Optimal Control of Granular Material

Yuichiro Aoyama, Amin Haeri, Evangelos A. TheodorouICRA 2023

TL;DR

This paper presents a novel framework for the optimal control of granular materials (like sand or grain) by combining physics-based simulation with machine learning. By reducing the complex, high-dimensional motion of thousands of particles into a small set of "modes" using PCA and training a Graph Neural Network (GNN) to learn their dynamics, the authors successfully use Differential Dynamic Programming (DDP) to calculate the precise movements needed to shape the material into a target configuration.

What Problem Does This Paper Address?

Granular materials are a nightmare for traditional control theory. They behave like solids, liquids, or gases depending on the forces applied, and simulating them requires tracking thousands of individual particles, each with multiple degrees of freedom.

In industrial applications—such as autonomous excavation, mining, or planetary rovers—robots need to interact with soil or grain efficiently. However, high-fidelity physics simulators (like the Material Point Method) are computationally "expensive" and far too slow for real-time trajectory optimization. Conversely, simpler models often fail to capture the complex emergent behaviors of these materials. The authors aim to bridge this gap: creating a model that is fast enough for optimization but accurate enough to be useful in the real world.

Methods

The authors propose a multi-stage pipeline to move from "slow physics" to "fast control":

  1. Data Generation: They use a high-fidelity Material Point Method (MPM) simulator with a Nonlocal Granular Fluidity (NGF) model to generate "ground truth" data of a box shaking granular particles.
  2. Dimensionality Reduction (PCA): To avoid the memory bottleneck of tracking every grain, they use Principal Component Analysis. They found that the motion of 1,000 particles could be compressed into just 8 principal modes while retaining 95% of the system's "energy" (variance).
  3. Dynamics Learning (GNN): They train a Graph Neural Network to predict how these 8 PCA modes evolve over time based on the box's movement. Because GNNs operate on the relationships between nodes, they are naturally suited for physical systems.
  4. Optimal Control (DDP): With a fast, learned model in hand, they apply Differential Dynamic Programming. DDP iteratively refines a sequence of control inputs (the box's acceleration) to minimize a cost function—essentially "shaping" the virtual sand into a target slope.

Datasets

The researchers generated their own synthetic dataset for this study:

  • MPM-NGF Dataset: 324 simulation examples of a 2D rectangular box containing granular material.
  • Variations: The simulations varied box widths, motion frequencies (up to 3Hz), and amplitudes to ensure the GNN saw a wide range of physical behaviors.
  • Scale: Each example consisted of 90 frames (1.5 seconds at 60Hz).

Key Insights and Ideas

  • The Power of Subspace Learning: You don't need to simulate every grain of sand to control the pile. By performing GNN training in the PCA-reduced subspace, the authors bypassed the massive VRAM requirements that usually make GNNs infeasible for large-scale particle simulations.
  • DDP Efficiency: DDP is particularly effective here because the "control dimension" (the box's movement) is very small compared to the "state dimension" (the particles). This allows for fast convergence compared to other optimization methods like SQP.
  • Closing the Loop: The authors didn't just test their controller on the AI model; they applied the resulting commands back to the original physics-heavy MPM simulator to prove the "learned" controls actually work on "real" physics.

Key Results

  • Accuracy: The framework successfully generated control sequences to form both steep and mild slopes.
  • Quantitative Performance: In "Example 2" (a mild slope), the Root Mean Square Error (RMSE) between the reconstructed target and the final state was as low as $2.20 \times 10^{-3}$.
  • Speed: By using the reduced model, the system can perform the "rollouts" (simulating future states during optimization) significantly faster than a full-scale physics engine.
  • Validation: When the DDP commands were fed back into the high-fidelity MPM simulator, the final particle shapes closely matched the targets, though with slightly higher error than in the learned subspace.

Limitations and Future Work

  • Local Minima: Like many trajectory optimizers, DDP can get stuck in local minima. If the target shape is too complex or the initial state is too far away, the solver might not find the global optimum.
  • Underactuation: The system is highly underactuated—you are trying to control 1,000 particles by moving just one box. The authors noted that increasing the number of PCA modes (to 16 or more) actually made the optimization harder because the solver struggled with the increased complexity.
  • Reconstruction Artifacts: PCA is a linear method. It struggles with sharp edges, leading to "rounded" corners in the reconstructed shapes.
  • Future Directions: The team suggests moving toward non-linear reduction methods like Autoencoders or Kernel PCA and exploring more robust cost metrics (like Chamfer distance) to better handle particle distributions.

Summary updated

Abstract

The control of granular materials, showing up in many industrial applications, is a challenging open research problem. Granular material systems are complex-behavior (as they could have solid-, fluid-, and gas-like behaviors) and high-dimensional (as they could have many grains/particles with at least 3 DOF in 3D) systems. Recently, a machine learning-based Graph Neural Network (GNN) simulator has been proposed to learn the underlying dynamics. In this paper, we perform an optimal control of a rigid body-driven granular material system whose dynamics is learned by a GNN model trained by reduced data generated via a physics-based simulator and Principal Component Analysis (PCA). We use Differential Dynamic Programming (DDP) to obtain the optimal control commands that can form granular particles into a target shape. The model and results are shown to be relatively fast and accurate. The control commands are also applied to the ground-truth model,(i.e., physics-based simulator) to further validate the approach.
BibTeX
@inproceedings{aoyama2023optimal,
  author = {Yuichiro Aoyama and Amin Haeri and Evangelos A. Theodorou},
  title = {{Optimal Control of Granular Material}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
}
Figure