Pet-NODE Modeling: Embedding Priors and Time-Series Features into Neural ODE

Abstract

Accurate modeling of dynamic systems is essential for robotics, enhancing system perception and control performance. This work tackles causal modeling challenges for mobile robots under complex uncertainties, including internal model inaccuracies and external environmental disturbances. Unlike first-principle or purely data-driven methods, we propose Pet-NODE, an advanced Neural Ordinary Differential Equation (NODE) framework that integrates physical priors with temporal features for high-fidelity system modeling. To further embed domain knowledge, we introduce a novel loss function with self-prediction objectives, ensuring adherence to physical principles. Extensive experiment evaluations, including ablation studies and comparisons against Nominal model, K-NODE and PI-TCN methods, demonstrate Pet-NODE’s robustness, interpretability, and superior localization accuracy on a self-collected wheeled robot dataset.