HOYALABA PERSONAL RESEARCH NOTEBOOK

FIELD NOTES BY ANHOYA

Learning,
one note
at a time.

Robot learning, papers read slowly, and the things I figure out by building.

21 NOTES IN THE ARCHIVELATEST ENTRY / 2026.10.19Read in Korean on Tistory ↗
Paper Figure 3 (NVIDIA, 2025): the GR00T N1 architecture. Eagle-2, with its language component frozen, encodes the image and instruction; embodiment-specific state and action encoders feed DiT blocks that cross-attend to those features; an action decoder produces the action chunk.

Paper Readings / 9 MIN READ

Reading GR00T N1: connecting the islands of robot data

A reading of GR00T N1, from its two-system architecture to the data pipeline that connects different robot embodiments.

A reading note

The notebook

ALL ENTRIES / 21
An orange-and-white cat tucked into striped bedding, from the original post

Engineering Notes · 7 MIN READ

Will my robot policy run on my hardware?

A practical look at inference hardware for ACT, Diffusion Policy, SmolVLA, and larger VLAs, including my own Franka experiment with 80 denoising steps on an RTX 5070 Ti.

Paper Figure 1 (Ha and Schmidhuber, 2018): a cyclist imagines a simplified bicycle and rider, illustrating an internal model of the world. Artwork from Scott McCloud's Understanding Comics, credited in the paper.

Paper Readings · 17 MIN READ

World Models: learning inside a dream

Reading Ha and Schmidhuber's World Models: why a VAE and MDN-RNN do most of the work, how an 867-parameter controller learns, and why a harsher dream can transfer better to reality.

Follow a thread

ONGOING SERIES