Embodied Intelligence in 2026

How to Read This Book

This is a reference text, not a narrative. It is built to be read in three ways.

As a map. Chapters 2 and 3 are directories of the Stanford and MIT robotics landscape — every significant lab, what it believes, what it has built, and where it is weak. Chapters 4 through 6 do the same for companies. If you want to know who works on tactile sensing, or who is betting on simulation, the directories answer that.

As an argument. Chapters 8 through 13 are analytical. They take the positions catalogued in the directories and set them against each other: how much robot data actually exists, which of the five collection strategies is winning, whether one policy can control many bodies, whether the humanoid form factor is engineering or theatre, and what the economics look like once you strip out the demo videos.

As a bibliography. Almost every claim is linked. Where a claim could not be verified, it is marked. There is a lot of confidently-stated nonsense circulating about this field; the flags are the most useful part of the book.

Epistemic status and method

Research was conducted in August 2026 across lab websites, arXiv, company engineering blogs, funding announcements, and trade press, then a separate verification pass re-checked thirty of the most load-bearing claims against primary sources. Where the verifier disagreed with the research, the disagreement is recorded in the text rather than smoothed over.

Three standing caveats:

1. Company-stated numbers are not audited numbers. "99% success," "100 million actions," "1,000 hours per day" are marketing claims until a third party reproduces them. Only one company in this book — Unitree — has published audited financials.
2. Valuations move faster than facts. Several 2026 rounds were reported but not confirmed by the company. These are marked reported.
3. Lab websites rot. Several MIT and Stanford groups have public pages years out of date while publishing actively, and vice versa. Where publication records and web presence conflict, the publication record was preferred.

A note on what "robotics" means here

This book covers manipulation and humanoid/general-purpose embodied AI. It deliberately includes adjacent fields — tactile sensing, biomechanics, prosthetics, world models, simulation — because the interesting arguments of 2026 happen at those boundaries. It deliberately excludes autonomous vehicles, drones-as-drones, and surgical robotics except where they bear on the general-purpose question.