Modelpublic

zjumty/so101-yolo11n-seg-plastic-cup

sourceHugging Faceagpl-3.0updated 23d agoView on Hugging Face

0likes89downloads

Model Card

SO-101 YOLO11n-Seg plastic cup model

This repository contains a YOLO11n-Seg model fine-tuned to segment the plastic_cup class in the SO-101 simulation scene. It also includes the synthetic dataset used for training and evaluation.

Repository contents

—best.pt: fine-tuned Ultralytics checkpoint.
—model/metrics.json: metrics from the held-out synthetic test split.
—model/training-config.yaml: portable training settings.
—model/evaluation/: test plots and prediction previews.
—dataset/: YOLO segmentation images, polygon labels, per-sample truth, and the dataset manifest.
—SHA256SUMS: hashes for the uploaded files.

Dataset

The dataset has 1,200 synthetic 640 x 480 images rendered from the SO-101 MuJoCo scene through task_camera. Object-ID segmentation supplied the label masks. The split is fixed by seed range:

Split	Images	Labels	Seeds
train	800	800	100000-100799
validation	200	200	200000-200199
test	200	200	300000-300199

The only class is plastic_cup (class_id=0). Samples include no-cup, one-cup, and multi-cup scenes. The manifest records the scenario, seed, visible instance count, image path, label path, and synthetic truth path for each sample.

Dataset generator commit: 2be8df09302feabffc7f028b16c90d06867f8055 (build(perception): pin training runtime and labeled overlays).

Training

The run started from yolo11n-seg.pt and used Ultralytics segmentation training with these main settings:

Setting	Value
image size	640
epochs	100
batch size	32
seed	20260831
deterministic	true
device	CUDA
AMP	false
optimizer	auto

Checkpoint SHA-256: f281d25258493e2c7c220dd1d84a7ca4f0501adf99ed4a921a065d74ace40781.

Synthetic test results

Metric	Boxes	Masks
precision	0.999735	0.999735
recall	1.000000	1.000000
mAP50	0.995000	0.995000
mAP50-95	0.995000	0.973662

These numbers describe the fixed synthetic test split. They do not establish accuracy on real cameras, unfamiliar cup appearances, or physical grasp success.

Use with Ultralytics

python

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("image.png")

The checkpoint inherits the licensing requirements of its Ultralytics YOLO11 base model. The repository is private because a separate license has not been declared for the synthetic dataset.