Human Hand Trajectory Prediction Using RNN Enhanced with KAN and Horizontal Vertical Exponential Gated Unit

Human motion prediction is a significant and challenging problem in the field of machine intelligence and a crucial component for achieving effective human-robot interaction. Robots need to accurately predict human future movements to safely plan their own trajectories, thereby cooperating effectively with humans. This paper proposes a framework for trajectory prediction based on an improved RNN neural network, combining the KAN network and Horizontal-Vertical Exponential Gating Units (HVEGU). This framework is specifically designed for human-robot interaction in medical dispensing scenarios, particularly when data representation capabilities are limited or hand motion trajectories are complex. We evaluate performance based on the average prediction error at each prediction step. Experimental results demonstrate that the KGRNN framework excels in predicting hand movements in medical dispensing scenarios. It maintains a low average error, around 7 cm. Compared to popular non-Transformer methods like LSTM, RNNIK-MKF, and WITRAN, KGRNN shows significant error reduction across various step lengths, with a maximum reduction of 26.67%. These findings indicate that our proposed KGRNN framework performs excellently in handling human-robot interaction tasks involving complex trajectories in medical dispensing scenarios.