3D printing and real-world task completion!

Goal: Print a backing for zima's body and begin collecting real world data

What I did:
- Reinstalled Freecad and moved cad models into the zima repo
- Set up 3d printer again
- Collected 10 episodes of real world data
- Fine tuned sim model using real world data
- ~35 epochs with 3000 samples, reduced LR to 1e-6 for resnet and 1e-5 for action head
- Connected main computer to wifi, allowing for tetherless operation
- Evaluated real world model

What worked:
- Printer setup went smoother than usual
- Havent forgotten how to use FreeCAD, so designing the backing wasn't too bad
-
- Backing printed successfully, all tolerances are fine
- After fine tuning on 10 episodes of real world data, the robot was able to successfully complete the task in the real world with an acceptable success rate (~30%), mainly succeeding in very controlled environments (rubiks cube against white wall backdrop)

What failed:
- Random 3d printer raspberry pi crash (possible brownout?)
- Accidentally stuck a cutter in the printer head while printing, so the backing turned out a ltitel lopsided, but still works just fine

Key learning:
- You can use "Part Binders" in FreeCAD instead of copying the entire part reference over when creating separate parts

Next session:
- Collect more real world data, fine tune on larger dataset (increase test data split size to 20% as well)
- vary the background the cube is in more (is the resolution just too low atp?)
- reduce the frame rate of the collected real world data, undersample it
- evaluate in real life after ~50 episodes are collected, document failure modes

Time spent: 4 hours