Smart Tactile Sensing Systems Based on Embedded CNN Implementations.

Area of Science:

• Robotics and Machine Learning
• Artificial Intelligence
• Sensory Systems

Background:

• Integrating machine learning (ML) into tactile sensing enhances data interpretation.
• Intelligent tactile sensing systems are crucial for advanced robotics and prosthetic devices.
• Efficient hardware implementation of ML models is key for real-time tactile data decoding.

Purpose of the Study:

• To implement and compare a convolutional neural network (CNN) model for tactile data decoding on diverse hardware.
• To evaluate the performance of embedded ML models for tactile sensing applications.
• To demonstrate the feasibility of real-time intelligent tactile data processing.

Main Methods:

• Development and deployment of a convolutional neural network (CNN) model for tactile data classification.
• Implementation of the CNN model on various hardware platforms for performance benchmarking.
• Experimental validation of classification accuracy and inference time.

Main Results:

• Model 3 achieved a classification accuracy of 90.88%, comparable to state-of-the-art methods.
• The proposed embedded implementation demonstrated a rapid inference time of 1.2 ms.
• The system consumed approximately 900 μJ, indicating energy efficiency.

Conclusions:

• Embedded implementation of intelligent tactile data decoding algorithms is feasible.
• The developed CNN model offers high accuracy and efficient inference for tactile sensing.
• This technology has significant potential for applications in robotics and prosthetic devices.