MuJoCo setup

Goal: Set up a full teleloperated simulation in mujoco with data collection

What I did:
- created a wheeled / tank robot model in MJCF
- debugged weird physics
- added camera functionality to simulation
- added keyboard teleoperation for data collection
- added hdf5 dataset collection for image - action pairs

What worked:

What failed:
- copying someone elses robot didnt work very well
- physics time steps were differing between my python code and the xml
- was this even an issue if it gets overridden?

Key learning:
- writing MJCF xml is tedious
- claude code is not that good at it
- mujoco is not so bad, easier to use than gazebo arguably
- hdf5 is a commonly used dataset format for robotics / scientific data
- h5py converts numpy arrays to hdf5 format

Next session:
- Collect 5 minutes of episodes (~30 episodes)
- task: search for and approach green cube (visual servoing)
- train a basic model to generate actions!
- resnet as fixed feature extractor
- train action head with 2-3 layers
- add CD to obsidian repo so zima devlog auto uploads to eryk.ca
Time spent: 4 hours