Multiscale Terrain Point Cloud Modeling for Off-Road Autonomous Ground Vehicle via Physics-Constrained Temporal Networks

Accurate and efficient prediction of the mobility of autonomous ground vehicle (AGV) in soil terrain environments remains challenging, primarily due to the inherent trade-off between physical fidelity and computational cost in existing approaches. To address this challenge, we propose a novel physics-informed neural network enhanced multiscale terrain modeling approach. This approach achieves soil macro-micro mechanical coupling through homogenization principles. The innovation of this study is the introduction of a physics constrained long short term memory architecture, which embeds soil constitutive equations as hard constraints and overcomes the physical inconsistency of purely data driven models through particle swarm optimization for adaptive loss weighting optimization. Experimental results demonstrate this approach achieves high precision predictions in critical mobility metrics. Compared to the traditional approaches, it enhances computational efficiency by a factor of 6.8 and reduces data dependency by over 40% compared to artificial neural network baselines. Additionally, cross-terrain validation confirms robust generalizability, providing a reliable foundation for subsequent navigation and control strategies of AGV operating in off-road environments.