Zima · 2025-11-21 · 4 hours
Goal: Collect more real world data and continue model fine tuning
What I did:
- Collected 20 more episodes of real world data, with varying scene setups and starting positions
- Further fine tuned model using 30 episodes of real world data
- Evaluated model in real life, mixed results
- motion blur in real life but not in simulation
- Tried fixing motion blur by reducing exposure time
What worked:
- Model trained with both real world and sim data was able to still complete tasks in sim
- data mixture is very sim heavy though, (210 vs 30 episodes) so this isnt too unexpected
What failed:
- model may be overfitting to the real world dataset, training and test accuracy very high (>90%) but task completion rate is mediocre
- funny artifact, policy seems to have overfit to a handful of training examples where the rubiks cube was placed near a carton of soy milk, and since the soy milk carton is more visible than the cube, the model gets tripped up and approaches the carton of milk sometimes
- Motion blur is harming model performance significantly in real life
- Rubiks cube is almost imperceivable at a distance
- Struggling to set manual exposure time successfully. Cameras not listening
- Battery died at end of session
- multi meter battery is dying, so readings are inaccurate, so have no idea if the BMS low-voltage protection kicked in
Key learning:
- Starting to really see the real world noisiness, simulation is very trivial compared to irl
Next session:
- create a battery voltage monitoring system (need to update and reflash MCU firmware)
- evaluate sim+real data policy in real life
- Add motion blur to the sim data?
- try debugging camera blur again
- set exposure time low and check if performance improves
- try mjpg 320x240, but moving the crop to center somehow
Time spent: 4 hours