Autonomous Feeding Robots: Inference of Eating Intentions via ResNet-LSTM with Gaussian and Variance Attention

The development of physically assistive robots for eating assistance has significant potential to improve the quality of life for seniors and individuals with disabilities. Effective autonomous feeding, however, depends on accurately inferring eating intentions. Many existing methods fail to adequately address human intention and the dynamic variability of behavior during eating. This paper presents a method for inferring eating intentions in human-robot interaction (HRI) by combining Residual Networks (ResNet) with Long Short-Term Memory (LSTM) networks, enabling robotic arms to feed autonomously. Inspired by observable facial movements during eating, we extract a key feature: the amplitude of the chin-to-nose distance from facial landmarks, which clearly indicates eating intentions. To mitigate noise and data corruption in input sequences, we use a Gaussian function as the convolutional kernel in the ResNet framework and integrate a variance attention mechanism in the LSTM's hidden layer to capture dynamic changes. Experimental results show our method achieves an accuracy rate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 7. 0 \%}$</tex> in intention inference. Real-world tests with a robotic arm and an RGB camera validate our approach's efficacy and real-time predictive performance.