Eryk Halicki Devlog Art

Real-world policy evaluation

Zima · 2025-11-22 · 4 hours

Goal: evaluate the policy in real life and iterate

What I did:
- reduced max wheel speed and compromised on camera resolution - fps tradeoff, reducing motion blur
- tried on device inference - fail
- added frame stacking / image history input to model, hoping to improve real world performance

What worked:
- fine tuning on just real world data for a few epochs improves real world performance noticeably

What failed:
- wasted more time trying to get camera driver to work, to no avail
- on device inference caused brownout?
- image stacking is starting to eat into my gpu memory, had to reduce batch size to 32 to not run out of mem
- Model with shortened action history (4) was not able to do proper search behavior

Key learning:
- Real world is difficult

Next session:
- Collect more real world data
- test new image stacked model in real life

Time spent: 4 hours