Activities of Daily Living Object Dataset: Advancing Assistive Robotic Manipulation with a Tailored Dataset

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Over 61 million adults in the US have disabilities, with many relying on wheelchairs and needing help with daily activities.
• Wheelchair-mounted assistive robotic arms can enhance independence, but complex controls hinder usability.
• Automating robotic control via deep learning requires specialized datasets for home environments.

Purpose of the Study:

• To address the lack of suitable datasets for training deep learning models for assistive robotics.
• To present a novel dataset of objects relevant to activities of daily living (ADLs) for robotic manipulation.
• To facilitate the development of user-friendly semi-autonomous control systems for assistive robots.

Main Methods:

• Curated over 112,000 high-quality images from COCO, Open Images, LVIS, and Roboflow Universe datasets.
• Focused on objects pertinent to daily living tasks for individuals with mobility impairments.
• Standardized annotations to YOLO Darknet format and enhanced data quality via YOLOv5x model filtering and manual validation.

Main Results:

• Developed a specialized dataset of over 112,000 ADL objects for assistive robotics.
• Ensured high data quality through rigorous filtering and validation processes.
• Provided a valuable resource for training deep learning models for robotic control.

Conclusions:

• The novel ADL object dataset is crucial for advancing assistive robotic technologies.
• This resource will enable the creation of more effective semi-autonomous control systems.
• The dataset supports innovation in human-robot interaction for home-based assistive applications.
• Empowering individuals with mobility impairments through enhanced independence and autonomy.