ROBOTIS/AIWorker-Arm-Segmentation-v1

AI Worker Arm Segmentation v1

SAM3-based segmentation for AI Worker arms and grippers

Model Overview

AI Worker Arm Segmentation is a SAM3-based model fine-tuned to segment the recording robot's visible arms, grippers, and attached cameras.

A single model supports all four camera views: left head, right head, left wrist, and right wrist. It produces instance masks, bounding boxes, and confidence scores.

The model runs locally with PyTorch and can support segmentation label generation and frame-by-frame video processing.

Model Details

Property	Details
Base model	SAM3 by Meta
Task	Instance segmentation
Camera views	Left/right head and left/right wrist
Input	RGB image
Output	Instance masks, bounding boxes, confidence scores
Runtime	PyTorch with BF16 on a CUDA GPU
Release date	November 10, 2026

Quick Start

Requires Python 3.12 and an NVIDIA CUDA GPU with BF16 support. Tested with PyTorch 2.10, CUDA 12.8, and an 8 GB GPU.

Installation

pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
pip install huggingface_hub

hf download ROBOTIS/AIWorker-Arm-Segmentation-v1 \
  --include requirements.txt predict.py inference_config.json \
  --local-dir aiworker-arm

pip install -r aiworker-arm/requirements.txt

Run Segmentation

python aiworker-arm/predict.py input.jpg --output-dir prediction

The model weights are downloaded automatically on first use. The same command works with images from any of the four camera views.

For an existing local checkpoint:

python aiworker-arm/predict.py input.jpg \
  --checkpoint /path/to/model.pt \
  --output-dir prediction

Output Files

File	Contents
mask.png	Combined foreground mask at the original image resolution; foreground is white and background is black
instances.npz	Individual masks, bounding boxes in XYXY format, and confidence scores

For videos, apply the model to each frame. The included example processes one image at a time.

Training and Evaluation

Fine-tuned on approximately 12,000 AI Worker images.

Evaluated on 1,317 development images with automatic/AI-assisted references, plus head- and wrist-camera videos. These references are not independent human ground truth.

Limitations

• Distant or partially occluded arm parts may be missed.
• Similar robots and dark background structures may be incorrectly selected.
• People, other robots, and handled objects are not intended targets.
• Only visible target regions are intended to be segmented.

Review predictions before using them as training labels.

License

Derived from SAM3 by Meta and distributed under the SAM License.