Tool Detector
YOLOv8m detection model for handheld workshop tools, trained on synthetic rendered data and fine-tuned on real annotated images.
Intended to give a UR5 robot arm (5 kg payload, ~850 mm reach) tool identity and bounding-box location for pick-and-place tasks.
Model Details
| Property | Value |
|---|---|
| Architecture | YOLOv8m |
| Input size | 640 × 640 |
| Classes | 6 (current fine-tuned model) |
| Base weights | COCO pretrained yolov8m.pt |
| Fine-tuned on | Real annotated images (Label Studio export) |
Classes
| Index | Name |
|---|---|
| 0 | allen_key_set |
| 1 | allen_key |
| 2 | screw_driver |
| 3 | hammer |
| 4 | cutters |
| 5 | pliers |
Usage
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
weights = hf_hub_download(repo_id="x-not/tool-detector", filename="best.pt")
model = YOLO(weights)
results = model("your_image.jpg", conf=0.3)
results[0].show()
Or via the CLI:
huggingface-cli download x-not/tool-detector best.pt --local-dir models/
yolo detect predict model=models/best.pt source=your_image.jpg conf=0.3
Training
Synthetic data generated with BlenderProc (Blender + Cycles): tools rendered as RGBA sprites at randomized pose/lighting, composited onto workshop background photos. Fine-tuned on real images annotated in Label Studio.
Source code and full training pipeline: github.com/NikolaasBender/Tool-Detector