Let Physics Walk: Russ Tedrake on Underactuated Robots, Contact, and Barefoot Commutes

Lex Fridman opens by asking Russ Tedrake for the most beautiful motion he has ever seen in an animal or a robot. Tedrake, a roboticist at MIT and vice president of robotics research at Toyota Research Institute, does not reach for Atlas or Honda’s Asimo. He names the passive dynamic walkers, especially the three-dimensional machine Steve Collins built with Andy Ruina at Cornell. You set it on a small ramp, give it a push, and gravity is the only power source. No controllers. No batteries. At the time, around 2001, it looked more natural and humanlike than any powered robot they had seen.

The simplest model is a rimless bicycle wheel: a hub and spokes, no rim. Push it and it still rolls downhill. Each spoke-strike dumps a little energy; each forward step puts a little back. Those two can balance into a stable periodic gait — too slow and it speeds up, too fast and it slows down. Strip the extra spokes, put a hinge in the middle, and you have a compass-gait biped. Ted McGeer started the exercise; Collins and Ruina took it to knees, arms, and a torso. The arms swing. The push produces something that looks like a stroll through the park.

Designing it sat on the boundary of art and science. The system was not perfectly modeled or controlled. Knee suction cups had to stick and then release at the right instant, a finesse you tune by hand. Honda’s Asimo was then the best walking robot in the world, a marvel of engineering that still walked with bent knees, as if on hot coals. Tedrake thinks Asimo gets a bad rap, and Atlas walked with bent knees too. What those machines were not was energy efficient. Control of that complexity often uses motors to cancel gravity, erase the world’s dynamics, and impose simpler, analyzable ones.

The passive walker was a proof that you do not have to do that. Let physics do most of the work and add a little energy when you need the flat. Tedrake’s line is almost a control slogan: instead of replacing the forces the world imparts, “let the world push us around and we go with the flow.” Everybody at the time said the eventual answer still involved control. The point was that the way they were doing control might not be the way they should.

On the animal side, Tedrake’s favorite example comes from George Lauder’s water-tunnel work at Harvard and MIT. Rainbow trout swimming upstream in mating season hang behind rocks, which looks like loafing and might be an energy trick. The researchers put a cylinder in a tunnel to throw a Kármán vortex street, then watched a live fish from above. Off the rock, the fish has a characteristic gait, as recognizable as a person’s walk. Behind the rock the gait changes: it resonates, surfing between the vortices.

A live fish still leaves the brain in the story. The clincher, and one of Tedrake’s favorite experiments, used a dead fish tied by the mouth to the rock so it would not wash into the grates. You would expect it to flop in the wake. Instead it starts swimming upstream — no brain, no motors, no control. The mechanics of the body resonate with the vortex street. Efficiency is not the only goal in nature, he notes; first you try not to get eaten, and only then, other things equal, you save energy. Animals can also adapt mechanics to recover efficiency in unusual situations.

He does not like a sharp line between control and mechanics. A robot with the motors off still has passive dynamics; gravity still works. Springs and dampers are mechanics — muscles and tendons are springs and dampers. What does active work is the part that puts energy in. The controller’s job is sending commands to the motors. The divide is whether you chose to command or left the motors off. Bodies underestimate how much the mechanics matter. In his early twenties Tedrake lost a finger, his fifth metacarpal, a bone you use opening jars and bumping into things. The fourth metacarpal was not used to that contact and hurt. Over a couple of years the bone remodeled. Geometry changed. Adaptation is not only in the brain.

Fridman offers a popular story: humans walk on two feet so they can carry food back to share. Tedrake has read the books on the evolution of walking and of the human body, and he is wary. It is easy to make bad evolutionary arguments, and most of them probably are. What often decided which pre-Homo sapiens group survived was not who ate well in plenty, but who happened to be in the right place when the ice came, or who foraged a food that still worked when the glaciers arrived. A million variables are in play, and the information in the stories is thin.

History, he agrees, often turns on events that look meaningless until you look back — someone hit in the head with a rock because somebody slept with the wrong person, then tribes go to war. The same skepticism applies to diet. A community of people, including Fridman at the time, make evolutionary arguments for eating only meat; Tedrake’s mother thinks that is deeply unhealthy. He eats what he eats because he feels good doing it. Nassim Taleb’s books on randomness and *The Black Swan* make the point that a few random events probably did the deciding.

He spent years on bipedal walking because it was hard and because it challenged control theory in ways he thought mattered — not because he would tell you to start a biped company. The most compelling practical argument is that the world is built for the human form, so a humanoid is a decent way to work in the world we have today, including natural places other form factors struggle with. In the long run, he suspects, we will build environments for our robots, and that argument falls away.

Tedrake still runs barefoot, to his wife’s chagrin. He did not start as a runner and then convert. He went to the Dynamic Walking Conference, where biomechanics people and walking-robot people share a room, heard talks on barefoot mechanics, and only later read *Born to Run*. He learned long distance after he already knew the theory, so he did not have to unlearn a shod gait. He will not evangelize; some people run beautifully in shoes. For him it is a long-term game. You can run ten miles in shoes, feel fine, and discover at night that your knees hurt. Barefoot, a hard landing hurts immediately. You change gait, often subconsciously. Put a shoe back on after many barefoot miles and the gait gets worse.

His goal is to still be running in old age, not to win races, so short-term scrapes are a cheap trade for knees that last. Fridman had the opposite path: he read the book, realized he was overstriding and hammering the heel, and spent a long time practicing tiny steps to songs at about 180 beats per minute. Barefoot is the thorough version; cadence alone fixes a lot. Tedrake’s advice for beginners is one word: slowly. The people who get hurt are excellent runners who take the shoes off and try the old distance and speed. He lucked out because he could not run far at first. He carries aqua socks or running sandals and puts them on if it is ugly. The foot has as many muscles and sensors as the hand, and we stick it in a hard shoe. Barefoot is more proprioception. Cuts heal in a week; a blown knee ends running. Feet also get tough. He can run over almost anything now. He points people to Barefoot Ken Bob’s book as the best description he has seen.

He lives about twelve miles from MIT and has made a game of the commute: run in and bike home, the reverse, rollerblades, and lately a rowboat on the Charles, which is a long way around. People ask how he justifies the time. For him it is when he thinks. He works in the morning, then needs the ride in to let it settle before meetings, and the ride home to put the day down. He is not a fast runner; his fastest splits were to daycare, when they charged by the minute. He plans the day around the commute and rarely feels the two at odds, because the thinking time makes the work better. David Goggins’s cult of chosen suffering is the opposite of what he wants. You should finish feeling better than you started. COVID took the commute; he was happy walking around town with his wife and kids. The point is getting away from the keyboard long enough for things to compress.

Asked for the most beautiful idea in the field, Tedrake does not name an algorithm. He has been lucky to sit with serious control theorists, and the clarity they bring to messy problems had a lasting effect. A couple of weeks before the taping he spent a day on a Zoom robotics conference full of good ideas, then had a late meeting with a favorite theorist. They went from abstract what-ifs to extremely precise statements about systems not much simpler than the ones he actually cares about. People underestimate the power of that kind of thought.

Deep learning is amazing, he uses it heavily, and it has changed the world. It also makes it easy to get things working without thinking as critically. As an educator, his challenge is making sure students still taste the more rigorous path. You could train a network to invert a matrix; you should not. Rigorous thinking is knowing the scope and limits of the tools. Take a class in analysis and force yourself to think carefully even about the rational numbers. The exercise travels. Simplifying a system into a “toy” is not a bug. If Newton and Galileo had deep learning, ran the experiments, and handed the world a neural net, Tedrake does not think we would be as far as we are. The simplest explanation of a phenomenon is worth something. He wants another Newton, even if networks can fit Schrödinger-like equations, because harder tasks still need simpler models.

Simple does not mean low-dimensional. Classical model reduction hunts for the minimal state dimension that hits an error bound; he thinks state dimension is the wrong complexity metric. The example he lives with is a robot hand picking up an object. The equations look wildly complex not because of the free-flying hand but because contacts turn on and off. A high-dimensional but simple description of the hand in air is fine. Write a different dynamical system for every polygon on the hand and every polygon on the object, in and out of contact, and the combinatorics explode. He wants a more intuitive-physics summary of that, and he is optimistic machine learning will help get there.

The DARPA Robotics Challenge followed the Grand Challenge and Urban Challenge that put a spotlight on self-driving. Gill Pratt, then at DARPA, pitched disaster response after Fukushima: if robots had been ready to send in, the accident might have been contained, and there were later moments people wished they could send machines. Boston Dynamics would supply Atlas, a 400-pound humanoid. Hardware teams brought their own robots; software teams first had to win the Virtual Robotics Challenge for the right to use Atlas. Tedrake did not fully know what he was signing up for. The call mixed rough terrain, valves, a drill, a hole in a wall. The contest could have been about walking or about perception and planning. Locomotion over complex terrain ended up mattering a lot. Autonomy was always the argument: a degraded link, a few bits, and a robot that can take “go there, do this” instead of a human picking every footstep.

He had just gotten tenure and a sabbatical most people use to rest. He used the back half to push algorithms they liked, and to do real software engineering. The team lived in the lab. One exhausted memory: Atlas hanging limp on a harness while humans laid cinder blocks so the robot could take its little stroll. The virtual contest was dramatic on its own. Gazebo ran in the cloud. Organizers wanted real-time physics because a human was in the loop, and right up to days before, the simulator was not quite there. That suited Tedrake: his controller was a fat optimization, also about 0.7× real time. Then, three days out, an email: good news, the simulator is now 1.0×.

If the quadratic program did not solve, the robot fell, and perception would not save it. That night he found Frank Pimenta still in lab and they wrote a new solver together. They did not shrink the problem. They warm-started: at high rate the last solution is close to the next. A thin, fast layer checked whether a nearby solution still stabilized the robot, and only then fell back to the commercial solver. You could not really test the fallback, because fallback meant the robot would literally fall. They ran the contest on that ledge. Later, days before the physical event, the robot fell in a way it never had. They had one machine running almost all the time and still not enough hours. The winning team had two robots, top-rate engineering, and a testing cadence nobody else matched.

Testing, for him, now runs from unit tests on a function to integration tests of the whole stack. System-level specs for a humanoid are hard to write. On the real robot you cannot run the same experiment twice. You hope the logs caught the failure; you will probably never reproduce it. The working philosophy is Monte Carlo: run as many trials as you can, bias the environment toward the things you fear, and accept that you are doing a somewhat random search. The failures that get you are often too nuanced to put back in simulation.

Not every fall was software. Once Atlas sat at a door waiting for a green light until the batteries died and it tipped over backward. The hardest task, in Tedrake’s view, was getting out of the Polaris. Driving was relatively easy. The organizers gave them a 400-pound humanoid and a small vehicle the robot did not fit. Feet would not go under the steering column, so it straddled the center, one foot in the passenger seat and one in the driver’s, and steered with the left hand. Then it had to park and stand up from a crouch in a cramped cabin with no door. After those videos, Fridman said, getting in and out of a car — especially injured — never looked the same. It is a real control problem.

They ran something close to a NASA pre-launch checklist. On day one they had to turn off the left-foot controller that pressed the gas, then turn on the balancer. First-day nerves: someone missed the box. The robot was in a precarious pose, sneaking a long leg out, when the driving controller told the left foot to move. On open ground they had videos of hitting Atlas with a ten-foot pole and watching it recover. In the car there was no place to step, so secondary balance was off. A foot hitting the seat might have been survivable. What killed them was the jostle: the tailbone, only a little off the seat, hit, and the other foot came up. Nothing in the plan covered butt on the seat and feet in the air.

Off script, even state estimation — the system that fuses sensors into a belief about where the robot is — did not know the pose, so its predictions were wrong. Atlas shook violently and fell face-first, into the destination zone. Strictly speaking they scored the egress point. The crowd laughed. Could Atlas learn to handle that pose? That is still a big challenge. Boston Dynamics and the world’s legged teams are excellent when contact is at the feet, often point feet. An elbow on a wall is a different problem. Mature stacks keep the controller from doing something stupid. Humans use incidental contact: you get pushed, you put a hand out, you make a brand-new contact. Quadrupeds do this with feet. The field is mostly afraid of contact anywhere else. An entire branch of planning is collision-free motion planning: elaborate algorithms so the robot never touches the world. The fear is that contact is still a hard control and sensing problem.

Fridman wrestles and trains jiu-jitsu, and he asks when a robot will beat a human in a match. Tedrake grants the hardness: there is an opponent, a game-theoretic, two-way dynamical system. Strip the opponent and look only at reasoning about contact, and there is reason to be optimistic, but the best understanding is still limited. He has been moving toward manipulation, where contact is richer. Deep learning policies can look decent through contact, and there are models that predict what objects will do. Deciding quickly from those models is still unsolved. Humans probably do not solve an optimal contact sequence on the fly either. You learn basic skills slowly, then refine. Planning the optimal sequence on a humanoid model in real time may be a problem that is never solved.

What fascinates Fridman in fighting is the looseness after expertise: throwing large weights almost without effort, much of it the body learning. Tedrake hears hierarchy and composition, which current systems capture poorly. Reflexes at the bottom, vacation-planning at the top, and something magical in between. Josh Tenenbaum’s pick-three game is the picture: clutter on a table, pick a phone, a shoe, a tissue box; lift the first with the right hand, the second with the left, and use those as tools to pick up the third. Tedrake has never picked up a phone with a shoe and a water bottle. The first time is slow; most of the time you figure it out. Flexibility — well-oiled controllers in familiar territory, models when you need to plan — is the impressive part. Expertise also refines the internal model: muscles and balance you did not even know you had.

RoboCup’s 2050 goal is a robot soccer team beating the human world champions. Fridman asks which comes first, that or a UFC champion. Tedrake says the harder technology is the wrong question; who started the organization matters. RoboCup has decades of community and increasingly humanoid robots. If you want robot MMA, start now. And if robots really play soccer under the written rules, they should play unlike humans — exploiting loopholes, actuators, and fast feedback. The DRC already showed how hard it is to write rules nobody can game. Fighting is a strange case anyway: Atlas kicks you in the shins and the match is over.

Fridman brings up the *Black Mirror* episode of armed robot dogs. Tedrake never watched it, but he knows when it aired. On an ordinary Monday he showed MIT faculty early Boston Dynamics Spot footage and watched horror cross the room. He had shown clips like that for years. The episode had changed how people saw the videos. He is not a psychologist. Roboticists do have a duty to think about social effects. He used to joke that you only had to run five minutes and the battery died; now it lasts more than an hour. Fear is partly cultural: Americans his age grew up on *The Terminator*, Japanese kids on *Astro Boy*. He will not dismiss people who see Spot on patrol as dystopia — the future is uncertain — but technically we are not there, software and drones can already do damage, and he is a technological optimist: we will create, we will screw some things up, and we will invent our way out. Part of the fear is anthropomorphism, granting self-awareness the machine does not have. On existential risk he likes Rodney Brooks’s line: it will not be robots versus people, because we will be robot people — phones, hearing aids, pacemakers, prosthetics, and eventually interfaces. The smartphone is already the robot to worry about, and the problems are human ones: misinformation, attention, how we treat each other online.

What is a robot? The standard definition — computation plus the ability to do mechanical work — is close. Robotics has a curse: once something really works, we stop calling it a robot. A home dishwasher is a beautiful mechanism with chips doing complicated things, and nobody thinks of it as one. The field always has to solve the next problem and rarely gets to celebrate. Factory arms are a huge success and still do not jump to mind. His half-joke: a robot is automation that fails frequently. Control needs differ wildly: a countertop machine that never touches you, a car, a drone, a walker, a hand. His passion is dynamic interaction with the world. Contact is what separates those problems from classical triumphs — unstable fighter jets, power grids, spacecraft landing — which at least enjoy mostly continuous change. A small aileron command makes a small path change, so you linearize and bring linear algebra.

Making or breaking contact is discontinuous. Order and mode of contact change the outcome in complex ways. Simulators often assume rigid bodies, and that produces paradoxes. A three-legged stool that touches at points has a unique set of normal forces from force and moment balance. A four-legged table, legs machined to the same height and sitting still, does not: conservation laws allow many combinations of support forces. That would be fine until Coulomb friction, where maximum friction scales with normal force. If you do not know the four normals, you do not know which way the table slides when you push — left, right, or straight. Soft contact can be numerically harder, so many engines stay rigid or mix the two. That is why we do not have fighter-jet-quality controllers for manipulation: at the instant of contact the equations do not tell you what happens next.

The standard repair is a differential inclusion, a set-valued equation: in this configuration, with these forces, a set of things could happen. Non-smoothness arrives when you make or break contact, or when static friction becomes slip and velocity jumps. Robust control is the textbook answer — a controller that still does the job no matter which way the world wants the table to go — at the cost of a heavier burden. Worse is combinatorics. Each contact mode, sticking or sliding, which body parts touch, is a different system. Tedrake does not have time to enumerate a humanoid’s contact configurations, and a human never visits all of theirs either. A flat foot already has heel and toe, four modes; add bumps and timing between left and right toes and it gets finer. Enumerate every order a dexterous hand touches a water bottle and you are dead. Walking got away with the approach because feet touch in a few places. Dexterous hands have not.

People say robotics is limited because contact simulation is inaccurate. Tedrake thinks that is partly true for the wrong reason. We know the governing equations, foibles and all; we should be able to simulate them accurately. Writing a truly good simulator takes a professional engineering team, which is different from a thriving open-source community. That is why Drake exists. He does not like calling it only a simulator. It is three things: an optimization layer in C++ and Python over commercial solvers for LPs, QPs, SOCPs, mixed-integer programs; a systems modeling language in the spirit of LabVIEW or Simulink, more explicit about state and parameters than ROS, so you can run in one process, do control design, rewind and replay; and a physics engine of multibody equations plus rendering, with an almost obsessive test suite of tumbling satellites and spinning tops, plus work on how to simulate contact. ROS made it easy to snap parts together. A Mr. Potato Head stack is hard to think about clearly. Ask more of each part and you can verify, and learn against, the whole.

Toyota Research Institute has put serious money into Drake. Michael Sherman leads a team of dynamics people, and a manipulation team uses loading a dishwasher to study hard corner cases. If they only cared about picking a plate out of a sink, Tedrake says, they could write a controller with no simulation and call it done. What they want is the path to a product: any dish, anybody’s kitchen, enough confidence to deploy. Deep learning perception, complex dynamics, a controller, and a planner all sit in the loop. The question is how to put that complexity through engineering discipline and verification. The DRC taught him you do not throw the last gap over the fence for someone else to harden. It may change how systems get written. If the robot breaks a dish in testing, how do you record it? You cannot rerun the real experiment. Do you bring that one-off back into simulation, change the controller, and make sure it never happens, or add it to a distribution and trust averages?

Finding corner cases has to get better. Control theory has talked about active experiment design for forty years; today people call it curiosity, exploration versus exploitation. The older idea is more specific: characterize uncertainty, then design the experiment that yields the most information — the trajectory that best teaches a parameter. Scaling that to a deep net plus a planner is hard. They have worked on falsification and rare-event simulation, hammering the simulator if it is trustworthy, spending time where the controller is bad. Imagine always testing a car in downtown New Delhi. Internally they joke about a Black Swan generator: not just rare events, but highly impactful ones. Humans are not great at asking what is the worst that can happen either, except there are billions of us writing blog posts about stupid things we did. TRI’s fleet-learning bet is that one humanoid does not have time to visit all its states, but a fleet of similar machines might, and their computation should be set up differently from ours. Fridman notes people like watching a system improve — Tesla owners tolerate a dumb assistant that is learning — and dropping a plate does not kill anyone, which is why home robots are a more forgiving teacher than cars.

Roombas already cruise, Alexa and Google Home already sit under the counter; arms that do something useful are a matter of time. Toyota’s entry point is aging in place: help older people stay home, stay active and social, improve quality of life. It may not be the most lucrative first market, and it is a delicate one, but it may be a problem that has to be solved anyway. Older people are often less comfortable with technology, which Tedrake finds interesting: the robot learns the human while the human learns the new thing. Fridman’s parents have a blank slate — they have not seen *Black Mirror* — so they ask what it can do for them. Drones left academia and became an industry; autonomous cars are in that wave; manipulation will hit logistics first and the home shortly after. He thinks it will pop.

Tedrake dislikes robots that are afraid of whole-body contact. Look closely: the places they actually touch are soft — pads on fingertips, rubber on soles — and then the arm is aluminum. Hitting a table with a nearly rigid arm brings the rigid-contact mess back. Softness changes the mechanics. Point contacts become patches with torsional friction and distributed load. Two points on an egg, with enough force to hold it, can crack the shell; wrap the whole hand around it and the load spreads. Softness does not make the computational problem worse. Mathematics has been biased toward rigid bodies for a long time. Soft contact is smoother: small actions do not produce discontinuous disasters, and things are less likely to fly off. Soft hardware also wants sensors. A TRI team puts cameras behind the skin of a soft hand and gets rich tactile data. That matters in manipulation: you work hard on head-mounted vision, identify the object, reach, and at the critical instant your hand occludes every camera. Without touch you are blind when it counts.

His MIT graduate course is Underactuated Robotics; lecture one is basically explaining the title. Fully actuated means as many actuators as degrees of freedom; take one motor away and you are underactuated. A friend teases him: more funding, would he work on fully actuated robots? No. The world hands us underactuated machines. He is one. He has more muscles than joints in places, and still cannot directly actuate the location of his center of mass. Jump, and no motor connects that mass to the ground. Passive walkers are the extreme of taking all the motors away. Every walking robot sits in the middle: some joints push and pull the unactuated degrees of freedom. Even with both feet down, unless you have suction cups you cannot accelerate the center of mass downward faster than gravity. Take a step and you are in the subtle regime. Dynamic robots have to embrace it. Manipulation is underactuated even when the arm is fully actuated, because you do not have an actuator on the cup. Button a shirt and the degrees of freedom of the shirt make a control theorist queasy. Those are the problems that excite him, because they break foundations he has thought about for years. The technical approach that has borne fruit is to write decisions as optimization, usually numerical optimal control, and let physics do more of the work.

Perception belongs in that loop. They had a complicated stack in the DRC, and cameras on a project with Sertac Karaman flying fixed-wing vehicles at high speed through forests when lidar was too heavy. Deep learning changed what perception can do. The naive use — network looks at cameras and outputs object pose — is often the wrong spec. Is the first step of buttoning a shirt to estimate the full state, including the wrinkles in the back? Some dimensions matter to the task; many are unobservable and irrelevant. That reopens state representation. In the lab they chop carrots instead of onions, which stink and photograph badly. Move a rigid object and you want pose. One chop and you have a hundred pieces. Does the controller need pose and shape of each? Probably not, and more chops only add pieces. The Lagrangian state exists; the relevant state is a latent, compressed representation — not necessarily low-dimensional, but easier for a person or a controller to handle. If contact mechanics among onion pieces explode combinatorially, the robot sits and thinks until the soup is cold.

Fridman argues you can fall in love with a robot through language; his father insists on touch and smell. Tedrake stays practical first: if a machine does real mechanical work near people who need help, it has to stop being afraid of contact. He loves Baymax in *Big Hero 6*, and TRI has people trying to build something in that spirit. A teddy bear in the house gets taught; a Terminator gets judged. Ted Adelson, a vision scientist at MIT, expected kids to pull him deeper into vision. What stunned him was touch: his daughter licked and mouthed everything, and he decided he had to study tactile sensing. Passive watching is not the same as doing an experiment. YouTube is a different problem from a robot that can apply force. Tedrake is not a human-robot-interaction researcher, but he believes in the value of a human element — Baxter’s eyes on an industrial arm. He had a connection to Atlas. In the finals they did heart surgery on it, melting a core an hour before competing, the robot hanging with its guts out. A hundred years from now those will look like early, ridiculous, important days.

Tedrake did not read much as a child and reads voraciously now. He recommends Yuval Noah Harari’s *Sapiens* and *Homo Deus* as mind-openers, and Taleb’s *The Black Swan*, with some reservations about the author. A more surprising pick is Mortimer Adler’s *How to Read a Book*. A lot of people hate it; he loved it, because he read it while drowning in low-quality information. arXiv dumps lightly reviewed papers every day. Very little of it is worth all of your time. When you find the piece that is, dig in, mark it up, keep a wrecked paper copy. Socrates and Plato said writing could never capture the beauty of dialogue. Adler’s reply is that a really good book *is* a dialogue with the author across time and space. Tedrake finds that romantic. Isaac Asimov, famous for fiction, spent much of his career on nonfiction and was called the great explainer. Teaching at MIT changed how Tedrake thinks. Fear of the lecture and live questions force ideas to be rock solid. Remote teaching is a new medium: he can push 3D into the browser while you have a search engine open. Robotics might benefit. Fridman adds that editing five or ten minutes of mush out of an hour can make a lecture shine.

Advice for a twenty-year-old: it is an interesting time to be a kid, and a lot of signs point to winner-take-all. The people who will excel are the ones who can think deeply about problems. Plenty of people will learn to query the internet nimbly. Leaders will be the ones who think more critically than that, and that is harder to learn. One path is mathematics and engineering. Start math early. Pull things apart and put them back together. Expect a long journey. You will never know everything. Stick with it, enjoy the ride, and stop to ask why things work. The world is so full of content it is easy to lose yourself. Fridman’s book is *Animal Farm*, which for him, born in the Soviet Union, carries a country’s history in a children’s story about pigs. Tedrake has technical books and algorithm papers he returns to, and he admits the habit is slipping: arXiv, blogs, GitHub, the new thing, and you lose the ability to master an idea.

He trained in computer engineering at Michigan and liked programming. At MIT, working with computational neuroscientist Sebastian Seung, the culture demanded more mathematically and in critical thought. He borrowed a book from his advisor’s office, sat by the Charles River, and thought: I am getting my butt kicked. Everyone who does this work gets that day. He never considered quitting. He was working hard and loving it. Almost every day someone around him shows him something he does not understand, and if he could just learn that piece, another part of the puzzle would click. Being willing to know what you can and cannot do, and loving the work of learning the rest, is the best part. The whole conversation keeps returning to the same claim. The gravity walker, the dead fish in the vortex, the egg in a full hand: do not rush to overwrite the world with control. See what the world already provides. Running twelve miles barefoot is the same logic. The body will give you feedback if you do not block the channel.