TactiGraph: An Asynchronous Graph Neural Network for Contact Angle Prediction Using Neuromorphic Vision-Based Tactile...

Area of Science:

• Robotics
• Computer Vision
• Sensor Technology

Background:

• Vision-based tactile sensors (VBTSs) provide high-resolution tactile feedback for robots.
• Conventional VBTSs suffer from high computational overhead due to fixed camera update rates and redundant data.
• Existing solutions often compromise on instrumentation costs or maintenance.

Purpose of the Study:

• To develop a neuromorphic vision-based tactile sensor (N-VBTS) for efficient contact angle prediction.
• To reduce computational overhead and instrumentation costs associated with tactile sensing.
• To improve the speed and efficiency of tactile feedback systems in robotics.

Main Methods:

• Developed a novel neuromorphic vision-based tactile sensor (N-VBTS) utilizing event-based cameras.
• Designed a graph neural network, TactiGraph, to process raw N-VBTS data asynchronously.
• Exploited spatiotemporal correlations in sensor streams for accurate predictions.

Main Results:

• TactiGraph achieved a mean absolute error of 0.62° in contact angle prediction.
• The N-VBTS demonstrated significantly faster and more efficient performance compared to conventional VBTS.
• The system requires only 5.5% of the computing time of traditional VBTS.
• The N-VBTS functions efficiently with or without an internal illumination source, reducing costs.

Conclusions:

• The proposed N-VBTS and TactiGraph offer a cost-effective and efficient solution for robotic tactile sensing.
• This approach significantly reduces computational load and improves prediction accuracy.
• The technology holds promise for advancing robotic manipulation and interaction capabilities.