π₀ → π₀.5: A Walked Tour

Six months. Same architecture. New training recipe. A policy that sometimes works in a kitchen it has never seen.

What actually changed between π₀ (Oct 2024) and π₀.5 (Apr 2025)? Almost nothing in the architecture. Almost everything in the training recipe. This site is a 25-minute walk through the five load-bearing ideas that make that sentence honest — starting at multi-head attention and ending at the open-world generalization claim, framed the way PI's own ablations frame it.

The detour through FAST and Knowledge Insulation is mandatory. FAST is the tokenizer that nearly became π₀-FAST — an autoregressive runtime sister of π₀ that didn't survive deployment — whose tokens now live on as a training-time auxiliary signal. Knowledge Insulation is how those tokens coexist with the flow-matching action head without corrupting the VLM. Together they are the recipe; π₀.5 is the result.

Each page anchors one mental model and one or two visualizations that earn their pixels. Stop wherever you have what you came for — every footer calls out the standalone takeaway. About five minutes a page; you can read the climax (Page 4) on its own if that is all you came for.

Sources: π₀ (arXiv 2410.24164), FAST (arXiv 2501.09747), π₀.5 (arXiv 2504.16054), Knowledge Insulation (arXiv 2505.23705), Real-Time Chunking (arXiv 2506.07339).

How to read this site

Two reasonable paths.

Linear (first read)

Reference (second read / lookup)

One-line thesis: π₀.5 = π₀ architecture + KI co-training + a heterogeneous data mixture. Inference is unchanged: continuous flow matching, no FAST tokens at runtime.