Datasets:

pngwn
/

microduck-detection-dataset

license: apache-2.0
task_categories:
  - object-detection
size_categories:
  - 2K<n<10K
pretty_name: Microduck Detection Dataset
tags:
  - object-detection
  - robotics
  - microduck
  - pollen-robotics
  - yolo

Microduck Detection Dataset

Single-class object detection dataset (class 0 = microduck, the 25 cm biped robot duck announced by Pollen Robotics / Hugging Face on 2026-08-27) in YOLO format, built to train a tiny (<5M param) detector.

Composition

Synthetic scene renders (train ~1,040 / val ~60): the robot from the open-source MuJoCo model in pollen-robotics/microduck_rl (full visual-mesh model, scene_allcollisions.xml), rendered headlessly with randomized pose (STAND/SIT/FOLD keyframes + joint noise + yaw), all four official colorways (Cream/Graphite/Lavender/Sky), camera distance/azimuth/elevation/FOV, lighting and floor color. Boxes are exact, obtained by projecting all 140 robot-body geom meshes into the camera.
Synthetic composites (train ~655 / val ~45): the robot cutout (segmented by a robotless render diff, feathered edges) pasted onto ~220 CC0 background photos (picsum.photos) with random scale, position, flip, brightness, 12% occlusion rectangles. Boxes are pixel-exact from the pasted alpha mask.
Synthetic negatives (train ~225 / val ~20): empty floor renders and background-only images with empty labels.
Real press-kit photos (8 train, 4 val): 13 official launch photos from the Pollen Robotics microduck press kit, pre-labeled by an ensemble of two zero-shot detectors (OWLv2 ∩ Grounding DINO, IoU≥0.35 agreement), threshold ≥0.4 (team photo keeps 2 boxes at ≥0.25 — it shows two robots). Train real photos are duplicated 15x to balance against the synthetic majority. microduck-skate.jpg had no confident box and is excluded from labels (kept for qualitative checks).

Splits

split	images
train	2,048
val	134

Files

images/{train,val}/*.jpg, labels/{train,val}/*.txt (YOLO: class cx cy w h, normalized)
data.yaml, gen_meta.json (per-sample generation parameters)

Caveats

Synthetic robot appearance comes from the open-source MuJoCo mesh model; real-photo coverage is only the 13 press photos, so expect the remaining gap to be real-world variation.
Zero-shot labels on the 4 real val images are model-ensemble-verified but not human-verified; treat real-photo mAP as approximate.