AI compressed billions of years of evolution into seconds to create 'Lego-like robots' that can recover even when they lose limbs

AI Evolved Lego-Style Robots That Thrive Outdoors — Even When Severed

For most robots, the leap from simulation to the real world is a fraught transition. Not so for a new breed of modular machines designed by artificial intelligence at Northwestern University.

Researchers in Sam Kriegman's Center for Robotics and Biosystems have created robots made from snap-together spherical modules — each containing a circuit board, battery, and motor — that are evolved entirely inside a computer simulation and then assembled and deployed without any retraining. The work, published March 6, 2026 in PNAS (DOI: e2519129123), marks the first time robots designed via evolutionary AI set foot outdoors and immediately performed.

"We compressed billions of years of evolution into seconds," Kriegman said. The approach works by having an AI algorithm propose millions of robot body plans, test them in simulation, and select the most capable. The winning designs are then built from off-the-shelf components and deployed directly into the real world — gravel paths, grassy fields, sand, mud, and uneven brick surfaces — and keep functioning.

The evolved metamachine comprises five snap-together modules, each mechanically simple with a single degree of freedom. Yet together they form an athletic system capable of navigating rough terrain. When damaged, the design philosophy shifts from repair to resilience: if a module is severed, each detached piece continues operating as an independent rolling agent and can rejoin the larger system when reassembled. Kriegman's 2023 work showed AI could design a small walking robot, but it could barely cross a table. This new system goes from tabletop to terrain.

Funding came from Schmidt Sciences' AI2050 program and the National Science Foundation (awards G-22-64506, FRR-2331581, FRR-2440412).

The system has real limits: modules cannot auto-reconfigure, reassembly is manual, and sensing is largely internal — tracking body position rather than building a rich picture of the environment. These are lab prototypes, not deployable products.

"These machines cannot yet fully reconfigure themselves, automatically absorb new modules, or rebuild after damage without outside help," ZME Science noted. Their sensing is also largely internal, focused on body position rather than a rich perception of the outside world.

The robots are born to run and refuse to die. But getting them to do something useful beyond surviving damage remains an open question.