allenai/molmoact ? reverse-engineered prompt

Build me a clean, working setup for MolmoAct so I can train and run the model from this repo without fighting the environment.

I want the project to install smoothly, preferably with the provided Dockerfile as the main path, and also have clear local setup steps for Python 3.11 and PyTorch. Make sure the repo can handle the full flow from preprocessing robot data into the action reasoning format, then fine tuning, merging LoRA if needed, running inference, and visualizing results. If there are missing pieces or old instructions, update them so a normal person can follow along. It should also be easy to reproduce the training and evaluation workflows mentioned in the README, including SimplerEnv and LIBERO style evaluation. If anything depends on current external docs or checkpoints, look them up online and wire the code and instructions around that.