blackbeedrones/imav-2025-gate

IMAV 2025 Gate Detection - YOLOv11n

Gate detection model for IMAV 2025 Indoor Competition - Mission 1.

Competition Context

The 16th International Micro Air Vehicle Conference and Competition (IMAV 2025) took place in San Andrés Cholula, Puebla, Mexico. The competition theme was "Search and Rescue", inspired by Mexico's seismic activity and the need for micro air vehicles in disaster response scenarios.

Indoor Arena

3D view of the 10m × 10m indoor arena. Tunnel-like spaces are formed by gates in blue (1.5m), green (1m), and red (0.5m).

Target Object

Gate Specifications:

• Tube diameter: Ø 38.1 mm (1.5")
• Total height: 2m (base + window)
• Three tunnel sizes:

Tunnel	Window Size	Base Height	Points
Wide (blue)	1.5m × 1.5m	0.5m	1
Medium (green)	1.0m × 1.0m	1.0m	2
Small (red)	0.5m × 0.5m	1.5m	3

Mission 1: Enter the Room

The MAV must navigate through one of four entry options:

1. Free passage (0 pts): 1m wide open path
2. Wide tunnel (1 pt): 5 aligned gates, 1.5m × 1.5m window
3. Medium tunnel (2 pts): 5 aligned gates, 1.0m × 1.0m window
4. Small tunnel (3 pts): 5 aligned gates, 0.5m × 0.5m window

Each tunnel is 2m in length (5 gates aligned horizontally).

Performance

Metric	Value
mAP@50	0.995
mAP@50-95	0.991
Precision	1.0
Recall	0.999

Training Curves

Confusion Matrix

Validation Predictions

Model Formats

Format	File	Use Case
PyTorch	gate_yolov11n.pt	Training, fine-tuning
ONNX	gate_yolov11n.onnx	Cross-platform inference
TensorRT	gate_yolov11n.engine	Jetson Orin Nano Super

Training Configuration

Parameter	Value
Base model	yolo11n.pt
Epochs	27 (early stopping)
Image size	640×640
Batch	Auto
Optimizer	Auto
LR	0.01 (cosine)
Augmentation	Mosaic, RandAugment

Full config: train/args.yaml

Usage

mirela-sdk

from mirela_sdk.ai.detection import Detector

detector = Detector("blackbeedrones/imav-2025-gate:best.pt")
detector.load()

result = detector.detect(image, conf=0.5)
for det in result:
    print(f"Gate: {det.confidence:.2f} at {det.center}")

Ultralytics

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict(image, conf=0.5)

References

• IMAV 2025
• Rulebook
• mirela-sdk