factory-safety-isaac YOLO weights
YOLO weights for factory-safety-isaac โ a body-camera-only factory safety agent built on NVIDIA Isaac Sim. These are the two trained detection models the project uses; no other model files are hosted here (MediaPipe's hand landmarker and the Qwen2.5-7B LLM are pulled from their own upstream sources at first run, see the project README).
Files
| File | Model | Classes | mAP50 | mAP50-95 | Precision | Recall |
|---|---|---|---|---|---|---|
warehouse_v3/best.pt |
YOLO26s, 960px | 19 (hazard/safe states, 8 tool types, cart, worker, cone, danger sign) | 0.898 | 0.689 | 0.920 | 0.827 |
pinch_v1/best.pt |
YOLO, 1-class lightweight | pinch_point (conveyor pinch-point detector) |
0.966 | 0.760 | 0.983 | 0.910 |
warehouse_v3 per-class mAP50
spill 0.920 ยท spill_marked 0.937 ยท stack_unstable 0.957 ยท stack_stable 0.960 ยท ext_fallen 0.907 ยท ext_blocked 0.909 ยท ext_ok 0.938 ยท worker 0.950 ยท cone 0.976 ยท danger_sign 0.950 ยท hammer 0.862 ยท screwdriver 0.828 ยท saw 0.806 ยท power_saw 0.901 ยท pickaxe 0.845 ยท shovel 0.819 ยท wrench 0.887 ยท drill 0.915 ยท cart 0.794
Trained on 7,700 synthetic images generated in Isaac Sim (Replicator auto-labeling), 70% bodycam / 30% free-viewpoint.
Usage
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
path = hf_hub_download("IBDPLab/factory-safety-isaac-yolo", "warehouse_v3/best.pt")
model = YOLO(path)
Or with the project's own scripts (scripts/run_patrol.py --weights <path>,
scripts/run_patrol.py --pinch-weights <path>).
License / provenance
Trained entirely on synthetic data (NVIDIA Isaac Sim warehouse assets + Poly Haven CC0 tool scans). See the main repository for full training/data-generation details and limitations (real-photo generalization has not been fine-tuned).
ํ๊ตญ์ด
factory-safety-isaac์ฉ YOLO ๊ฐ์ค์น์ ๋๋ค. factory-safety-isaac์ NVIDIA Isaac Sim ๊ธฐ๋ฐ์ผ๋ก ๊ตฌ์ถํ, ๋ฐ๋์บ ์์๋ง ์ฌ์ฉํ๋ ๊ณต์ฅ ์์ ์์ด์ ํธ์ ๋๋ค. ์ด ์ ์ฅ์์๋ ํ๋ก์ ํธ์์ ์ฌ์ฉํ๋ ํ์ต๋ ํ์ง ๋ชจ๋ธ 2์ข ๋ง ์ฌ๋ผ๊ฐ ์์ต๋๋ค. MediaPipe ์ ๋๋๋งํฌ ๋ชจ๋ธ๊ณผ Qwen2.5-7B LLM์ ์ต์ด ์คํ ์ ๊ฐ ์๋ณธ ์ถ์ฒ์์ ๋ด๋ ค๋ฐ์ต๋๋ค(ํ๋ก์ ํธ README ์ฐธ๊ณ ).
ํ์ผ
| ํ์ผ | ๋ชจ๋ธ | ํด๋์ค | mAP50 | mAP50-95 | Precision | Recall |
|---|---|---|---|---|---|---|
warehouse_v3/best.pt |
YOLO26s, 960px | 19๊ฐ (์ํ/์์ ์ํ, ๊ณต๊ตฌ 8์ข , ์นดํธ, ์์ ์, ๋ผ๋ฐ์ฝ, ์ํ ํ์งํ) | 0.898 | 0.689 | 0.920 | 0.827 |
pinch_v1/best.pt |
YOLO, ๋จ์ผ ํด๋์ค ๊ฒฝ๋ ๋ชจ๋ธ | pinch_point (์ปจ๋ฒ ์ด์ด ๋ผ์ ์ง์ ํ์ง) |
0.966 | 0.760 | 0.983 | 0.910 |
warehouse_v3 ํด๋์ค๋ณ mAP50
spill 0.920 ยท spill_marked 0.937 ยท stack_unstable 0.957 ยท stack_stable 0.960 ยท ext_fallen 0.907 ยท ext_blocked 0.909 ยท ext_ok 0.938 ยท worker 0.950 ยท cone 0.976 ยท danger_sign 0.950 ยท hammer 0.862 ยท screwdriver 0.828 ยท saw 0.806 ยท power_saw 0.901 ยท pickaxe 0.845 ยท shovel 0.819 ยท wrench 0.887 ยท drill 0.915 ยท cart 0.794
Isaac Sim์์ ์์ฑํ ํฉ์ฑ ์ด๋ฏธ์ง 7,700์ฅ(Replicator ์๋ ๋ผ๋ฒจ๋ง)์ผ๋ก ํ์ตํ์ต๋๋ค. ์์ ๊ตฌ์ฑ์ ๋ฐ๋์บ 70%, ์์ ์์ 30%์ ๋๋ค.
์ฌ์ฉ๋ฒ
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
path = hf_hub_download("IBDPLab/factory-safety-isaac-yolo", "warehouse_v3/best.pt")
model = YOLO(path)
ํ๋ก์ ํธ ์์ฒด ์คํฌ๋ฆฝํธ๋ก๋ ์ฌ์ฉํ ์ ์์ต๋๋ค:
scripts/run_patrol.py --weights <๊ฒฝ๋ก>, scripts/run_patrol.py --pinch-weights <๊ฒฝ๋ก>
๋ผ์ด์ ์ค / ๋ฐ์ดํฐ ์ถ์ฒ
์ ๋ ํฉ์ฑ ๋ฐ์ดํฐ(NVIDIA Isaac Sim ์ฐฝ๊ณ ์์ + Poly Haven CC0 ๊ณต๊ตฌ ์ค์บ)๋ก ํ์ตํ์ต๋๋ค. ํ์ตยท๋ฐ์ดํฐ ์์ฑ ์ธ๋ถ ์ฌํญ๊ณผ ํ๊ณ๋ ๋ฉ์ธ ์ ์ฅ์๋ฅผ ์ฐธ๊ณ ํ์ธ์. ์ค์ฌ ์ด๋ฏธ์ง์ ๋ํ ํ์ธํ๋์ ์์ง ์งํํ์ง ์์์ผ๋ฏ๋ก, ์ค์ ํ๊ฒฝ์์์ ์ผ๋ฐํ ์ฑ๋ฅ์ ๊ฒ์ฆ๋์ง ์์์ต๋๋ค.