SHolic/hurocup-yolo-models

HuroCup YOLO Models

YOLO26 detectors for five FIRA HuroCup events (archery, basketball, marathon, weight lifting, united soccer). They are the perception models of the vision–control architecture described in Efficient Integration of Vision and Control for Resource-Constrained Humanoid Robots, deployed on a ROBOTIS OP3 and a Hiwonder AiNex.

The training images and labels are in SHolic/hurocup-yolo-dataset.

Models

Every run is a YOLO26 model trained with Ultralytics 8.4.80 for 100 epochs at batch size 16. Each directory holds the best.pt weights, the OpenVINO and ncnn exports of best.pt, and the training record (args.yaml, results.csv, curves and confusion matrices).

Deployed models (imgsz 320)

Directory	Event	Classes	Base	Deployed on
archery	Archery	Target	yolo26n	OP3, AiNex
basketball	Basketball	Ball, Basket	yolo26n	OP3, AiNex
marathon	Marathon	forward, left, right	yolo26n	OP3
marathon_marker	Marathon	marker	yolo26n	AiNex
weight_lifting	Weight lifting	Bar	yolo26n	OP3, AiNex
united_soccer_clean	United soccer	Ball	yolo26n	OP3, AiNex

The two platforms use different marathon models. The OP3 classifies the three direction signs directly. The AiNex is about 350 mm tall, so its camera sits low and a distant arrow subtends too few pixels to classify. It therefore runs the class-agnostic marathon_marker detector, trained on AprilTags and arrow signs as a single marker class, which keeps the marker trackable at range; AprilTag detection and arrow classification take over the turn decision once the robot is close.

Ablation and auxiliary runs

Directory	Purpose	Base	imgsz
archery_640, basketball_640, weight_lifting_640	Input-resolution comparison	yolo26n	640
archery_s, basketball_s	Model-capacity comparison	yolo26s	320
united_soccer	Three-class united soccer detector (obstacle, football, goalpost)	yolo26n	320

The united soccer model used on both platforms, and referred to as united_soccer in the paper, is united_soccer_clean. The three-class united_soccer run is not part of the reported results.

Accuracy of the deployed models

Best epoch by Ultralytics fitness, on each event's validation split (number of validation images in brackets).

Model	mAP@0.5	mAP@0.5–0.95	Precision	Recall
archery (17)	0.995	0.993	0.992	1.000
basketball (15)	0.987	0.843	0.940	0.951
marathon (45)	0.981	0.894	0.930	0.942
weight_lifting (15)	0.991	0.766	0.937	0.992
united_soccer_clean (15)	0.988	0.842	0.969	0.933
marathon_marker (7)	0.813	0.447	0.857	0.806

Each event has 50–300 self-collected images (see the dataset card), so these figures describe fit to a small, single-site dataset rather than general-purpose detection accuracy.

Files

<run>/
├── args.yaml                      training arguments
├── results.csv, results.png       per-epoch metrics
├── Box*_curve.png, confusion_matrix*.png, labels.jpg
└── weights/
    ├── best.pt                    PyTorch weights (best epoch)
    ├── last.pt                    PyTorch weights (last epoch)
    ├── best_openvino_model/       OpenVINO IR, used on the OP3 (Intel i3-6100U)
    └── best_ncnn_model/           ncnn, used on the AiNex (Raspberry Pi)

All exports were made with Ultralytics 8.4.80 at imgsz 320 (640 for the _640 runs), with end2end: false, so NMS is not part of the exported graph.

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

path = hf_hub_download("SHolic/hurocup-yolo-models", "archery/weights/best.pt")
model = YOLO(path)
results = model.predict("image.jpg", imgsz=320)

To use an export, download its whole directory and pass the directory to YOLO:

from huggingface_hub import snapshot_download
from ultralytics import YOLO

root = snapshot_download("SHolic/hurocup-yolo-models", allow_patterns="archery/weights/best_ncnn_model/*")
model = YOLO(f"{root}/archery/weights/best_ncnn_model")

On the robots the models run with Ultralytics 8.4.45 and OpenVINO 2026.1.0 (CPU plugin) on the OP3, and Ultralytics 8.1.31 and ncnn 1.0.20260526 on the AiNex.

Limitations

• Trained on 50–300 images per event, collected in one laboratory under its lighting, with the competition props used there. Expect lower accuracy on other fields, props or lighting.
• The models detect competition objects only; they are not general-purpose detectors.
• basketball loses accuracy at high IoU thresholds (mAP@0.5–0.95 0.843), which follows from the 320 input size.

License

The models were trained with Ultralytics and are released under AGPL-3.0, the license of the Ultralytics framework.

Citation

@article{sheng2026efficient,
  author  = {Sheng, Haoliang and Zheng, Xingyu and Lau, Meng Cheng},
  title   = {Efficient Integration of Vision and Control for Resource-Constrained Humanoid Robots},
  journal = {IEEE Access},
  year    = {2026},
  note    = {Under review}
}

This work builds on our earlier vision system:

@article{lau2026modular,
  author  = {Lau, Meng Cheng and Pottier, Nicolas},
  title   = {A Modular Vision System for Practical Object Detection on Resource-Constrained Humanoid Robots},
  journal = {Biomimetics},
  volume  = {11},
  number  = {6},
  pages   = {363},
  year    = {2026},
  doi     = {10.3390/biomimetics11060363}
}