Embodied Turing Machines Stateful Code for Robot Recursive Self-Improvement paper:

Stateful code for robot recursive self improvement is the embodied version of the agent harness problem. Durable state, clear improvement loops and measurable gains matter as much for factory robots as for enterprise agents.

Robots rewriting themselves. Nothing could possibly go wrong.

gating merges on beating rerun noise is a nice touch. most agent self-improvement loops i've seen happily keep whatever got a lucky seed

recursive self-improvement loop is ambitious. does the stateful code persist across hardware failures, or does a robot reset lose its learned improvements?

Robots self-improving, nothing to see.

Robots keeping a shared code library so they stop repeating their mistakes. That's more self-improvement than most New Year's resolutions manage 🙏

Stateful robot loops go sideways once sensors start lying

Robots upgrading themselves now. We're cooked.

Gostei muito!