Zima · 2025-12-11 · 5 hour
Goal: Lock in on finals and improve Zima data collection system to include all relevant manipulation data (arm delta commands, joint state, multi camera, etc.)
What I did:
- explored hugging face LeRobot repo
- https://github.com/huggingface/lerobot
- https://huggingface.co/docs/lerobot/index
- Lots of standardized systems available
- dataset format (already connected to pytorch)
- async_inference server
- policy training
- read lerobot dataset reference
- https://huggingface.co/docs/lerobot/lerobot-dataset-v3
- read lerobot hardware integration guide
- https://huggingface.co/docs/lerobot/integrate_hardware
- Continued reading NLP and CV foundational papers
What worked:
- Locked in on exams
- Got through entire CLiP paper in one sitting
What failed:
- N/A
Key learning:
- Lots of dirty data is better than a little clean data
- LeRobot provides a Robot base class for custom hardware
Next session:
- Begin writing Zima LeRobot interface
- Expose a non ros2 interface? (TCP probably)
- on board computer runs ros2 and tcp interface
- external applications can interface with zima over ros2 (ex. run teleoperation on mac, run )
Time spent: 5 hour