AI-evolved adaptable robot is almost impossible to destroy

Most robots are fragile in exactly the ways their creators don't anticipate. A warehouse bot built for smooth concrete becomes an expensive doorstop the moment it encounters an uneven threshold. A robotic dog that climbs stairs beautifully will limp uselessly if it loses a leg. The assumption baked into every design is that the operating environment will cooperate — and then it doesn't.

Northwestern University researchers have spent years working on a different idea: machines that don't need the world to cooperate. Their latest result, published in the Proceedings of the National Academy of Sciences, is a modular robot made from Lego-like units that can reconfigure themselves, recover from catastrophic damage, and keep moving across rough outdoor terrain. According to Northwestern's press release, the robot can survive being cut in half or chopped into pieces and continue functioning.

"These are the first robots to set foot outdoors after evolving inside of a computer," said Sam Kriegman, an assistant professor of computer science, mechanical engineering, and chemical and biological engineering at Northwestern's McCormick School of Engineering, who led the study. The claim is specific and verifiable — and it's the real story.

The project starts with a question that sounds almost naive: what if we stopped trying to design robots for specific tasks and environments, and instead gave AI the building blocks and let natural selection do the rest? Kriegman's team built half-meter-long modules, each containing a motor, battery, and computer inside a central sphere. Alone, a module can roll, turn, and jump. But the interesting behavior emerges when they combine — the assemblies can bound like lizards, undulate like seals, or move in ways that have no clean analogue in the natural world.

According to the research described in Northwestern's press release, the team ran an evolutionary algorithm inside a computer simulation. The algorithm generated thousands of possible configurations, tested each against virtual obstacles and extreme terrain, kept the best performers, and iteratively bred new designs by combining or mutating them. What emerged were body plans no human engineer would have chosen — gaits and structures optimized for survival in conditions the researchers themselves hadn't predicted.

The team then physically assembled the best designs and took them outside. According to Northwestern, the metamachines — three-, four-, and five-legged configurations — traversed gravel, grass, tree roots, leaves, sand, mud, and uneven bricks. They jumped, spun, and righted themselves when flipped. When a leg broke off, the remaining modules adapted and kept going. The severed leg could roll independently and rejoin the assembly.

The "indestructible" framing is vivid shorthand, but it's slightly overcooked. The robot cannot see. It has no outward-facing sensors, cannot map its surroundings, and does not know where it is going. Its intelligence is largely internal — detecting orientation and the positions of other modules. It is slow and moves awkwardly. Even Kriegman describes the current version as not yet particularly useful for real-world tasks. The research goal is proving the concept: that AI-evolved, physically modular machines can survive and adapt outside a lab. That's genuinely novel.

The comparison to science fiction writes itself. The module design — a sphere with two rotating arms — resembles the swarm robots in Disney's "Big Hero 6," tiny machines that combine into larger structures capable of reshaping themselves at will. The real version is cruder. But the direction of travel, if the metaphor holds, is roughly right.

The broader implication for robotics is worth sitting with. Most industrial and service robots remain essentially fragile instruments: highly capable within their designed parameters, useless outside them. The assumption that operating environments must be controlled — that floors must be smooth, that damage is failure — has shaped decades of robot design. Kriegman's approach suggests another possibility: robots designed to absorb unpredictability rather than be defeated by it.

Whether that philosophy scales to precision-demanding domains like warehouse logistics, surgical assistance, or infrastructure inspection remains an open question. The modules are mechanically simple, which helps with durability but limits fine control. And the evolutionary algorithm was optimized for movement efficiency in researcher-designed virtual environments. The real world is messier.

But the instinct behind the work is sound. The history of robotics is full of machines that worked beautifully in the lab and collapsed in the field. Taking evolution as a design methodology — letting selection pressure, rather than human intuition, determine what a robot should look like — is one answer to that gap. These robots aren't truly indestructible. But they come closer than most to something that doesn't need the world to be perfect.

The paper — "Evolved legged metamachines that run on autonomous modular legs" — appears in PNAS (DOI: 10.1073/pnas.2519129123). Co-first authors are PhD students Chen Yu, David Matthews, and Jingxian Wang, all members of Northwestern's Center for Robotics and Biosystems. The research was supported by Schmidt Sciences AI2050 and the National Science Foundation.

This article draws on reporting by Etiido Uko at New Atlas, which first covered the research, verified against Northwestern University's official press release and the underlying PNAS paper.