Modular robots evolved by AI survive being chopped in half and keep running

Chop one of Sam Kriegman's robots in half, and you will not get a broken machine. You will get two smaller machines, each one rolling, crawling, and searching for its other half. When the pieces find each other, they reconnect and resume running.

This is not a design trick. It is an emergent property of what Kriegman's team at Northwestern University calls legged metamachines -- modular robots whose body plans were not drawn by human engineers but evolved inside a computer. The work, published March 6 in the Proceedings of the National Academy of Sciences, represents the first time AI-evolved robots have been deployed outdoors and tested against real-world terrain.

Lego blocks with brains

Each building block is a half-meter-long leg module shaped like two sticks joined by a central sphere. Inside that sphere sits everything the module needs: a circuit board, a battery, and a motor. On its own, a single module can roll, turn, and jump. But the real capability surfaces when modules snap together.

To find the best configurations, Kriegman's team used an evolutionary algorithm that mimics natural selection. They gave the algorithm the modular legs as raw material and a single goal: design a robot with efficient, versatile movement. The algorithm generated body types, simulated each one, kept the best performers, discarded the rest, and bred new designs by combining or mutating survivors. Depending on the configuration, modular legs became actual legs, spines, or tails.

The output was strange. No human engineer would have conceived these forms. Some designs undulate like seals. Others bound like lizards or spring like kangaroos. The algorithm did not care about aesthetics -- it cared about locomotion, and it found solutions that fall well outside conventional robotics thinking.

Gravel, mud, tree roots, and a deliberate amputation

The team assembled the best three-, four-, and five-legged configurations and took them outside. The metamachines ran across gravel, grass, tree roots, leaves, sand, mud, and uneven bricks. They jumped and spun in the air. When flipped upside down, they righted themselves without any retraining or manual intervention.

Then came the damage tests. When a leg broke off, the remaining modules adapted their gait and kept moving. The severed leg, still a functional robot in its own right, could roll independently and eventually rejoin the group. A metamachine is, in effect, a robot made of robots. There is no single point of failure.

"They can survive being chopped in half or cut up into many pieces," Kriegman said. "When separated, every module within the metamachine can become an individual agent."

From tabletop walkers to outdoor runners

The new work builds on earlier research from Kriegman's lab, where the team developed the first AI algorithm capable of designing robots from scratch. That previous system compressed billions of years of simulated evolution into seconds and produced a small, flexible walking robot. But those early creations could do little more than crawl across a table. They could not sense their own bodies or coordinate movements among parts.

The leap from table to terrain required solving both the hardware and software problems simultaneously. The modular legs needed to be mechanically robust enough for outdoor conditions while remaining simple enough to snap together without specialized tools. The control system needed to handle gaits for body plans that had never existed before -- plans that the algorithm invented, not the engineers.

"Evolution can reveal new designs that are different from or even beyond what humans were previously capable of imagining," Kriegman said. "So, we really wanted to study how and why it works. The best way -- or at least the most fun way -- is to evolve structures in realistic conditions."

What this does not yet do

The metamachines are a proof of concept, not a product. The modules rotate around only a single axis, which limits the complexity of tasks they can perform. The evolutionary algorithm optimized for locomotion, not for manipulation, sensing, or communication with humans. The outdoor tests demonstrated durability and adaptability on natural terrain, but the robots were not performing useful work -- they were simply moving.

Scaling the approach presents open questions. Larger metamachines with more modules would need more sophisticated coordination algorithms. Power remains a constraint: each module carries its own battery, and runtime is limited. The team has not yet demonstrated long-duration autonomous operation or the ability to self-repair broken internal components rather than simply working around missing limbs.

Still, the core insight is significant. By combining physical modularity with AI-driven body design, the researchers have produced machines that do not merely execute pre-programmed movements but adapt their behavior to their current physical configuration. That is a different paradigm from conventional robotics, where body and controller are designed as a fixed pair.

The road ahead for evolving machines

The practical implications extend beyond novelty. Search-and-rescue operations, environmental monitoring, and exploration of hazardous terrain all demand robots that can survive damage and keep functioning. A robot that can lose a limb, adapt, and continue moving is fundamentally more useful in unpredictable environments than one that becomes inert after a single mechanical failure.

The co-first authors on the study are Chen Yu, David Matthews, and Jingxian Wang, all Ph.D. students at Northwestern's Center for Robotics and Biosystems. The research was supported by Schmidt Sciences AI2050 and the National Science Foundation.