Essays and notes on physical AI, robot learning, and the mathematics behind generative models.
An essay on why robotics, the least mature branch of AI, may also be the most open and the most democratizing. It argues that physical AI—through its accessible physical interface, its dependence on operations, hardware, and global data collection, and its still-unmonopolized state—offers a chance to direct AI toward the common good of humanity.
A visual guide to the π0 and π0.5 model architectures. This post provides detailed architectural diagrams illustrating how Flow Vision-Language-Action models process multimodal inputs to generate robot actions through flow matching.
A comprehensive exploration of the Fokker-Planck equation and its applications in diffusion generative models, including proofs, special cases, and connections to flow matching processes. Download the PDF.