The Columbia lab trying to build robots that eat, heal and reproduce
Source · Technology desk
— Summary
An FT Magazine feature visits the Creative Machines Lab at Columbia University, where professor Hod Lipson and his students are building "robot metabolism" — modular robots that can assemble themselves from simple parts, repair themselves, and, they hope, eventually reproduce. The lab's current building blocks are white plastic rods called "truss links" (developed by PhD graduate Philippe Wyder) and a new generation of triangles (master's student Sylvester Zhang). A paper in Science Advances last summer described what the group called the first robot system to grow from single parts into a full 3D robot while systematically improving itself — without external machinery. Lipson's life quest: identify "the 20 building blocks to make all possible robots", a deliberate parallel to the roughly 20 amino acids that encode biological life.
The context is that robotics has become AI's lagging twin. Large language models swallowed the hype, the capital and the PhDs; robots, Goldfeder says, "doesn't work yet", which is precisely what makes it interesting to this group. Lipson's thesis is that the next leap cannot come from bigger models: "the brains have moved forward, and now it's time for the bodies to catch up — in nature, there is never a mind without a body." The lab's exemplar project is "a machine that manufactures another machine, and that machine just walks right out". The commercial horizon Lipson sketches is radical: consumers would one day buy not a finished robot but a bagful of modular blocks that self-assemble into whatever task they are asked to perform.
The reality is less tidy. At Penn, Mark Yim — a pioneer of modular robotics — concedes that the 2000 promises of "versatility, affordability and reliability" remain unmet: "they can't do anything, they're super expensive and they break all the time." At MIT, CSAIL director Daniela Rus calls the reproducing-robot demonstrations "very simple as compared to the promise". Even so, the Columbia lab shipped a small victory at a weekly meeting witnessed by the FT: eight of Zhang's triangles wriggled millimetre by millimetre across a linoleum corridor, snapped into a snake, then split into two tetrahedra that folded themselves up — a primitive act of robotic reproduction. Source: Financial Times, 18 April 2026, Oliver Roeder.
The story in one line. Hod Lipson’s Creative Machines Lab at Columbia is assembling the conceptual scaffolding for a future in which robots are not monolithic appliances but self-replicating living-like systems — a long bet on “robot metabolism” at the exact moment AI capital has moved elsewhere.
Key numbers
Lab: Creative Machines Lab, Mudd Hall, Columbia University — director Hod Lipson.
Lipson’s design target:~20 building blocks for all possible robots — chosen to echo the ~20 amino acids in biology.
Key publication:Science Advances, summer 2025 — first demonstration of a robot system that grows from single parts into a full 3D robot while self-improving, without external machinery.
Current platforms: plastic “truss links” (Wyder’s rods) and new triangles (Zhang).
Humanoid price benchmark:$60,000 Unitree G1 (MIT demo — fell on its face).
Moon dust experiments: replica lunar regolith sintered at $35/kg at the lab, preparing a NASA grant application.
Modular robotics’ 2000 promises (Mark Yim, UPenn): versatility, affordability, reliability — all still unmet.
Historical lineage: von Neumann’s automata theory; Samuel Butler’s 1863 “Darwin Among the Machines” letter; Homer Jacobson’s 1958 self-replicating train set.
Why it matters
Three things stand out for a sceptical reader. The first is a genuine technical advance: the lab’s documented ability to have simple modules assemble into a larger robot, reshape itself, and continue improving — without a human operator or external fabrication machinery — is the closest thing yet to a general-purpose “embryo” for robotics. Reproduction, as the piece notes, is the ultimate form of self-repair. If scaled, that turns robots from disposable capex into something more like a population that can regenerate its own stock.
The second is an industry observation. Yim’s candid “they can’t do anything, they’re super expensive and they break all the time” is the correct way to read most modular-robotics demos of the last 25 years. Robotics is no longer computation-bound; it is hardware-bound, and the real-world physics gap between simulation and a chunky plastic triangle crossing a linoleum corridor has not closed much. Daniela Rus’s polite verdict — “very simple as compared to the promise” — is the broader academic consensus on self-reproducing robots.
The third is capital-allocation signal. The reason a Columbia basement lab still makes progress where billions of private dollars have not is that commercial robotics has concentrated on narrow, high-margin niches (warehouse picking, surgical assist, humanoid demos). Lipson’s wider “kingdom of machines” pitch is, for now, a public-science project — the kind of work that only survives if governments and universities underwrite it, which is exactly the infrastructure he worries is being hollowed out.
Takeaway
Treat robot metabolism as a long-dated option rather than a near-term investment thesis. The breakthrough that matters — cheap, reliable modular robots that self-assemble at useful scale — is probably a decade or more away, and the hardware bottleneck is real. But the idea that a truly general-purpose robot will look less like a humanoid Gabe and more like “a bagful of robot” deserves to be part of any serious forecast of physical AI. The immediate market signal to watch: whether any of the Columbia/Penn/MIT modular labs secure NASA or DoD contracts on top of the moon-dust pipeline.
Source: Financial Times, 18 April 2026, Oliver Roeder.