Northwestern Engineers Build AI-Evolved Modular Robots That Survive Being Chopped in Half and Keep Running
Legged metamachines evolved by AI traverse rugged terrain, self-right when flipped, and continue operating after losing limbs, in research published in PNAS.
Overview
A team at Northwestern University has created the first modular robots that can be assembled from interchangeable parts, evolved by artificial intelligence inside a simulation, and then immediately deployed outdoors where they run across rough terrain, recover from severe damage, and keep moving. The study, published March 6 in the Proceedings of the National Academy of Sciences, introduces what the researchers call “legged metamachines” — stick-like robots composed of autonomous modules that snap together like building blocks.
“These are the first robots to set foot outdoors after evolving inside of a computer,” said Sam Kriegman, assistant professor of computer science, mechanical engineering, and chemical and biological engineering at Northwestern’s McCormick School of Engineering. “They are rapidly assembled and then quite literally hit the ground running.”
How They Work
Each module is a half-meter-long structure resembling paired sticks connected by a central sphere. Inside the sphere sits a motor, a battery, and a circuit board — what Kriegman describes as “muscle,” “metabolism,” and a “nervous system,” respectively. Every module can rotate around a single axis and is a fully functional robot on its own, capable of rolling, turning, and jumping independently.
When multiple modules are connected, the resulting metamachine moves in ways that depend on its configuration. The team tested three-, four-, and five-legged designs, each producing distinct locomotion styles. According to the Northwestern press release, some configurations undulate like seals, others bound like lizards, and still others spring forward like kangaroos.
Notably, the robots carry no external sensors — no cameras, lidar, or microphones — yet still demonstrate adaptive behavior in unstructured environments, as RoboHorizon reported.
Instant Evolution
To find the most effective body plans, the researchers used an evolutionary algorithm that mimics natural selection. The system generates candidate configurations by mixing and matching modules, simulates each design in a virtual physics environment, retains high performers, discards poor ones, and iteratively breeds new designs through mutation and combination. According to EurekAlert, the process compresses “billions of years of evolution into mere seconds.”
The algorithm produced body shapes no human engineer would have conceived. Rather than defaulting to familiar dog-like or humanoid forms, the AI generated strange new “species” of machines optimized purely for locomotion efficiency, as described in the Northwestern announcement.
Damage Resilience
The most striking capability is the metamachines’ resilience to physical damage. Traditional robots typically fail when a single component breaks. These machines, by contrast, can survive being chopped in half or cut into multiple pieces, according to Northwestern. If flipped upside down, they instinctively right themselves and continue their journey.
When a leg is severed, the remaining modules adapt their movement pattern to compensate for the missing limb. The detached leg itself continues rolling independently and can rejoin the group, as reported by Tech Briefs. The team tested the robots across gravel, grass, tree roots, leaves, sand, mud, and uneven bricks.
What We Don’t Know
The researchers acknowledge that the robots are “not yet useful” for practical applications, positioning this as foundational research rather than immediately deployable technology, RoboHorizon noted. Key open questions remain: how well the metamachines scale to larger module counts, whether adding sensors would improve or complicate the evolved behaviors, and how the approach compares to other modular robotics platforms in terms of real-world endurance.
The study also does not report specific speed metrics or quantitative comparisons with conventional legged robots, making it difficult to assess raw performance against existing platforms.
Significance
The research was funded by Schmidt Sciences AI2050 and the National Science Foundation, according to EurekAlert. It builds on Kriegman’s earlier work designing AI-generated robots from scratch, though previous iterations could only walk across tabletops without environmental sensing or self-coordination.
The paper, titled “Agile legged locomotion in reconfigurable modular robots,” points toward a future in which robots can be rapidly assembled, repaired, and recombined for autonomous operation in unpredictable environments — from disaster response to planetary exploration — without complex setup or retraining. Co-first authors Chen Yu, David Matthews, and Jingxian Wang are Ph.D. students in Northwestern’s Center for Robotics and Biosystems.