Residual Depthwise-Separable Temporal Convolutional Network for Robot Localization

Abstract

Robot indoor localization is essential for navigation and autonomous operation in GPS-denied environments. In complex indoor spaces, wireless signals are affected by multipath propagation, non-line-of-sight (NLOS) blockage, and spatial layout changes, making channel state information (CSI)-based localization challenging. This paper proposes a Residual Depthwise-Separable Temporal Convolutional Network (DS-ResTCN), a lightweight CSI-based model for robot indoor localization. The method converts raw complex CSI into a mixed amplitude–phase representation composed of amplitude, sine-phase, and cosine-phase features, which mitigates phase discontinuities and preserves complementary channel information. DS-ResTCN learns localization-related features from CSI sequences using a compact residual temporal model and predicts the two-dimensional robot position. Experiments on the WSR-Toolbox-Dataset show that DS-ResTCN achieves mean localization errors of 0.6493 m, 0.8116 m, and 0.4484 m in LOS, NLOS-Convex, and NLOS-Nonconvex scenarios, respectively. It outperforms MLP, CNN-1D, TCN, LSTM, GRU, and Vision Transformer (ViT) baselines. DS-ResTCN also uses only 0.2855M parameters and 69.4765M multiply-accumulate operations (MACs), showing that compact temporal modeling is effective for CSI-based robot indoor localization.