Path Optimization for Construction Machines Based on Real-World Soil Dynamics

The automation of construction machinery, such as bulldozers, has gained significant attention in recent years. Effective automation requires accurate modeling of soil behavior and plan path planning. Conventional approaches rely on physics-based simulators, which demand significant expertise in soil mechanics and substantial computational resources for their design and operation. To overcome these limitations, we propose a data-driven framework that models soil dynamics directly from real-world observations, eliminating the need for complex simulator design. We developed a robotic testbed to replicate soil spreading work, collected soil behavior data, and introduced an efficient learning method based on localized prediction and data augmentation. Using this approach, we trained a neural network capable of accurately predicting post-push soil profiles. Our data-driven dynamics model is approximately 15.5 times faster than conventional simulators. Additionally, we trained a path planning network within the learned environment and validated its performance on the physical testbed, achieving a high fill rate of 90.8%. This work demonstrates a practical and efficient alternative to traditional simulation, highlighting the potential of real-world modeling for construction automation. A preliminary version of this paper was presented at SAMCON 2025(1), where it received the Outstanding Paper Award.