Physical AI, from task spec to fleet
Fifteen pipeline stages and the interfaces between them: what each stage consumes and produces, at which rates and in which formats, which methods and tools are current as of October 2026, and how the pieces connect into one running robot. Written for engineers who know the parts and want the whole.
Master pipeline
How to read this site
Stage pages follow one template with 16 sections: where you are, purpose, interface contract, mental model, methods, decision guide, tool table, where it runs, technique tags, reference stacks, running example, pitfalls, numbers, papers, open problems and a self-check.
Integration views cut across stages: an end-to-end trace of two systems, the what-goes-where matrix, the modality × annotation matrix, the on-robot runtime architecture, a paradigm comparison, three reference architectures, the frontier scan, a learning roadmap, a capstone self-test and a glossary.
Filters in the top bar (Cloud, Sim, Edge, RL, IL, Classical, Frontier, Open-source only) apply to every table, method list and diagram on the current page. Search (press /) covers all stages, tools, papers, terms and questions.
Conventions
- FRONTIER
- Method or tool that became important in 2025-04 → 2026-10; linked to the frontier entry.
- Maturity
- research · pilot · production
- Release dates
- From official release pages; “year inferred” marks GitHub listings without a year; “rolling” means no tagged releases.
- Claims
- Company-reported numbers are labelled; disputed or unverified claims say so.
- Superseded
- Kept and struck through in tool tables, so the history stays visible.
The fifteen stages
Integration views
Changelog
Maintained in data/meta.js. See MAINTENANCE.md for how to update stages, tools and frontier entries, and sources.md for all sources by stage.