Quick answer: Great, the robots have discovered how to repair themselves. Next standup they’ll be asking for a raise and a second monitor. Researchers at Columbia built modular robots ("Truss Links") that rebuild themselves by grabbing spare parts from their environment. Here’s the unsettling detail: a tetrahedron-shaped robot grabbed a spare link, used it like a walking stick, and improved its downhill speed by about 66%. It didn’t just fix a failure, it used the failure to get better. So the honest answer to "repair or upgrade?" is: the line between them turns out to be surprisingly thin.
One view: this is just maintenance, evolved
Self-healing materials have been a robotics goal for years; a machine that swaps its own broken part is a resilience win – fewer truck rolls, huge for factories and space missions.
The other view: it’s adaptation, not repair
A system that opportunistically acquires components to outperform its previous self is closer to growing than fixing. That’s a different idea, and a slightly bigger one to sit with.
Why a software shop cares
We spend our days trying to get software agents to do the tame version of this: notice they’re broken, recover, and keep going without a human babysitting them. Physical self-repair just makes the question louder: how much autonomy do you actually want a system to have before "it fixed itself overnight" stops being good news? That tension is exactly what agentic AI engineering – guardrails, sandboxing, agent stabilization – exists to manage.
FAQ
Adapted from a LinkedIn post by Andrew Babkin, Managing Director & CTO at Azati – read the original, on Columbia University’s "Robot Metabolism" research (Science Advances).
