unitreerobotics/unifolm-world-model-action — reverse-engineered prompt

Please get this project into a state where I can actually use it without fighting the setup. I want a working Python repo for UnifoLM WMA 0 that can do the main things shown in the README, install cleanly, download the released checkpoints and a small example dataset, run inference with a pretrained model, and show the world model predicting future robot interaction video while also producing actions.

Please also make sure the data prep flow works for my own data in the LeRobot v2.1 style, with clear commands for converting data and starting training in decision making mode or simulation mode. If the repo already has deployment support for Unitree robots, keep that available but optional so I can test everything offline first.

Clean up anything broken, fill in missing glue code, and leave me a simple quick start with the exact commands to run the demo, training, and inference. Look up current docs online if you need to.