Dual-Path Inertial Odometry with Temporal Attention

ICLR 2026 Conference Submission24954 Authors

Keywords: Dual-path IMU odometry; Temporal attention fusion; Cross-backbone improvement

TL;DR: Dual-path IMU odometry fusing raw and SG-filtered signals via temporal attention cuts RONIN error by 10%, improving robustness to devices, sampling rates and backbones (ResNet, TCN, LSTM) with minimal overhead.

Abstract: We present a dual-path inertial odometry framework that processes the IMU stream through two parallel branches. One branch works directly on raw measurements to preserve high-frequency transients, while the other applies a Savitzky–Golay filter to enforce smoother, Newton-consistent motion and reduce drift. The outputs are fused online by a compact temporal-attention mechanism that adjusts their relative weights according to the motion dynamics. On the RONIN dataset, our method reduces final position error by about 10% compared with the previous state of the art, and this advantage persists across four smartphone models and three sampling rates. Integrating the dual-path block into other backbones yields similar gains — for example, roughly a 10% error reduction for a ResNet-based odometry network — and produces consistent improvements for both TCN and LSTM baselines, suggesting the approach generalizes across architectures.

Primary Area: learning on time series and dynamical systems

Submission Number: 24954