thinks they are a boon, getting safer and cleaner. Every technology has a story like that.

The Hawaiian shirts began as contrarianism. Someone told him they were old-fashioned and he could not wear them. He stopped for about a week, then decided he was not going to let them tell him what to do, and has worn them almost every day for fifteen or twenty years. It took him a while to realize he was a contrarian; the trait can be a useful tool. During the DARPA years a Marine named Ed Tovar, still around, answered "you can't do that" with "why not?" Raibert asks it all the time. Advice to young people is the unconstrained version of the same habit. When MIT undergraduates asked what they should do, he told them to imagine no resource, opportunity, or skill constraints, start there, and see how close they can get. People think they are in a narrow channel. It is usually wider than they think. Special work can take years. Some people think he is in a rut after forty-plus years on locomotion; part of the institute's inspiration was to do something else. Cognitive function is harder to show on YouTube than a backflip. Fridman argues that sucking at a task is compelling — clumsy, confused, slowly less terrible — and that robots are a mirror for what makes humans special. Raibert agrees there is charm in it. Asked the meaning of the whole thing, he does not claim a cosmic answer. You have to have fun while you are here. It would be a waste not to.