GitHub radar
What an AI agent built in one hour from a robotics research paper
This repo is an unofficial build of a research paper on humanoid robot control, written by an AI coding agent in one hour from the paper's text alone. The guide shows what that kind of agent work can teach you.
An AI agent wrote this code from the SuperMimic research paper, with no official code from the authors. The model controls a simulated 29-joint Unitree G1 humanoid robot, reproducing 40 motions: dancing, running, jumping, falling, and getting up.
When it helps
Useful if you have Linux and a powerful NVIDIA GPU like the RTX PRO 6000 or H100 and want to explore robotics. Not useful on a laptop or mid-range gaming GPU — the hardware bar is too high, and the author himself calls the code "too sloppy."
Pros
- Open MIT license — free to use and modify
- Model learned to reproduce 40 motions: dance, run, jump, fall, get up
- A clear example of an AI agent implementing a research paper in an hour
- Author openly warned about the code's risks and roughness
Cons
- Needs heavy hardware: Linux, NVIDIA RTX PRO 6000/H100-class GPU, Docker, Isaac Lab
- Written in an hour with no code review, the author calls it "too sloppy"
- Won't run on a laptop or mid-range gaming GPU
How to set it up — step by step
- 1Check you have Linux and a powerful NVIDIA GPU (RTX PRO 6000 or H100) — without it, nothing runs.
- 2Install Docker with GPU support and clone zita-ch/supermimic_vibed from GitHub (https://github.com/zita-ch/supermimic_vibed).
- 3Run tools/setup.sh to build the Isaac Lab image and the supermimic Docker container.
- 4Run tools/download_data.sh to fetch the 40 LAFAN1 motions and the G1 robot model.
- 5Start training with CPUS=0-23 ./run_train.sh lafan1_main and track progress via tmux and the eval.csv logs.
- 6No such hardware? Send your agent the text in the block below to repeat the trick on your own task.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
If you want to try the same spirit without robots: ask your agent to spend one hour building a small, risky prototype from the description of someone else's paper or a hard task in your own field — not to install this repo, but to copy the move.
Heavy hardware: you need Linux, an NVIDIA GPU (the README mentions the RTX PRO 6000 or H100), Docker with GPU support, and the Isaac Lab framework — a regular laptop or a mid-range gaming GPU will not do here.
Open on GitHub▌ More finds