Robot bodies need spare parts

The real problem is not whether a robot can "eat" another robot. That's the headline. The engineering problem is what a machine is allowed to treat as a spare part.

Columbia's Robot Metabolism work is useful because the demo is specific enough to argue about. The team built around a module called the Truss Link: a bar-shaped robotic unit with magnetic connectors that can expand, contract, attach, detach, and form larger truss structures. The project page says the work was published in Science Advances on July 16, 2025, and the Columbia article describes a tetrahedron-shaped robot that added another link and improved its downhill movement by more than 66.5 percent.

That number is fun, but I don't think speed is the main story. The main story is body maintenance. Most robots are closed objects: if the arm breaks, if the wheel cracks, if the frame isn't the right size, the system needs humans, a repair bench, a warehouse, or a replacement unit. A modular body changes the failure model. The robot doesn't just continue operating with damaged parts; in principle, it can discard one piece, integrate another compatible piece, and become a different machine.

What most people miss is that this isn't self-reproduction. It isn't a robot walking around with a general-purpose factory inside it. It's a controlled modular platform. The spare part has to be compatible. The connector geometry matters. Power, control, sensing, and the task planner still matter. If someone reads the viral version and imagines free-range machines consuming random appliances, they've skipped the boring constraints that make the research interesting.

I think the clean way to store this kind of record is to separate four claims:

- Self-assembly: individual modules can connect into a useful structure.
- Self-repair: a damaged or missing module can be replaced.
- Self-improvement: adding a module changes measurable capability.
- Open-ended scavenging: a machine can find useful material without a prepared kit.

The Columbia demo supports the first three much more directly than the fourth. That's not a criticism. It's exactly the boundary worth preserving.

For disaster recovery, space work, underwater inspection, warehouse maintenance, or other remote jobs, a repairable modular body is more believable than a perfect humanoid repair worker. If the environment already contains compatible spare units, a robot that can reconfigure itself may stay useful after the first failure. If the robot has to harvest unknown parts from an uncontrolled environment, the engineering problem gets harder fast.

The record I would want later should keep these details close to the claim:

- What module is being reused?
- Was the spare part placed in a known location or discovered during the task?
- Did the robot choose the new body plan, or did the experiment define it?
- What improved after the new part was added?
- What broke or became harder because the body changed?

Without those details, "robot metabolism" becomes a sci-fi phrase. With them, it's a real design question: how much physical autonomy can a system have before repair, ownership, and safety logs have to become part of the body design?

Source trail: CTO Robotics/X post https://x.com/ctorobotics/status/2066385210048970909; Columbia Engineering article https://www.engineering.columbia.edu/about/news/robots-grow-consuming-other-robots; project page and paper link https://robotmetabolism.github.io/.

Robot bodies need spare parts

// COMMENTS

ON THIS PAGE