A Novel NDLV-xLSTM Model for Trajectory Prediction in Table Tennis

• Jun Wang,
• Xin Xiong,
• Xi Hu,

Vol. 19, No. 11, November 30, 2025
10.3837/tiis.2025.11.004, Download Paper (Free):

• LSTM
• Multi-view
• NDLV
• Trajectory prediction
• Transformer
• xLSTM

Abstract

The accurate prediction of table tennis trajectories plays a vital role in applications, including intelligent decision-making systems, robotic control, and human-machine interaction. However, due to the presence of latent dynamic factors that are difficult to observe directly, traditional models are difficult to achieve high robustness and generalization in real scenarios. To address the aforementioned problem, this paper proposes a novel three-dimensional trajectory prediction model that integrates a Nonlinear Dynamic Latent Variable (NDLV) mechanism with an extended LSTM (xLSTM) architecture. This model leverages three-dimensional trajectory data collected by a multi-view vision system and introduces a physically-driven NDLV mechanism to capture the hidden, time-evolving influence of latent dynamic factors on trajectory development, i.e., this model employs a hybrid backbone comprising a two-layer Scalar LSTM (sLSTM) and a Matrix LSTM (mLSTM) structurally, which is guided by dynamic control signals generated by the NDLV mechanism. It enables joint modeling of temporal dependencies and spatial correlations, which significantly enhances the ability to represent non-stationary trajectory patterns. To validate the effectiveness of the proposed model, our experiments are conducted on multi-step trajectory prediction tasks. The experimental results show that the proposed model outperforms traditional Transformer, CNN, LSTM, and xLSTM models in both accuracy and computational efficiency. It demonstrates the proposed model's strong adaptability and practical potential in modeling high-speed, complex target trajectories such as those in table tennis.

Statistics

Cite this article