It took Atlas fourteen years to walk the way Playter wanted

The gait Boston Dynamics wanted was ordinary on purpose. The army had hired the company for a robot that could wear a protective suit and walk through a toxic-gas test as a person would, so the motion had to look natural: the leg reaching out, heel striking first, then rolling onto the toe instead of slapping down flat. They built an early machine they called the PETMAN prototype, and Playter still likes how it moved. Even then, getting the robot to land on a fully extended leg — the way people actually walk — refused to come together.

He did not see the walking he had been waiting for until 2022, on a newer generation of Atlas. The team was developing a different process for walking-control algorithms, and the natural-looking motion showed up as a byproduct, not as a pose they had copied by hand. The PETMAN prototype was around 2008. That is ten to fifteen years of work for a robot to do something every able-bodied viewer already knows how to do.

Lex Fridman frames the rest of the conversation around that patience. Playter is CEO of Boston Dynamics, the company behind Atlas and the robot dog Spot, and he has been there from the MIT years. He took his PhD in aeronautical engineering in 1994 at the Leg Lab, wrote a thesis on robot gymnastics, and programmed a biped to do what the show calls the world’s first 3D robotic somersault. The viral clips — dancing, opening doors, throwing heavy bags, backflips — sit on top of that longer method.

Playter first fell for the work in a basement. Visiting MIT for a PhD, an aero professor told him to go see Marc Raibert under the AI lab. Raibert showed him robots, including one doing a somersault, and Playter’s reaction was immediate: robots can do that? He was already a gymnast, already in aeroastro because flight and movement fascinated him, and here was the open problem that still occupies the company — how a legged machine balances and actually gets around.

Gymnastics had given him a phrase for what he was looking at: letting your body do what it wanted to do. Once you are good enough, you put yourself in a position where inertia and momentum push you the rest of the way. That was what Raibert was building toward with springy pogo-stick machines — a bouncing spring-mass system with its own natural frequency, then intent layered on top without overpowering the physics. Playter calls the coordination beauty, or synergy. The point is the same. If the body is set up right, the physics solve part of the problem, and some of the stabilizing feedback lives in the hardware rather than in the computer.

The lasting lessons from those years, he says, are still company values. Have the courage of your convictions and go after a big problem because it is interesting, not because you can already see the money — dynamic legged locomotion, at the time, looked like a multi-decade bet that nobody had solved. The other lesson is reduction. Raibert would show videos of animals running across the savanna or climbing mountains, then start with a pogo stick, because balancing that stick already contained the fundamental difficulties. Solve those, and you can eventually extrapolate to something that gallops like a horse.

Fridman adds an observation Playter recognizes. He meets a lot of PhD students worn down by the process; roboticists work brutal hours and still look happy, in a way he otherwise associates with people who skydive a lot. Playter’s version is more specific. Attrition at Boston Dynamics is very low. People come and love the pursuit, partly because a machine moving in the world is easy to identify with. He does not think the happiness is unique to robotics — anyone doing work they love works harder — but he feels lucky the company could build a business around that pull.

In the early days, Playter says, what made the group unusual was working at that principled level: find the central idea, then build software and hardware around it, with feedback control simple enough to run in real time at 1,000 hertz on the computers of the era. Not many teams could do both. The other special habit was rougher. Use the robots, break them, repair them, test and repeat. If you are working with a very expensive machine you bought and cannot fix, you treat it with kid gloves and you cannot make progress. You have to be able to break something. That, he says, is how the work stays fearless.

Fridman remembers the opposite feeling: an RC car he had loaded with a computer and cameras, code written on too little sleep, a full-speed slam into a wall, and a private mix of sadness, regret, and new fear. Playter’s reply is practical. It depends on whether you could have built the car again. In graduate school he learned to run a lathe and a mill, broke a unique part in the lab, and realized he did not have to buy a replacement. He could make it. That was an enabling feeling. It might take more time than he wanted, but he was no longer afraid.

Asked whether the dynamics were more art or science, he refuses the split. There is real science in a spring-mass bouncing system and in taking those principles from one leg to two or four. Humanoids add an art channel. The form is so complex that you lean on intuition about how a body should look while it moves — knowledge that lives in your own joints before you can write it down. A coach, he says, can see a tiny change in an athlete and sometimes cannot quite articulate it, only ask the athlete to feel it. Humanoids make that kind of seeing available. A stranger shape would not.

Fridman points to *Ex Machina*, whose robot is played by a ballerina, as the standard for motion that looks efficient, easy, and simply right. Early robots look wrong in the other direction: not because caution is ugly, but because the movement feels unnecessarily inefficient. Playter’s account is that Boston Dynamics machines attract people because the inherent dynamics are closer to right. Some early walkers were essentially trying not to fall, always stopping the tipping motion. The insight of dynamic stability is to go with it. Let yourself fall, then catch yourself with the next foot. Get those physics expressed in the machine, and viewers read the result as lifelike, elegant, natural. If you get the physics right, he argues, the gait is also likely more efficient, stable over a wider range of conditions, and more beautiful at the same time.

That is why walking, not the viral flips, is the stubborn problem. On the army contract they needed a natural-looking gait and got a good one on the PETMAN prototype, including the heel strike. The remaining snag was the straight leg. People land more or less extended, then flex the knee on contact. Fully extended, the leg is up against a mathematical singularity: it cannot go further in that direction, and the calculations for torque and position get ugly. Early work boiled the body into simpler models that did not contain all of that complexity. More recently the team can take the full physics of the robot into account and handle those strange configurations.

Fridman asks whether the deeper issue is that the robot is underactuated — that it is always falling. Playter agrees the word is right. You cannot push the machine in any direction you want. There is no gravity hook, and no god’s hand, applying an external force on demand. The only external forces come through the feet, and they depend on where you put them. Natural human walking lives inside that constraint, which is why a gait that looks alive has to be invented there rather than imposed from above.

The humanoid form is attractive and, Playter says, a headache. You have a big upper body with a lot of mass and inertia; throwing that around makes balance harder, and picking up something heavy makes it harder still. Early quadrupeds at the Leg Lab and the company had a built-in simplification: a large rigid torso and very light legs, so swinging a leg barely disturbed the body. Atlas does not get that gift. The legs are big and heavy. Swing them and everything else moves. That coupling is what makes the platform so much more complicated.

Recent Atlas videos are explicit about this. The robot throws a heavy bag, handles a barbell, moves objects that are an appreciable portion of its body weight, because a useful machine has to move stuff around. Playter thinks that still differentiates them from other humanoid efforts, most of which are not picking up heavy things yet. The hard part is anticipation. Reach for an object and your center of mass is already shifting; pick it up and the mass shifts again; turn, and the inertia has to be in the model; throw it, and all of that has to be planned. The robot needs an expectation of the weight and has to think a couple of seconds ahead about how to manage body plus load without falling. That was a big change. With the new tools, some of the motions in the latest video were created in a matter of days. It used to take about six months to do anything new on the robot.

Once the object is in the hands, he argues, you do not want to overfit every shape. The gripper has to conform to the surface, but after that it is mostly mass and inertia that matter — especially if you are going to throw a tool bag onto scaffolding and only need the center of mass. A rigid body is easier; the loose-filled tool bag they used already worked. Harder cases still waiting include a large jointed object or a bunch of loose wire or rope. Carrying another robot is mostly a joke: they did a Christmas skit in which two Spots held up a third trying to put a bow on a tree. Asked how much Atlas can curl, Playter laughs that it probably cannot curl as much as a person yet. A metric he prefers is the box jump. Atlas managed something like a meter, a height his chief technical officer struggled to reach. The next generation, he says, is heading into a regime where some joints and motions will be stronger than a person.

Jumping and flipping evolved faster than walking. The first somersaults were almost handmade: pick a trajectory, decide how hard to throw, iterate. In those early experiments Playter would still borrow from his own backflip intuition and suggest technique changes. The team does not need that anymore. They now run model-predictive control, in which the robot thinks a second or two ahead, solves for optimal trajectories from A to B, and explores options such as moving an arm a little more one way. Optimization, and sometimes learning, has accelerated the behaviors even though the mathematics underneath is heavy.

The striking part is that this loop still runs while the whole robot is flying. Physics is just Newton’s laws. The “sick trick” — a front somersault with a half twist — is the example Playter likes: he has seen Atlas land in different configurations and still stabilize. If takeoff was a little screwy, the controller projects forward in real time and adjusts. People do a version of the same thing on a gymnastics trick they have set up to be repeatable. The catch, which Fridman had already named, is underactuation at its most extreme. In the air you cannot change momentum, because you cannot apply an external force or torque. You are not even a drone. You cannot hover. You are going to hit the ground. Be ready.

Doing a backflip with a body that large took nerve, and they broke a few robots trying. That is the build-it, break-it, fix-it strategy again: repeated failures expose weak points, then you redesign so the next machine does not break so easily. When people ask why you would bother flipping a robot, Fridman quotes Kennedy on the moon. Not because it is easy. Because it is hard. Playter laughs and agrees.

The 2015 DARPA Robotics Challenge asked humanoids to drive, get out, open a door, find and shut a valve, use a tool, climb stairs, and cross rough terrain, with communications that could drop entirely so the robot had to proceed on low-bandwidth guidance. Boston Dynamics’ role was to build a first-generation Atlas — Playter thinks they made maybe ten — and hand machines and baseline software to teams that had won a programming contest. Carnegie Mellon and others brought their own robots. Boston Dynamics did not compete in the maze. They supported everyone else. It was humbling. The tasks were not that hard for a person. Mostly the robots fell down. Some finished and won. The tears Fridman remembers from his couch, eating popcorn, were the same ones the builders felt watching years of work fail in public.

A general-purpose platform, Playter notes, is cursed: Atlas was a bit too big to get in and out of that particular car, which was probably harder than the driving. Until now, successful robots have been built for a specific task in a controlled environment. The future usefulness of robots depends on dealing with uncertainty and unexpected situations, and that is just darn hard. One failure he cannot forget is a robot that missed a door handle, assumed the grasp, and threw itself over when the expected force never came back. People take that recovery for granted, drunk or sober. Robots still have to be taught it. That is why the team leans on model-predictive control: predict a little ahead, and if the world disagrees, predict again.

The control stack is layered. Joints typically run around 1,000 hertz, measuring position or force and commanding a hydraulic or electric actuator with little extra math. A slower loop, maybe 100 hertz, looks at whole-body physics. Perception and vision sit slower still. All of it has to share computer time in a consistent real-time budget; as computers improve, the forward models can get more sophisticated. The software story matters as much as the math. They test in physics-based simulation first, and it was only several years before the interview that the same code ran in simulation and on the hardware. They have about four modern Atlas robots and roughly forty developers fighting for time on them. The simulation tools themselves started at MIT, failed as a commercial product, and survived as an internal advantage — especially contact with the ground, which off-the-shelf tools did not handle well. Dextrous manipulation, Playter warns, will be harder still, because objects slip in the hand.

Fridman wants to know whether to say what or who when the subject is BigDog. Playter will not draw a hard line around it, him, or her. Organic movement makes people identify with the machine and attribute feeling, the way they get attached to cars. That can help human-robot interaction, and the company has a little fun with it — a recent Atlas video starts with the robot looking around for a bag of tools after a worker on scaffolding asks for them. It can also be a curse. Viewers have only seen living things move this way, so they attribute a level of intelligence the robots do not yet have.

The company started in 1992. For about a decade it was not a robotics company; it sold simulation. A stint writing software for Sony’s Aibo got them slightly involved again, and then a DARPA contract around 2003 asked for a robot dog. BigDog was the first machine they built with everything on board: power plant, computers, hydraulics that needed cooling, no tethers. The transition was rough. Early versions used off-the-shelf go-kart engines, gas-powered two-strokes chosen for weight, often without mufflers because they did not care about noise. The robots stood about belly or chest high and weighed maybe a couple of hundred pounds. Over about five years they went from flat ground to rocks, inclines, mud, and slippery mud, and convinced themselves that legged locomotion in a robot could actually work. MIT quadrupeds had always been tied to a lab pump and a lab computer. BigDog walked in the world and balanced.

Once DARPA saw a working legged robot, more programs followed. They built LS3 to carry heavy loads — an 800-pound machine designed for a 400-pound payload — and Cheetah to find out how fast a robot could run. Leg length and payload became parameters they knew how to scale. They also started a humanoid thread. Around 2012, with government interest cooling, they sold the company to Google so the work could continue. Spot was born from Larry Page asking what the product would be. The logical answer was a quadruped, but smaller, without a gas engine, and probably not hydraulic. That was the start of an electrically actuated robot.

Page liked the toothbrush test: products you use every day, ideally a consumer machine. Playter did not think that was the next step. A household robot probably needed to cost a few thousand dollars; they were building machines that cost hundreds of thousands, maybe a million, two at a time. The company’s philosophy has been to build the best hardware you can, so you are discovering the hard problems of balance rather than fighting a crappy machine. Industry can pay. If Spot costs $100,000 or more, that is not a large expense next to a factory that has to keep running. Consumers will not buy that robot. A cheap home machine may come later, he thinks, but the path runs through excellent industrial ones you then learn how to simplify.

The scare in going from R&D to a commercial company is cultural: will the same people who love hard scientific problems care about quality, reliability, and cost? Playter’s takeaway is that they have good engineers, and manufacturability is just another kind of problem. Reliability, though, is a different sport from the high-stakes demo. They run fleets 24/7 in the offices to find rare failures. Visitors see a dozen Spots walking the hallways on autonomous missions — off the dock, around the building, collect data, sit back down — because that is what a brewery or factory actually needs. On a weekly basis the robots accrue something like 1,500 or 2,000 kilometers of walking and more than 1,000 hours of operation, a volume he does not think anyone else can match. They are still learning to cast parts instead of milling them from billet, to mold plastic, and to build the same robot twice. Final assembly for Spot and Stretch is in a Waltham factory about a mile from headquarters, while design changes from the reliability tests keep landing on the line.

The hard problem of the next decade, he says, is mobile manipulation. Spot’s arm is almost as complex as the robot, an attachable payload with motors, sensors, and a camera in the hand so the machine can close on a grasp the way it already manages its own feet. In software version 3.3, an autonomous mission can include going through a door: on a tablet you mark the handle, which way it hinges, whether it opens in, and the rest is up to the robot. Ontario Power Generation used Spot to throw a high-power breaker that can produce a lethal arc flash; the first teleoperated pass took about an hour, and Boston Dynamics automated the sequence — flip a cover, press a button, ratchet a socket — because it was the next logical manipulation problem, not because it would sell a pile of units. Bigger customers who want ten, twenty, or thirty robots get to set that kind of direction.

Making money as a robotics company is notoriously hard. Early Boston Dynamics lived on contract R&D. Google, SoftBank, and now Hyundai were owners willing to invest through to products. Playter’s discipline is applications broad enough to sell thousands, then tens of thousands; sell only hundreds and you commercially fail. The bet for Spot is industrial monitoring. A chip-fab line down because a vacuum pump failed can cost a million dollars a day. Thermal cameras, acoustic imagers for compressed-air leaks, gauge reading, and vibration measurements are standard preventive work at AB InBev, GlobalFoundries, electric utilities, and nuclear plants. They had about 1,100 Spots in the field. Profitability, he says, wants something like 1,000 to 1,500 a year.

Stretch started from the market, not from the company’s blood. By some estimates a trillion cardboard boxes move around the world each year, many of them by hand, and a mobile robot that can shift them looked obvious early. As soon as they talked to potential customers, people offered to order twenty. First deliveries went out in January of the year of the interview to DHL and Maersk; Gap was getting robots as they spoke, with seven or eight other customers already committed to between ten and twenty each, a couple hundred units in all. Playter thinks Stretch will go. Unloading trucks is only the start. Palletizing, depalletizing, loading — any box-moving task in a warehouse — and he can see thousands, probably tens of thousands, of machines.

It looks like a strong arm on a mobile base about the size of a pallet, because pallets are what warehouses are made of. It is not a legged robot. That was a deliberate, slightly awkward commitment: their first non-balancing machine, and a challenge because the easier problem still had to be done well. The first logistics concept was a balancer. That robot is Handle, a machine Fridman calls epic, out of a science-fiction movie. Its history, Playter says, connects the whole family. Once Atlas could stand on one foot and use its arms the way a person on a tightrope uses momentum, they sketched a two-wheeled robot with a long arm that could reach a box at the top of a truck and a tail for counterbalance. Wheels on legs, a tail, an arm, all moving together — people said it looked like an ostrich — because they had figured out the mathematics of momentum control and wanted to see it in a new body.

Handle is beautiful. It moves as gracefully as anything they have built. It was also the wrong machine for logistics: too slow, not enough boxes per hour. Letting it go was a hard-nosed business decision, a sign they would not build only for the beauty of the mathematics. Stretch has a heavy base and a giant battery good for two shifts, sixteen hours, which also keeps a 50-pound box from tipping the robot. It is omnidirectional, with suspension for gaps in the floor. Inside a shipping container in summer it can be 120 degrees, and the work is backbreaking. At DHL, Maersk, and Gap, the people who used to unload trucks are becoming the robot operators. Playter will not quote a price; he will say customers typically see a return in two years. With Stretch the target was obvious. With Spot it took a while to find.

Playter would like Spot-like machines in the home. He thinks the path is industrial applications, mass manufacturing, and a broad set of skills, with performance first. Social-robot companies that started from cute interaction mostly did not survive. Safety is the other gate: if you are going to interact with children, you had better be safe, and today they still count on some standoff so you do not pinch a finger. A home robot would also have to know beer from Coca-Cola and clear dishes without breaking them. Fridman pushes a narrower dream — a companion with no chores, a dog that is a friend. Playter does not dismiss it. Sony sold 100,000, maybe 150,000 Aibos, then suspended and later revived the line; people treated them as pets. ChatGPT-like conversation might open social robots again. It is probably not Boston Dynamics’ path, because they want to start from utility and add the rest.

Hyundai is the primary investor; SoftBank still owns 20 percent. They are not, Playter says, fixated on driving the company to profitability as soon as possible. The longer brief is what mobility means and whether they can shape it. He describes the internal goal as self-destination: without a profitable business, somebody else drives the ship, but they intend to keep building the next generation of technology at the same time. Elon Musk and Tesla’s Optimus announcement, in that light, was welcome. When Musk speaks, people listen, and a decade of Boston Dynamics work suddenly had a brighter light on it. Humanoid startups proliferated; some cite a former Boston Dynamics employee as a fundraising credential. Competitive juices are flowing. The latest Atlas video — scaffolding, a thrown toolbox, a flip — was meant to show not only parkour but the ability to pick up a 50-pound block on a robot that weighs about 150 pounds. Next they want more dextrous hands and coarse two-handed work in manufacturing and construction, reorienting unwieldy objects rather than stuffing chips into boards.

Some higher-level machine learning, Playter believes, will be commoditized. A vision model for a new door handle or gauge ought to be something you download; they already have an API, with partners such as Levatas and Reply providing gauge-reading software. Reinforcement learning is in the house for locomotion and manipulation, and he can imagine new behaviors eventually coming from outside, though that still requires a deep understanding of the robot. Large language models surprise him like everyone else. He is anxious about disinformation, an unintended consequence of social media that these tools could make worse. He is not very worried about the same models on a robot doing a manual task. You can judge whether the job got done. Physical reality is a good verifier; a theoretical answer from ChatGPT may or may not be true. The week of the interview he had already seen a Levatas demo that hooked a language tool to Spot for spoken commands.

The Boston Dynamics AI Institute is a separate organization led by Marc Raibert, founder, former CEO, and Playter’s old MIT advisor. Raibert wanted unadulterated research without the confinement of commercialization; Hyundai agreed to fund it. Boston Dynamics will keep doing nearer-term research, on a horizon Playter puts around five years. The institute is for over-the-horizon problems, the way legged locomotion once was a multi-decade bet.

Playter holds up a Wall Street Journal piece about ChatGPT illustrated with a humanoid that looks as if it will replace you. He hates the collage. People are afraid of legged robots, he argues, because they have been taught to be afraid for more than a hundred years. The word robot comes from Karel Capek’s 1921 play *Rossum’s Universal Robots*, and in that first depiction the robots take over at the end. People love to be afraid; fiction made takeover feel like the logical progression. Fridman thinks a super-intelligent system in a visible body is less threatening than software that can pretend to be human at arbitrary scale. Playter adds that humanity already knows how to certify machines. It does not yet have that experience with software that can propagate worldwide in an instant. The fear of the latter should not, he says, impede getting robots out, because you can judge when a robot is being safe.

On jobs, he reaches for a familiar history. Every large technological change brings automation anxiety. Someday people will look back at unloading trailers by hand and ask why they ever did it that way, and there are a lot of people in that job now. His version of the off-ramp is that the same workers can operate the machines, the way farming moved from hand tools to equipment without a general regret. Demographics, in his view, make the timing right: the first world is losing population, Europe already worries about staffing logistics, and after COVID those jobs are getting harder to fill. He is more alarmed, with Fridman, at language models drafting the text that consultants and journalists used to write — while noting the models still do not know how to be right.

Boston Dynamics authored a letter, co-signed by companies including Unitree in China, Agility in the United States, and ANYmal in Europe, stating they will not put weapons on their robots. The prompt was commercial machines already showing up in videos with guns strapped on, remotely operated. A bright line, Playter says, is better for an industry that needs trust in order to be let in. The plan is to engage lawmakers and regulators; he has a policy director whose job is those conversations. Asked what it takes to be CEO after starting as a PhD engineer, he puts courage first: courage to tackle hard problems, and courage to lead people smarter than he is. Adaptability is the rest of the story. He never aspired to the job; for probably twenty years he was the main person hiring, essentially head of HR, stepping into whatever was not being done. The interview he liked asked candidates to present work they cared about so you could see whether they had really done it. Google’s standardized coding questions compare people at scale and, he thinks, tell you less about who they are.

He does not spend much time on robot consciousness and thinks it is remote for the machines they build. People read intelligence and emotion into physical gestures. Large language models, as he understands them, are elaborate word completion, a statistical averaging of common responses, and he is dubious that this has much to do with consciousness or with the kind of absolute knowledge he associates with truth and emotion. You cannot average something true and something untrue and get something moderately true. His advice to young people is the Leg Lab principle that became a company value: follow your curiosity, love what you are doing, do not plan too far ahead, and be willing to pivot — the way he knew, on first seeing Raibert’s robots, that he had to go do this. In a hundred years he expects robots to be ubiquitous and doing hard labor, but not a *Wall-E* cruise ship of unused people. People need to work to feel productive. The complementary future he wants keeps the creative problem-solving, and the simple labor that is well done, on the human side. A long-running study he mentions, perhaps from Harvard, found that companionship makes for a happier life. Machines might simulate that. He is not convinced they will ever be sentient. Whether a robot can sit with you and bring up a shared memory, he says, remains to be seen.