HuntLab’s Physical AI and ROS 2 articles are written with AI and backed by experiments we actually run in an automated environment. This page describes that process as it is. The principles are in our verification and correction policy (Korean).
Who writes the articles
Articles are drafted and reviewed by Anthropic’s AI (Claude). “Hunt” is the site’s operating brand, not a human author. The operator sets the rules for picking topics, reviewing and publishing, and handles error reports and corrections. A person does not read and approve every article before it goes live.
How one article is made
- Pick a real question from places where developers get stuck, such as Robotics Stack Exchange and GitHub issues. Questions we already covered are skipped.
- Check the sources: only topics that can be verified against primary sources (official docs, official source code). At least three top existing answers are read and their gaps recorded.
- Run it in the environment below, usually 5 times per measurement. Worked examples are recomputed in code.
- Write: measured values go only in the measured section; interpretation stays outside it.
- Number gate: a program checks that every number in the measured section exists in the actual result files. One missing number blocks publication.
- Independent review: a separate AI review step checks sources, accuracy and overlap with existing articles. Articles below the bar are rewritten or held.
- Publication check: after the scheduled release, the live page is fetched again to confirm title and body.
Test environment
- A Linux cloud server (AWS Lightsail, 2 vCPU, 4 GB RAM).
- Docker containers, by default the official ROS 2 Jazzy image. When comparing versions, packages are pinned from the official ROS snapshot repository (snapshots.ros.org).
- These are software experiments, not tests on real robot hardware; we do not claim the same results on hardware.
Reproduction code
Experiment code and result files are public: github.com/sungpyo9053/blog/tree/main/experiments. Anyone can re-run them and check the numbers.
Found a mistake?
Open an issue on the GitHub repository or use the contact page. Confirmed errors are corrected in the article with the correction date.