FabienDanieau/reachy-mini-detection

Reachy Mini detector (YOLO11n) — round 2

Single-class object detector (class: reachy_mini) fine-tuned from COCO-pretrained YOLO11n on 1644 training / 80 validation images. Dataset v2 mixes:

• posed views rendered from the official rigged Blender model (pollen-robotics/reachy_mini_blender): varied head/body/antenna poses, camera angles and lighting, composited onto real photo backgrounds — FabienDanieau/reachy-mini-detection-dataset-v2
• the round-1 composites (product-photo cutouts): FabienDanieau/reachy-mini-detection-dataset

Validation on the v2 val split: precision=0.972 recall=0.938 mAP50=0.965 mAP50-95=0.890 @ imgsz=640. Same-model comparison on the v1 val split: precision=0.875 recall=0.750 mAP50=0.810 mAP50-95=0.626.

Files:

• yolo11n_reachy_mini.pt — PyTorch weights (desktop / Reachy Mini Lite use)
• yolo11n_reachy_mini.onnx — ONNX export
• ncnn_model/ — NCNN export for Raspberry Pi (Reachy Mini Wireless, CM4)

NOTE: still trained on synthetic data only; expect a drop on real camera frames. Fine-tune on ~100-200 real captures from the robot's camera for best results.