Skip to main content

Mechanical Intelligence

A live textbook on information in mechanisms

Author

Zach J. Patterson

Published

21 November 2026

Preface

This is a live textbook on how mechanisms carry and process information about the world: how much a robot’s body knows about the forces acting on it, how that knowledge depends on its design, and how much of it reaches the sensors a controller can use. It starts from Patterson (2026), which compared actuator transmissions in legged robots with information-theoretic measures, and it grows as new results come in.

How the book is organized

Part I builds the main tool, the energy metric. It explains why the body’s state is measured in units of energy before noise is added, proves what that buys (invariance under any change of coordinates, the right behaviour when a spring becomes rigid, results that do not depend on a finite-element mesh), and works through the leg designs of the running example. Part II asks whether energy is the right choice: it proves that energy is the only noise model with a short list of physical properties, compares the alternatives, and builds a thermally consistent body in which a design’s job is to route information to its sensors. The research log holds dated notes, the roadmap and the changelog.

How the book changes

The book is not planned in advance. New results first appear as dated research notes in the research log. A note moves into a chapter once its result is checked, its proofs are labelled, and it connects to the existing material; the note stays as the dated record. The roadmap lists the open threads and what happens next on each, and the changelog lists what changed in each release.

Every page opens with its status:

  • exploratory: a new result, checked by code but not yet connected to the rest;
  • draft: connected and proved, still being revised;
  • stable: unlikely to change except for corrections.

What you can rely on

Every number in the text is computed by the Python package mibook in this book’s repository, and a test in tests/ pins it, so a number that drifts fails the build. Figures are generated by code cells when the book is built. A proof that is not complete is titled “Proof sketch” and says what it assumes. Every citation is marked by whether its source was read in full or only through abstracts and citing works (Appendix A — Status of the literature claims).

Reproducing the book

The repository is github.com/CyPhiLab/mechintelligence. With uv installed:

uv sync
uv run pytest
uv run quarto render

The computations depend on the analysis library from zpatty/quadruped-mech-info, pinned to a fixed commit.

How to cite

Cite the version you read, since pages change between releases:

Patterson, Z. J. Mechanical Intelligence: A live textbook on information in mechanisms. Version 0.1, 2026. https://mechintelligence.cyphilab.com/

Acknowledgments

The text and code were developed with the help of Claude Code. Every result is checked by the code in the repository.