AI evolved legged robots reconfigure run and survive damage

by Clarence Oxford
Los Angeles CA (SPX) Mar 29, 2026
Northwestern University engineers have created modular legged robots with what they describe as athletic intelligence, capable of operating outdoors, adapting to damage and reshaping themselves into new configurations as needed. Called legged metamachines, the systems consist of autonomous Lego-like modules that snap together in many combinations, with each module functioning as a complete robot containing its own motor, battery and computer.

Individually, a half meter long module can roll, turn and jump, but the most versatile and resilient behavior emerges when several are connected into a larger machine. Depending on how the modules are arranged, the combined robot undulates like a seal, bounds like a lizard or springs similar to a kangaroo, and it can flip itself upright when overturned, hop over obstacles and perform mid air spins. Because the overall machine is literally a robot made of robots, broken sections do not become dead weight but remain active, continue moving on their own and can later reattach.

To discover effective body designs, the team used artificial intelligence to evolve robot configurations rather than hand designing them around familiar animal like templates. The researchers started from an evolutionary algorithm that imitates natural selection, seeding it with digital models of the modular legs, which resemble a pair of rods joined by a central spherical hub that houses the control electronics, power and actuator. The algorithm was tasked with finding body plans that provided efficient and versatile locomotion and then simulated thousands of candidates, keeping the best performers, discarding weak designs and generating new variants through virtual mutation and recombination.

Over many iterations, this Darwinian style search, accelerated by computation, produced new robotic body shapes that human engineers had not anticipated. In the evolved designs, modular legs can act as legs, spines or tails depending on their position in the larger structure, and the resulting machines coordinate rotation about a single axis at each joint into complex whole body motions. According to the researchers, this process compresses an evolutionary timescale that would take millions or billions of years in nature into minutes or hours of simulation time.

The team then built physical three , four and five legged metamachines that embodied the highest performing evolved designs. In outdoor field trials, these robots ran over rough terrain including gravel, grass, tree roots, leaves, sand, mud and uneven brick surfaces, without requiring elaborate calibration or retraining for each new environment. The machines demonstrated the ability to jump, spin, right themselves when flipped upside down and continue traveling across unstructured ground.

A key result is that the metamachines retain functionality even when damaged. If a leg breaks off, the remaining structure automatically adjusts its gait and continues moving, while the detached module is still able to roll independently and potentially rejoin the group later. The researchers report that every module can sense its surroundings, move, compute and learn, enabling the overall system to be rapidly assembled, repaired, redesigned and recombined while still operating in the field.

The work extends earlier research from the same lab in which an AI algorithm was used to design small soft walking robots from scratch, demonstrating that evolution inspired computation can instantly produce viable machines. Those earlier robots could perform only basic walking on a tabletop and could not sense or coordinate their own bodies, but they showed that AI driven evolution can generate functioning designs markedly different from conventional engineering solutions. The new legged metamachines apply similar principles at larger scale and in more realistic conditions, integrating sensing, control and modular actuation into robust outdoor capable platforms.

By combining physical modularity with AI based design, the researchers argue that robots can be made less like rigid pre programmed tools and more like adaptive systems that behave analogously to evolving life forms. They suggest that such metamachines could be rapidly configured for new tasks or environments and offer inherent resilience for applications where components are likely to fail but the mission must continue. The team frames the project as a step toward machines that do not just survive exposure to the real world but adjust and reorganize themselves in response to it.

Research Report:Agile legged locomotion in reconfigurable modular robots