Let the physics do the work: the idea behind graceful legged robots

The Short Version

Graceful legged robots came from a change of attitude toward physics more than from faster computers. Russ Tedrake, Robert Playter and Marc Raibert each describe the same move: stop using motors to cancel what a body naturally does, and instead add a little energy and intent to motion that gravity, springs and momentum already supply [1][2][3]. Playter puts it as letting the robot tip and then catching it with the next foot, and Tedrake points to walking machines with no controller at all [1][2]. That is why a pogo-stick hopper with a three-part control rule could balance in the early 1980s, long before today's computing [3]. The sources also show the limits of the idea: flat ground, heavy humanoid bodies and contact with the world still needed prediction, optimization and decades of work [1][2][3].

Table of contents

What They Said

Russ Tedrake: machines that walk with the motors off

Asked for the most beautiful motion he has ever seen, Tedrake, an MIT professor and vice president of robotics research at the Toyota Research Institute, picks the passive dynamic walkers, especially the 3D walking machine Steve Collins built with Andy Ruina at Cornell, building on work Tad McGeer started [1]. You set it on a small ramp and give it a push. "No controllers, no batteries whatsoever," he says, and at the time it looked more natural and human-like than any robot yet built [1].

He explains why it works with the rimless wheel: a bicycle wheel with the rim removed [1]. Each time a spoke hits the ground it loses a little energy, and each step down the slope gains a little [1]. The two effects balance into a stable periodic motion: too slow and it speeds up, too fast and it slows down [1]. Remove all but two spokes, add a hinge, and you have a compass gait walker; Collins and Ruina took the idea all the way to a robot with knees, arms and a torso [1].

His contrast is Honda's ASIMO, which he defends as a marvel of engineering but which, like MIT's own Atlas work, was not energy efficient [1]. Conventional control, in his description, takes whatever the world is doing to you, pushes back to "erase the dynamics of the world," and imposes dynamics chosen because they are mathematically simple [1]. The walkers showed another path: perhaps we should be "letting the world push us around and we go with the flow" [1]. In his words, "You can just let physics do most of the work" [1].

He is careful about what the demonstration did not show. The machine only walked down a ramp, and to walk on the flat you have to add energy at some point [1]. Everybody, even at the time, said the answer was to do it with control; the point was that "maybe the way we're doing control right now isn't the way we should" [1]. Designing such a machine is also partly art, with small tricks like tuning suction cups at the knees, and it took a lot of falls [1].

His favorite animal example makes the same case. Researchers George Lauder at Harvard and Michael Triantafyllou at MIT put a cylinder in a water tunnel to recreate the vortices behind a rock, and a live trout behind it changed its gait and appeared to surf between them [1]. Then they tied a dead fish behind the cylinder. "The dead fish basically starts swimming upstream," Tedrake says. "Completely dead. No brain, no motors, no control" [1].

From this he draws a working definition. Mechanics is what happens with the motors off: gravity, springs and dampers, which is what muscles and tendons act like [1]. Control is commands that add new energy [1]. He argues we have underestimated how important dynamics are [1].

The same thinking gives his MIT course its name. A robot with as many motors as joints is fully actuated; take one away and it is underactuated [1]. "The world gives us underactuated robots, whether we like it or not," he says, and "I'm a human. I'm an underactuated robot," since no motor connects your center of mass to the ground when you jump [1]. The passive walkers are the extreme case, where you take away all the motors [1]. He sums up that "the philosophy there is let physics do more of the work," with optimal control and numerical optimization as the main technical tools [1].

Robert Playter: the gymnast's version

Playter, chief executive of Boston Dynamics, came to the idea through his own body [2]. Visiting MIT to find a PhD lab, he was sent to see Raibert in the basement of the AI lab, where he watched a robot do a somersault [2]. As a former competitive gymnast he recognized it at once: a skilled athlete learns to let the body do what it wants to do, putting it in a position where momentum and inertia carry it naturally [2].

Raibert was building machines that did the same, starting with a springy pogo-stick hopper whose spring-mass system had its own natural frequency [2]. The trick was to add intent and control without overpowering those natural dynamics [2]. Playter says this is more efficient, and that it "becomes easier to control in its own way because the physics are solving some of the problem itself" [2]. He goes further: stabilizing mechanisms can occur simply by virtue of the physics of the body, so "It's not all in the computer" [2].

His clearest statement concerns falling. Early walking machines tried hard never to tip over, always stopping the tipping motion [2]. The insight of dynamic stability is to go with it: "let the tipping happen. Let yourself fall, but then catch yourself with that next foot" [2]. Getting that physics expressed in the machine, he says, is something "people interpret as life-like or elegant or just natural looking," and it likely ends up more efficient too [2].

He ties the principle to a method. While Raibert showed videos of animals galloping across the savanna, he started with a pogo stick, because solving balance there contained the fundamental problems that could later extend to a four-legged gallop [2]. In the early days the special skill at Boston Dynamics was finding that central principle and building a feedback controller "simple enough that it could run in real time at a thousand hertz" [2]. The science, Playter says, lies in principles like the spring-mass model that let the team extrapolate from one leg to two or four [2].

Marc Raibert: reject the slow robot

Raibert, who founded Boston Dynamics and ran the Leg Lab at Carnegie Mellon and then MIT, describes the moment of decision [3]. At a biomechanics conference on legged locomotion he saw a six-legged robot that always kept three feet on the ground and moved very slowly [3]. He thought that was wrong, because people and animals balance dynamically, predict what will happen in each step, and use the springiness of muscles and tendons to recirculate energy [3]. He decided to try the opposite [3].

The result was a planar hopper working around 1982 and a 3D version in 1983, built with engineer Ben Brown [3]. The control recipe was surprisingly simple [3]. First, regulate the bouncing by estimating hop height and adding or removing energy, since energy moves between the spring, altitude and velocity [3]. Second, decide where to place the foot relative to the center of mass, like a pole vaulter who must be in the right spot when the pole plants [3]. Third, keep the body upright by applying torque between leg and body while the foot is on the ground [3].

Raibert credits the body directly. He discusses McGeer's passive dynamic walkers, which walked downhill with no computer, as evidence that the body participates in motion and that good designs account for what the machine naturally wants to do [3]. He and Brown studied kangaroos and ostriches and filmed and digitized horses, and his job with Brown's early spring-filled sketches was to insist on simplifying [3]. The company's early motto was that you have to run before you can walk [3].

He adds two things the others say less about. Natural-looking movement comes partly from good hardware, and people who think hardware innovation is finished are wrong [3]. The other part is dynamic, predictive control instead of backward-looking servoing [3].

Where They Agree, and Where They Don't

The consensus. All three reject the strategy of overpowering a body's natural motion. Tedrake names it as erasing the dynamics of the world; Playter as stopping the tipping; Raibert as the slow robot with three feet always down [1][2][3]. All three cite the same proof of principle, a walker with no computer [1][3]. Playter's and Raibert's accounts connect directly, since Playter was Raibert's graduate student and made a two-legged robot do a 3D somersault at MIT [2][3].

A caveat on the shared origin. Playter and Raibert trace the idea to the Leg Lab [2][3]. Tedrake, in this conversation, traces it to the passive walkers of McGeer, Collins and Ruina, and the summary and passages of his conversation do not mention the Leg Lab [1]. The principle is common to all three; the lineage is shown for two.

Elegance versus power. Tedrake's ideal is a machine driven by gravity alone, and he counts MIT's Atlas work among the inefficient approaches [1]. Raibert still loves hydraulics for unmatched strength for their weight, and his robots hop and run aggressively [3]. These are different readings of the same rule: Tedrake stresses how little energy is needed, Raibert how boldly a machine can move once it balances dynamically [1][3].

How graceful Atlas is. Playter says he saw the natural walking he expected from Atlas about 2022 [2]. Raibert says Atlas's running may now be close to human, but its walking still is not quite as graceful [3].

Art or science. Tedrake calls passive walker design partly art [1]. Playter calls the work part science and part art, with the art showing up in humanoids, where everyone has an intuition about how a body should look in motion [2]. Tedrake also argues hardest for rigor, warning that deep learning "makes it easy to get things to work without thinking as critically about it" [1].

Learning versus classical control. Raibert notes that the most impressive performances, including most of Atlas's, still come from traditional control like model predictive control, and expects the best results from combining it with learning [3]. Playter says his company uses reinforcement learning for locomotion, manipulation and perception [2]. Tedrake uses deep learning heavily and still wants the rigorous foundations taught [1].

What comes next. Raibert believes manipulation needs the same shift today: after 50 years of careful geometric grasping, robots are nowhere near human dexterity [3]. Tedrake says "we mostly in robotics are afraid of contact on the rest of our body, which is crazy" [1]. Playter calls mobile manipulation the hard problem of the next ten years [2]. They agree on the direction.

Why It Works

Energy balances itself. In the rimless wheel, energy lost at each foot strike and energy gained from the slope settle into a steady rhythm that corrects itself when disturbed [1]. Stability comes from the mechanics.

Springs recirculate energy. Raibert's hopper moves energy between the spring, altitude and velocity, and the controller only tops it up or trims it [3]. Playter describes the same spring-mass system as having its own natural frequency [2].

Falling is the motion. Dynamic stability treats tipping as part of walking, with foot placement doing the catching [2][3]. Playter says this keeps a machine stable over a wider range of conditions than refusing to tip [2].

A resonant body does work. The dead trout shows mechanics alone extracting motion from a flow [1]. Tedrake adds a personal example of bodies adapting: after he lost a finger, the bone next to it remodeled over a couple of years to handle the new contact, so "the idea that somehow it's only our brain that's adapting or evolving is not right" [1].

Simple control was enough to start. Because physics solved part of the problem, the first controllers could be simple and fast [2]. The sources show how little computing was involved: the passive walkers used none, and the Leg Lab quadruped relied on a room-sized VAX computer off board [1][3]. This supports the claim that balance came before computing power, with one qualification the sources themselves make. Later gains leaned on computation: Atlas now uses model predictive control, projecting a second or two ahead, in real-time loops that Playter says keep improving as computers get faster, and Tedrake's team had to rewrite an optimization solver overnight to keep its controller running in real time [1][2].

Where It Applies, and Where It Doesn't

The principle applies most cleanly to locomotion with a rhythm: hopping, running, walking down a slope [1][3].

It needs help on flat ground and with intent, where energy must be added and choices made [1]. It gets harder as the body gets heavier and less symmetrical. Playter explains that legged robots are underactuated because the only outside forces come through the feet, and that a humanoid's heavy upper body and heavy legs make balance much harder than on a quadruped with light legs [2]. A fully straight leg is a mathematical singularity that complicates the torque calculations [2].

It does not by itself handle contact. Tedrake describes making or breaking contact as discontinuous, with the combinations of which surfaces touch exploding in number [1]. At the DARPA Robotics Challenge, MIT's Atlas fell when a foot controller left on by mistake put it in a position where, as he recalls, "nothing in our plans had ever told us what to do if you're butts on the seat and your feet are in the air" [1]. Playter recalls a robot that missed a door handle and threw itself over because it expected the door to push back [2].

It also does not decide what sells. Handle, a balancing robot Playter calls graceful like nothing else they have built, was dropped for being too slow at picking boxes, and its replacement is, in his words, "not a balancing robot" [2].

Tedrake applies the idea to his own running: barefoot, a hard landing hurts immediately, so the gait corrects itself, often subconsciously [1]. He warns that the biggest mistake is experienced runners taking their shoes off and immediately trying their usual distances [1]. Anyone changing how they run should get advice from a qualified professional first.

What To Do With It

1. Ask Tedrake's question of any moving system: what does it do with the motors off? That part is mechanics, and it is free [1].
2. Start with the simplest machine that contains the core problem, as Raibert did with a pogo stick before any gallop, and as the rimless wheel does for walking [1][2].
3. Design the controller to add energy and intent to the natural motion instead of cancelling it [1][2].
4. For balance, let the body tip and solve for where the next foot goes [2][3].
5. Break the control problem into a few plain rules; Raibert's hopper used three: bounce height, foot placement, body attitude [3].
6. Predict a second or two ahead and re-plan when reality differs, the approach both Boston Dynamics accounts describe [2][3].
7. Expect to break hardware and budget for it. Playter says "it's good to break stuff," and Raibert advises spare parts and a repair crew over a gold-plated robot on the shelf [2][3].
8. Test far more than feels necessary. Tedrake credits the winning DARPA team with having two robots and more disciplined testing [1].
9. Stop designing robots to avoid touching the world; Tedrake's verdict is "I think we're building robots wrong and being more soft and more contact is important" [1].
10. Learn the fundamentals rigorously before reaching for learning methods, which is Tedrake's advice to students [1].

Sources

1. Russ Tedrake on Underactuated Robotics, Contact, Touch and the DARPA Robotics Challenge · Lex Fridman Podcast
2. Robert Playter on Boston Dynamics: Atlas, Spot, Stretch and the Business of Legged Robots · Lex Fridman Podcast
3. Marc Raibert: Boston Dynamics and the Future of Robotics · Lex Fridman Podcast