Reliable Vision-Based Grasping Target Recognition for Upper Limb Prostheses.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Computer vision shows promise in wearable robotics for tasks like grasping but faces challenges such as noisy data and complex environments.
• Existing methods struggle with real-world variability, limiting the reliability of vision-based systems in prosthetic applications.

Purpose of the Study:

• To develop a robust, vision-based framework for upper limb prosthesis control, specifically for grasping assistance during arm reaching.
• To enhance the reliability and accuracy of grasping target recognition in challenging, real-world scenarios using advanced machine learning.

Main Methods:

• Leveraged Bayesian deep learning (BDL) to create a novel framework capable of quantifying uncertainties in vision-based grasping.
• Developed a probability calibration network to integrate diverse uncertainty measures for real-time decision-making.
• Utilized a 3-D simulation platform to rigorously test algorithm performance under various challenging conditions.

Main Results:

• The BDL framework successfully measured model and data uncertainties, improving grasping target recognition.
• A probability calibration network effectively fused uncertainty measures for reliable online decision-making.
• The system demonstrated enhanced performance in a simulated 3-D environment, addressing common practical challenges.

Conclusions:

• The developed vision-based framework significantly improves the reliability of grasping assistance for upper limb prostheses.
• Integration into a shared control system demonstrated fluent and effective target reaching and grasping for human participants.
• This approach offers a promising solution for enhancing the functionality and usability of prosthetic devices.