Recursive Self Improvement for Hardware

Can a design engine get measurably better at design with every machine it designs?

The dog's body search, played back: 288 bodies over 24 generations, each trained and scored in 6 scenarios. Shown, each generation's best; the line is the best so far.

Why it is hard

Hardware is designed by a chain of specialists, and each handoff discards information: one person chooses the body, another writes a controller for it, and a third sizes the structure against loads they estimate. Automating one stage is common. Making the whole chain one loop that improves itself is not.

Our angle

One pipeline from task description to printable structure, with the body, its controller and its structure designed together. A language model frames each problem and runs the stages, and verification gates check every stage's output before the next uses it. When a run fails silently, a new gate is written into a shared library, so every later project starts with a better tool.

Where it stands

The first complete pass designed a quadruped from a task description to printable, stress checked geometry in about 2.5 hours on one GPU, searching 288 bodies for a 17.9 kg robot with 4.84 kg of generated structure. Its most useful finding is how such a pipeline fails: silently, with complete and plausible output. Still to come: a fixed benchmark that makes recursive measurable, and a machine built and working in the real world.

Papers

Interactive paper / Jul 2026Body, Brain and StructureThe whole pipeline on a quadruped, from task description to printable structure. Every figure is live.Open
Paper / In preparationThe authoring layerHow the language model frames the problem, drives the stages and writes new gates. Interactive edition to follow.