CV Model for ARCC S2

I got bored on Friday, so I decided to see if I could train a model using largely synthetic data, with a little bit of fine-tuning at the end on some actual data I have. Used 1000 images for pre-training and ~4000 for fine-tuning. It’s a YOLO26n-pose model, estimating color, icon, and light bar keypoints. I’ll probably train a YOLO27 variant if there’s interest once that releases.

I wouldn’t consider it SOTA in the realm of open source, that’s still SZURP’s model IMO, but if anyone wanted to play around with a model that can do the new 3v3 robot lineup, go ahead :+1:

v2_final.pt.zip (5.8 MB)

Unrelated, but I put together a list of some good CV & Auto-nav resources I came across a few months ago. That can be found here: Notion

For testing this on the 3v3 lineup, the key practical question is how you handle missing or occluded keypoints. Does the model expose per-keypoint confidence, and do you recommend masking low-confidence color/icon/light-bar points before PnP? A short note on keypoint ordering and missing-point handling would make the zip easier to compare with SZURP’s model.