Real-world robustness improvements

Goal: Make model reliably execute search and approach strategy in real life

What I did:
- Collected more real world demonstrations
- evaluated model in real world again
- optimized data loading speed by maxing out cached episodes + parallel workers
- added partial action chunk execution to try smoothing out movements

What worked:
- Model performed better in real life with image history + larger action history
- Training on all data for 2 epochs, then fine tuning on one epoch of only real world data had good results
-

• Model able to reliably execute search and approach strategy
  • ~70-80% completion rate for in distribution situations
  • ~20-30% completion rate for out of distribution situations
    
    
    
    What failed:
• Model struggles to detect objects far from camera
  • almost certainly due to low resolution, a rubiks cube ~1.5 meters from the camera is ~10 pixels wide, all unique features probably unrecognizable at this resolution
• Model struggles in out of distribution settings
  • when the background is very different from most of the dataset, sometimes obvious cube detections in the camera do not trigger an approach movement
  • Can probably be resolved by using a more diverse dataset, and a more generalizable model

Key learning:
- More data + more context for model + cleaner visual input -> acceptable model performance

Next session:
- Begin implementing transformer in pytorch, and continue reading papers
- create a battery voltage monitoring system (need to update and reflash MCU firmware)

Time spent: 6 hours