Physical Motion Classification in Assistive Industrial Robotics Using Support Vector Data Description
Abstract
Today's industrial assistive robots improve worker safety, productivity, and cooperation. Robots can't understand, predict, or respond to human behavior without good body movement classification. Traditional classification methods struggle with real-time adaptability, uneven data, and noise. This paper addresses faulty classification approaches that cannot handle limited-sample or one-class learning in assistive robots. Standard machine learning methods fail in industrial settings due to the lack of large, balanced datasets. This restriction makes it challenging to design intelligent robotic systems that assist humans in adapting to changing situations. Physical Motion classification using Support Vector Data Description (PM-SVDD) is our novel solution to these issues. PM-SVDD models usual motion patterns and identifies outliers that may indicate undesirable or aberrant motions using SVDD's one-class classification feature. Kernel-based mapping and temporal feature extraction boost discriminative performance in many industrial situations. A custom assistive robotics dataset and conventional motion datasets were employed for experimental validation. With an average classification accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 4. 7 \%}$</tex>, PM-SVDD outperforms SVM and k-NN in scenarios with limited negative data. Due to its low computational delay, the approach is ideal for real-time robotics. Finally, the PM-SVDD technique for physical motion classification in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{A I}$</tex> is strong and effective, enabling more intelligent and safer human-robot interaction.