RASD

Robotic Arm Segmentation Dataset. 3,269 manually annotated images of robotic arms in diverse environments.

Robots that share a workspace with other arms or with people need to know where robotic arms are in an image, to avoid collisions and to plan motion. When this work started, no public dataset targeted robotic-arm segmentation. RASD (Robotic Arm Segmentation Dataset) fills that gap with 3,269 manually curated and annotated images that cover many kinds of robotic arms in industrial and everyday settings. It was introduced together with a segmentation method that refines a pre-trained model’s mask with spatial and semantic cues from a large vision model (RoboticCC 2025).

Robotic-arm segmentation examples
Examples from the paper: input images, a DeepLabV3 baseline, the proposed method and the ground-truth masks.

My role

I coordinated the construction of the dataset by a ten-person team from several research groups and institutions:

  • collected and screened the images;
  • deployed and administered the CVAT annotation server, including accounts, permissions and task allocation;
  • organised regular review meetings to agree on a common annotation standard and to check the quality and consistency of the masks.

Access

The dataset is available on Hugging Face under the Apache-2.0 licence; access requires accepting the dataset’s conditions.