By kind · 按形式
Note笔记
One topic, worked through: concept, mechanism, experiment — 5 notes.
- Note2026.09Action Tokenization: Discretizing Continuous Actions
The chain "continuous action → discrete token → continuous action", unpacked into the background each link requires — π₀-FAST's discrete autoregressive route against FSQ, Binning and flow matching.
- Note2026.08Autoregressive Models: BERT / GPT and Prefix-LM
Which sequence-modelling paradigms exist, what separates BERT from GPT, and why robot VLAs such as π₀-FAST use a hybrid of the two called Prefix-LM.
- Note2026.08Flow Matching ODE Solvers, Illustrated
Sampling an image is numerically integrating an ODE. Five classic solvers — Euler, Midpoint, Heun, RK4, Dopri5 — taken apart, with real integrations behind every animation.
- Note2026.08Fourier & DCT: From Frequency to Energy Compaction
From Fourier series to the DCT's energy-compaction property — why a DCT makes a smooth signal's energy pile up onto a handful of low-frequency coefficients.
- Note2026.08Lie Theory: An Intuitive Introduction
Lie theory carries optimization over curved spaces — rotations, poses — back into ordinary vector calculus, from manifolds and tangent spaces to hat/vee, exp/log and a hand-written SO(3).