Ai designs a robot that crawls on after being chopped in half
A machine-learning agent with no taste for bipedal theatrics has quietly sketched, iterated and 3-D-printed a modular creature that keeps moving even when its own limbs are blasted apart. Northwestern University engineers, who let the algorithm breed 3,000 generations of virtual bodies overnight, watched the winner wobble out of the printer at dawn: a stack of hexagonal pods, each carrying its own motor, battery and microcontroller, that can shrug off a shotgun blast of gravel or a deliberate bisection.
The design that no human would sketch
Evolutionary algorithms work like selective breeding for silicon. The code mutates joint angles, segment lengths and actuator torque, keeps whatever shambles fastest across a simulated sandbox, then spawns again. No aesthetic filter, no Hollywood humanoid fetish, just pure locomotive utility. The result looks like a centipede crossed with a Rubik’s cube—ugly, until you see it reassemble after falling off a cliff. “We never told it to survive mutilation,” says project lead Sam Kriegman. “It simply discovered that redundancy beats elegance in the real world.”
Each pod talks to its neighbors through a 915-MHz mesh. Lose half the stack and the remainder vote on a new gait pattern within 400 milliseconds. The current prototype, built from carbon-fiber-infused nylon, costs $142 in parts and walks away from a table saw that lops off 60 % of its body mass. Power draw drops, speed halves, but mission continuity remains. Try that with a $200 k titanium-legged biped.
From disaster zones to asteroid regolith
The team’s DARPA funders picture swarms of these polybots slithering through reactor rubble or melting ice caves on Europa. No central brain means no single point of failure; every slice is a complete robot that happens to share a spine. Add cheapness and you get expendable scouts that can be air-dropped by the hundreds, instructed to melt together into bridges when they hit the ground.
Inside the clean room I poked a dismembered segment with a soldering iron. It twitched, oriented its LEDs toward the heat source, and inched away like a wounded starfish reclaiming life. My stomach flipped—not at the gore, but at the casualness of resilience. We are used to fragility; this thing treats damage as a suggestion.
Kriegman shrugs off terminator anxiety. “The algorithm cares about joules per meter, not world domination.” Still, the lab has already filed for export-control review. A robot that refuses to die is dual-use by nature: first responders love it; military planners love it more. The same code that saves earthquake victims can also clear a minefield while taking fire.
Next month the team ships twenty units to FEMA for flood-channel tests. If they come back in one—well, in however many pieces still function—mass production starts in a converted auto-parts plant outside Chicago. Price target: $37 per segment. At that cost, disposability becomes a feature, not a bug.
Science fiction warned us about machines that replicate; nobody promised they would also be cheap, ugly and impossible to kill. The future arrived looking like broken pottery that refuses to stop crawling, and it is already cheaper than a weekend in Vegas.