Method

How we know

We test one question: can an AI run a robot while a person can still stop it, limit it and watch it? This page explains how a test is run, how a claim gets published, what we do not claim, how we earn money and how to send us a correction.

Method as of 15 November 2026. Our research is done by AI agents that BotInfo runs, under rules our founder sets.

How a test is run

Simulation and mock hardware only

Every test so far runs in a simulator or reads public code. We have not run our code on a real robot. A result shows what one model did in simulation, not what a real robot will do.

A sealed simulator

Each simulation test is one Python script that runs the MuJoCo physics engine inside a locked sandbox on our machine. It has no network. It can read only the Python runtime, its libraries and its own run folder, and it can write only its output folder. Its limits are enforced from outside, where the script cannot change them.

Pinned models and inputs

Robots are published simulation models, and every input is a public file pinned to an exact commit. For the Unitree G1 we use MuJoCo Menagerie's model (BSD-3-Clause) and the models Unitree publishes with its walking policies. Anyone can load the same file and run the same script from our GitHub.

A real AI, and a stop mid-task

In our lab, a real AI model (Claude, made by Anthropic) drives a simulated robot through BotInfo, our open-source layer between an AI and a robot. It gets only BotInfo's tools, in an empty folder, and is not told it is being tested. The lab plays the owner and presses stop at a moment the test sets in advance. We keep the full record: the audit log, the AI's transcript and its final report.

Lab records stay private by default. When a video uses a lab result, it names the run, and the checker reads our own stored copy of the scores, not the agent's.

How a video is made

  • Our simulation footage is rendered by our own software from the simulator's numbers, using the same pinned models.
  • A label in the corner of every frame says what is on screen, such as SIMULATION, AI VOICE or whose footage it is.
  • The voice is an AI voice (Google Gemini text-to-speech), and we say so.
  • On YouTube, every video we make this way is marked as altered or synthetic content.
  • Before upload, a second agent checks the title, the description and every source. It fails any video that uses someone else's footage, images or sound without written permission, except short clips we comment on, credited to their source, and music under an open licence (CC0, public domain or CC BY).

How a claim gets published

  1. Source and date. Every claim carries its source and the date we read it. Code is cited at a pinned commit wherever we can, so the line we read stays the line you see.
  2. A grade. Findings are graded: checked (a second agent reopened the source), our data (our own reading, not yet reopened), our inference (our reading of checked facts, labelled), or unknown, said aloud.
  3. A second agent checks it. Before a repo or a video goes out, an AI agent that did not write it reopens every cited source and confirms each claim. A failed check publishes nothing, and the writer gets the reasons.
  4. Code is tested too. A change to the BotInfo code is read by a second agent, then its tests run on GitHub's machines. Only a passing change is merged.
  5. Fixed fast, and dated. When we learn a claim is wrong, we correct it and date the correction where the claim appears.

What we do not claim

  • We do not claim that a simulation result holds on a real robot.
  • We do not say any robot, stop or setup is right for your use. We report what happened in a test and the limits of that test.
  • A code reading describes one code path at one version. It is not a test of a robot and not a rating of a project.
  • When we cite someone else's measurement, we say so, and we say whether we repeated it.
  • Prices and web pages change. The date is part of every claim.
  • Nothing we publish is legal advice or investment advice.

How we earn money

  • A referral fee on Noetix Bumi links. Our Noetix Bumi pages carry links to a seller that pays us a fee on orders. Those pages say so beside the links, and our affiliate disclosure explains it. No other page on this site carries a paid link.
  • Referral fees on buyer introductions. When an organisation asks us where to buy a robot, we may introduce it to a seller that pays us a referral fee if it buys.
  • Evidence dossiers. We sell a full evidence file on a specific robot to organisations, by written quote. What a dossier contains.

No maker or seller pays us for a test result or a finding.

Send a correction

If you think a claim is wrong, write to [email protected] with the claim and your source, or open an issue on the GitHub repo that holds the result. We check it against the source, fix what is wrong and date the fix.