Writing

Reinforcement learning from zero to every part of Squint, told as a story on our own arm's numbers, with the maths derived and the code beside it.

AWR and RECAP from the ground up, then how Larchenko used them to teach two arms to fold clothes, traced through his released code.

Why setups a policy always or never solves teach it nothing, and the sampler that fixes it at a million environments.

Action chunking is what made behaviour cloning work. It is also what leaves the policy blind for two seconds at a time.

A 450-million-parameter robot policy that beats a 3.3-billion-parameter one, and only about 1.6 million of its weights are new. We took it apart on our own arms.

The transformer variant, taken apart one shape at a time on our own arm.

Wrist, top and front cameras plus six joint values go in. A coherent chunk of future robot commands comes out.