Eryk Halicki Devlog Art

ACT policy training

Zima · 2026-02-22 · 3 hours

Goal: Try training a lerobot act policy

Work Done:
- Got familiarized with lerobot policy training parameters
- Created automated scripts for training and uploading models on vast.ai instances
- Trained ACT policy, available at https://huggingface.co/ehalicki/zima-rubiks-cube-act (most likely not functional yet)

Successes:
- Automated scripts work, full policy training pipeloine works

Failures:
- very slow training (~4 hours or more for the recomended 30k steps on a A4000)
- How long would a VLA take to train with these kinds of budgets???

Next steps:
- Test policy deployment with remote policy server
- Collect more episodes

Time spent: 3 hours