End-to-End Deep Learning by MCU Implementation: An Intelligent Gripper for Shape Identification.

Area of Science:

• Robotics and Automation
• Machine Learning
• Sensor Data Processing

Background:

• Object identification is crucial for intelligent robotic systems.
• Real-time processing of sensor data on embedded systems presents challenges.
• Vibration signals contain rich information for object recognition.

Purpose of the Study:

• To develop a real-time system for object shape identification using a microcontroller unit (MCU).
• To implement a convolutional neural network (CNN) for classifying grasped objects based on sensor data.
• To demonstrate the effectiveness of an intelligent gripper system for real-time applications.

Main Methods:

• Raw acceleration data from embedded accelerometers was converted into images using Short-Time Fourier Transform (STFT).
• A Convolutional Neural Network (CNN) algorithm was employed for feature extraction and object classification.
• CNN hyperparameters were optimized for efficient hardware implementation on a Renesas RX65N MCU.

Main Results:

• The system achieved real-time processing and classification of raw sensor data.
• The intelligent gripper successfully identified the shapes of grasped objects.
• The implemented CNN model demonstrated effective performance on the MCU.

Conclusions:

• The proposed approach enables real-time object shape identification using a CNN on an MCU.
• The intelligent gripper system offers a viable solution for embedded robotic applications.
• This work highlights the potential of sensor-based AI for practical robotic tasks.