Bones That Want to Run

or, the long way around the valley

In the 1970s, paleontology had a problem with dinosaur posture. The textbook drawings going back to the early twentieth century showed Tyrannosaurus and the other large theropods upright like enormous kangaroos, dragging heavy tails on the ground behind them as a third support. The Dinosaur Renaissance of the 1970s — Bakker, Ostrom, Greg Paul — argued that the upright pose was an artifact of bad reconstruction. Bipedal dinosaurs, the new picture went, ran with the body held horizontal, the tail extended straight out behind, the tail acting as a counterweight to the head. The fossils alone could not settle the disagreement. Bones tell you where muscles attached. They do not tell you, by themselves, which muscles fired when.

What eventually settled it was simulation. Beginning in the late 1990s, biomechanics groups built virtual skeletons of theropods — Tyrannosaurus, Velociraptor, Allosaurus, the smaller ones — layered virtual muscles over the virtual bones, and asked a computer to find a gait. The procedure was a genetic algorithm. Try a population of muscle-firing patterns. Evaluate which patterns moved the body forward fastest without falling over. Breed the survivors. Mutate. Repeat. After enough generations of in-silico optimization, the dinosaurs ran. They ran horizontal, the tails held out behind for counterweight, exactly as the renaissance had been arguing. The bodies the computer was driving were the fossil bodies. The gait fell out of the algorithm. The renaissance was right.

This is the trick. Given a body and an objective, a way to use the body can be found. It can be found by an algorithm running on a computer. It can be found by an algorithm running on a brain — by a baby on a kitchen floor, working out how to use the legs it was issued; by Dick Fosbury, in his coach’s barn in Oregon in 1968, working out how to clear a high-jump bar by running at it backwards and arching his spine over it. It can be found by an algorithm running on a population — the four billion years of evolution that produced the bones in the first place. Different substrates. Same recipe. Try variations. Keep what works a little better. Try variations on those. Repeat. People call it different things in different fields: natural selection, reinforcement learning, stochastic search, the blind watchmaker. It is the only algorithm anyone has found that produces the appearance of design without a designer.

The algorithm works on babies the way it worked on the dinosaur skeletons, except faster. A nine-month-old does not know what walking is. Some part of its brain has been handed a body to operate; the body has muscles; the muscles can be twitched; the twitching produces sensations of tipping, of contact, of being suddenly unsupported. Through some combination of trial and pain and applause, the neural circuits controlling those muscles get tuned. Twitches that produce forward motion are reinforced. Twitches that put the carpet against the face are suppressed. Within a few months the child can stand. Within a year it can walk. Within two it can run, at which point its parents lose the ability to leave any room unsecured. Nobody told the child what walking was supposed to look like. The body and the brain figured it out together, by trying things and keeping what worked.

The algorithm is not perfect. The same trial-and-error process that finds good gaits also gets stuck. The recurrent laryngeal nerve in your neck loops down past your aortic arch and back up before reaching your voice box, fifteen feet of detour in a giraffe — a route that made sense in the gilled fish where it first ran, and that no mutation can shorten without breaking everything in the middle. The vertebrate eye has its photoreceptors pointed backwards. The QWERTY keyboard layout was designed in 1873 to slow typists down. None of these are good designs. They are local maxima — peaks that are higher than every other peak nearby but lower than peaks farther off. To find a higher peak you have to climb down first, into a valley, before you can climb up again. Climbing down hurts. The crowd boos when you try the backwards jump and miss the bar. Most jumpers, most of the time, do not make it across the valley. The ones who do — like Fosbury — make the rest of us look obvious in retrospect.

The Experiment

Below is the same algorithm at work, on a body the universe never made. The body is fixed: a small theropod-style trunk laid out horizontally, two legs hanging from a hip joint near its center, four hinge joints in total. The body does not change between trials. What changes is the controller — twelve numbers describing how to twitch each joint to a sine wave, which is a serviceable cartoon of what a nervous system does when it fires its muscles. Ten of these controllers are tried at once, in ten separate physics worlds, drawn on top of each other so you can watch all ten attempts in parallel. After ten seconds, the ones that did best on whatever you chose as the objective keep their numbers. The rest are replaced by mutated mixes of the survivors. Repeat. This is reinforcement learning, not natural selection — the body is not changing — but it is the same algorithm the dinosaur simulations ran on, the same algorithm Fosbury was running in his barn, the same algorithm a baby runs on the floor.

You do not pick winners. That would be intelligent design, which is the thing this page is built to argue you do not need. The only choice you make is what counts as winning. Should the body learn to run, jump high, climb stairs, or swim? Pick the objective and step out of the way. The rest is the algorithm.

Things to try:

It takes a while to converge, because the search is harder than the biomorphs and easier than evolution. A body has to coordinate four oscillators against gravity and not knock itself out by landing on its head. But the underlying mechanism is the one that ran in the dinosaur simulations when the gait fell out, in Fosbury’s nervous system in 1968, in a baby’s motor cortex on the kitchen floor, and in the four billion years of trial and error that built every body that has ever walked. Mutation is undirected. Crossover shuffles. Selection filters. The walkers are what is left.

And the algorithm has been running on you all along. The body you are using is mostly the body you were given. The objective you are pursuing is mostly the objective you settled on years ago. The trial-and-error system in your head that learned to walk has not stopped — it is still trying variations on how you use yourself, still keeping the ones that work, still mostly invisible to you. The question is not whether the algorithm runs. It does. The question is whether you let it find you a better gait or keep insisting on the one you have.