Data-Driven Bulldozer Blade Control for Autonomous Terrain Leveling

Abstract

Autonomous leveling of granular materials is a ubiquitous yet challenging operation in automated construction due to the complex physics governing the soil-tool interaction. This paper outlines a simulation-driven framework for optimizing low-level bulldozer blade control (pitch and height) to enhance leveling performance. The approach uses high-fidelity, physics-based simulations to generate training data. This data informs a Neural Network (NN) based reduced-order model that accurately predicts both the terrain evolution and the leveling operation duration in response to blade actions. A gradient-based, multi-objective optimization algorithm then utilizes the reduced-order model to determine optimal, time-varying blade control profiles, managing the trade-off between leveling flatness and operation time. The proposed method augments the state-of-the-art by producing policies that can readily level arbitrary soil pile configurations while avoiding vehicle immobilization and achieving better leveling efficiency. The system exhibits robustness to variations in initial pile geometry, and offers explicit control over the trade-off between leveling quality and operational efficiency. By integrating high-fidelity physics into the controller design and providing an open-source simulation pipeline, this work provides a low-level control solution that complements existing global path planning algorithms for autonomous construction operations. The project's resources, including code and media demonstration, are available at: https://anonymous.4open.science/r/Autonomous-Leveling/

The Challenge of Autonomous Terrain Leveling

While existing methods have demonstrated effectiveness in bulldozer path planning, they often overlook the nuanced interactions between the bulldozer and terrain. The consequences of this oversight are twofold:

1. It's difficult to predict the final terrain heightmap profile upon completion of the operation.
2. The vehicle may become immobilized by dense soil piles, leading to operational failure (as illustrated in Fig. 1b).

Our Simulation-Driven Approach

This paper introduces a simulation-driven framework for optimizing micro-level bulldozer blade control. Our methodology integrates high-fidelity physics into the control design process:

1. High-Fidelity Data Generation: We use the Chrono simulation engine, specifically Chrono::FSI and Chrono::Vehicle, to run physics-based terramechanics simulations. These simulations capture complex soil-tool interactions, vehicle dynamics, tire slip, and sinkage. This process generates comprehensive training data detailing how terrain evolves under different blade control actions (pitch and height). Figure 2 illustrates the data collection pipeline.
2. Differentiable Reduced-Order Model: The extensive data from simulations is used to train a Neural Network (NN) based reduced-order model. This NN, detailed in Figure 3, learns to predict the post-operation terrain heightmap and the operation duration given an initial heightmap and blade control inputs. The differentiability of this NN is key for optimization.
3. Gradient-Based Optimization: Leveraging the differentiable NN model, we formulate a multi-objective optimization problem. This problem aims to find optimal, time-varying blade control profiles that maximize leveling flatness (matching a desired flat terrain) while minimizing operational time. The optimization is solved using the Adam algorithm.

This approach allows for the determination of optimal blade actions tailored to specific initial terrain conditions, leading to improved leveling performance and efficiency.

Fig 2. Overview of the data collection process involving a simulated two-pass terrain leveling operation. Each simulation yields multiple training samples, linking initial heightmaps and blade control parameters to resulting heightmaps and operation durations.

Fig 3. Neural Network Architecture. The initial terrain heightmap is processed by a CNN encoder, and blade control inputs by an action-processing module. These are combined and fed through fully connected layers to predict the post-operation heightmap (via a CNN decoder) and operation time.

Experimental Results

We evaluated our autonomous leveling scheme through a series of experiments. The bulldozer performs two forward leveling passes, with the control actions for each pass optimized based on the current terrain heightmap. We compare our algorithm against fixed-blade strategies and assess its robustness and the trade-off between leveling quality and operational time. Leveling performance is quantified using the percentage change in Mean Squared Error ($\Delta$MSE) relative to a desired flat terrain.

Our experiments include:

1. Demonstration of various scenarios for same initial pile height
2. Performance comparison on varying pile heights
3. Analysis of the quality-time trade-off by varying optimization parameters

Fig 5. Algorithm performance demonstration with an initial soil pile height of 0.37m. Top row: Fixed blade parameters lead to vehicle immobilization. Second row: Another fixed blade setting results in insufficient leveling. Third row: First pass with our optimized control ($\Delta$MSE=26.20%). Bottom row: Second pass with optimized control achieves superior leveling ($\Delta$MSE=48.13%).

Fig 7. Algorithm performance across different initial pile heights. The proposed algorithm consistently achieves effective leveling ($\Delta$MSE scores) after two passes for various initial pile configurations.

Fig 8. Illustration of the trade-off between leveling quality ($\Delta$MSE) and operation time by varying the penalty parameter $\lambda$ in the optimization. A smaller $\lambda$ (e.g., $1 \times 10^{-6}$, right) prioritizes flatness, resulting in higher $\Delta$MSE (48.13%) but longer duration (12.08s). A larger $\lambda$ (e.g., $5 \times 10^{-4}$, left) prioritizes speed, reducing duration (10.82s) but also $\Delta$MSE (26.61%).

Resources

Acknowledgements