Floor check
A forecast can't see the floor Spot walks on. Our fine-tuned open model flags standing water in a camera frame so the dispatcher can hold that waypoint. It runs in this browser with transformers.js.
A forecast can't see the floor Spot walks on. Our fine-tuned open model flags standing water in a camera frame so the dispatcher can hold that waypoint. It runs in this browser with transformers.js.
Pick an image. The model downloads once, then runs in this browser with no server.
| Model | Accuracy | Wet recall |
|---|---|---|
| Always "dry" | 56.9% | 0% |
| CLIP zero-shot (baseline) | 71.8% | 72.2% |
| Our fine-tuned ConvNeXt-Tiny | 99.6% | 99.1% |
Honest number: 92.7% wet recall on images from a single source (41 wet, 2 scenes). The full score is partly inflated because dry and wet images mostly come from different datasets.
Off-domain: in our Python evaluation it flagged 2 of 6 dry NextEra kit images as wet (gauge p=0.51, thermal p=0.63). In the browser the gauge lands just under 0.5 because image resizing differs slightly; the thermal image is still a false alarm. Trained on public street scenes, not plant floors. Next step: fine-tune on real Spot frames.
Model: builtbyd3v/spot-wetfloor-convnext-tiny. Data: Mendeley water/wet-surface images (CC BY 4.0) and a CC0 Kaggle public-spaces set. Base model facebook/convnext-tiny-224 (Apache-2.0).