AI Didn't Build Dogs — It Built Something That Writhes

Cut a traditional robot in half and you have two broken robots. Cut one of Sam Kriegman's machines in half, and you have two robots that keep moving.

That's the image from a study published March 6 in the Proceedings of the National Academy of Sciences by researchers at Northwestern University — and unlike a lot of robot demos that end when something breaks, this one just keeps going. Chop it into quarters, and every resulting piece becomes an independent agent, still trying to complete whatever task it was given. It's a genuinely strange thing to watch, and the researchers lean into that: their AI didn't design something that looks like a dog or a human. It designed something that writhes.

"These are the first robots to set foot outdoors after evolving inside of a computer," Kriegman said. "They are rapidly assembled and then quite literally hit the ground running. They can survive being chopped in half or cut up into many pieces. When separated, every module within the metamachine can become an individual agent."

Kriegman — an assistant professor of computer science, mechanical engineering, and chemical and biological engineering at Northwestern's McCormick School of Engineering — calls the resulting machines "legged metamachines." The name is apt: each one is a robot made of other robots. The building blocks are half-meter modules that look like a pair of sticks joined by a central sphere. Inside that sphere is everything a robot needs — a circuit board, a battery, and a motor. Alone, a module can roll, turn, and jump. Connected to others, the combinations can bound like a lizard, undulate like a seal, or spring like a kangaroo.

The design process is where it gets interesting. Kriegman and his team didn't draw these body plans — they evolved them. Starting with the modular building blocks as a constraint, they ran an evolutionary algorithm that generated thousands of possible configurations, simulated each one in a virtual physics environment, kept the best performers, and iteratively "bred" new designs by mutating and combining the winners. Depending on the configuration, the modules became legs, spines, or tails.

"We simulated the Darwinian process of mutation and selection within a virtual, physical environment," Kriegman said. "This is survival of the fittest — accelerated by computers and made real by athletic modular building blocks."

The result of that simulated evolution wasn't constrained by what a human engineer would consider a reasonable body plan. "Evolution can reveal new designs that are different from or even beyond what humans were previously capable of imagining," Kriegman noted. "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."

That phrase — realistic conditions — is the actual news. AI-evolved robots have been demonstrated in labs and controlled environments before. What Northwestern's team did differently was take their evolved designs straight to the outdoors. The metamachines ran across gravel, grass, tree roots, leaves, sand, mud, and uneven bricks without retraining or complicated setup. They jumped, spun, and righted themselves when flipped. According to ZME Science, all of the physical designs learned a jump-turn maneuver requiring orientation before landing — a capability the algorithm developed in simulation that transferred directly to rough terrain.

The co-first authors are Chen Yu, David Matthews, and Jingxian Wang, all PhD students in Northwestern's Center for Robotics and Biosystems. Douglas Blackiston from Tufts University and the Wyss Institute at Harvard also contributed. The work was supported by Schmidt Sciences AI2050 and the National Science Foundation.

The resilience claim is where the video is most compelling. When a traditional robot breaks, it fails. When one of these metamachines loses a leg, the remaining modules adapt to the missing limb and keep moving. The severed leg rolls away independently — it's a complete robot on its own, capable of rejoining the team or acting as its own agent. The team has demonstrated this on video: a machine bludgeoned in multiple places, still crawling toward its objective. As Reuters covered, it's a step toward robots that can adapt to damage in the field rather than requiring immediate human intervention.

There are real limits. According to ZME Science, the machines cannot yet fully reconfigure themselves, automatically absorb new modules, or rebuild after damage without outside help. The modules have only one degree of mechanical freedom — rotation around a single axis — which keeps the design simple but constrains what gaits are possible. The step from a research demonstration to a field-deployable system is substantial.

The funding — Schmidt Sciences AI2050 and the NSF — signals the kind of basic research this is: fundamental science, not a product roadmap. Kriegman's lab has a track record here. In 2023, it demonstrated the first AI algorithm that could design a working robot from scratch in seconds. That earlier work produced machines that could walk across a table. This one goes outside.

For robotics researchers and funders watching the broader embodied AI space, the study is notable for what it suggests about the sim-to-real transition. Evolved designs have historically struggled when dropped into the physical world — the gap between simulated physics and real terrain is wide. That these machines learned to handle rough outdoor conditions in simulation and then did it is a meaningful data point, even if it's a controlled demonstration rather than a product claim.

The weirder claim — that breaking a robot doesn't break it, it makes more robots — is either a compelling feature or a philosophical reframe, depending on what you think robots are supposed to be. Kriegman's framing is unambiguous: he calls the circuit board a nervous system, the battery a metabolism, the motor a muscle. "Inside the sphere, the robot has everything it needs to survive," he told Northwestern's press team. Whether that language is metaphor or something closer to a design philosophy tells you a lot about how this lab thinks about the boundary between machine and organism.

The paper is C. Yu, D. Matthews, J. Wang, J. Gu, D. Blackiston, M. Rubenstein, and S. Kriegman, "Agile legged locomotion in reconfigurable modular robots," PNAS 123(10) e2519129123 (2026).