Towards soil-aware autonomous terrain shaping for bulldozing

This dissertation advances the goal of autonomous terrain shaping in unstructured environments with unknown and spatially varying soil conditions. Terrain shaping refers to the use of heavy machinery to modify the landscape to meet desired specifications, an essential capability in construction, mining, and military engineering. While recent progress has been made in automating earthmoving equipment, most existing systems are confined to structured settings with known terrain characteristics and rely on strong assumptions about soil uniformity. These limitations render them unsuitable for the uncertain and austere environments in which Army Combat Engineers often operate. Motivated by the need for robust autonomy under these conditions, this research focuses on bulldozing automation and addressing the critical challenge of operating effectively under variable, unknown soil conditions. A central contribution is the development of a physics-infused neural network (PINN) that estimates soil strength properties in situ from vehicle observations during earthmoving. This estimation method is validated in simulation and integrated into a terrain mapping system that fuses exteroceptive and proprioceptive sensing in a Bayesian framework to track terrain elevation and soil strength over space and time. These tools enable the system to maintain a spatial memory of soil conditions, supporting soil-aware control and planning. To explore sample-efficient learning for earthmoving, this dissertation also investigates model learning and predictive control for automated berm removal. Two modeling approaches are used to capture the coupled evolution of terrain and vehicle states using only a few minutes of demonstration data. The learned models are then leveraged in a model predictive control framework that optimizes over a parameterized action space. Though evaluated only in simulation, the system achieves near-human performance in reducing berm height over multiple passes, highlighting the potential of data-efficient learning for autonomous terrain shaping. To bridge the gap between simulation and real-world deployment, a compact track loader (REO-RCTL) was converted into a robotic platform, and the terrain mapping system was adapted to run online within a ROS 2 framework. A terrain smoothing controller was developed that tracks a locally fitted terrain plane, enabling interleaved cutting and depositing without a predefined surface profile. Despite limitations from actuator dynamics and localization drift, the system performed effectively in field tests. Experienced heavy equipment operators completed earthmoving tasks via beyond visual line of sight teleoperation, validating the practicality of the mapping and control methods and marking a key step toward autonomous terrain shaping in unstructured environments. Collectively, this work introduces a suite of estimation, mapping, learning, and control techniques that extend the capabilities of autonomous earthmoving systems in unstructured environments. By addressing key limitations of prior methods and emphasizing physically grounded, sample-efficient learning, this research lays the foundation for future systems capable of robust terrain shaping with minimal human supervision.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Wagner, William Jacob

Contributors dc:contributor

• Driggs-Campbell, Katie
• Soylemezoglu, Ahmet
• Hauser, Kris
• Kim, Joohyung

• Autonomous Earthmoving
• Robotics
• Machine Learning
• Construction
• Bulldozer
• Physics Infused Neural Networks
• Soil Property Estimation
• Terrain Mapping

Statement dc:rights

• Copyright 2025 W. Jacob Wagner

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129891

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129891