Particle Method for Generation of Hybrid Terramechanics Model Training Data

  • Karpman, Eric (McGill University)
  • Kovecses, Jozsef (McGill University)
  • Teichmann, Marek (CM Labs Simulations)

Please login to view abstract download link

There are two main approaches for modelling wheel-soil interaction in simulations. First, there are semi-empirical approaches that use equations based on physical principles as well as equations and coefficients determined based on fitting of experimental data. These approaches are computationally efficient and therefore are popular in the context of real-time simulations, but they can also be difficult to use for certain types of simulations due to assumptions that are baked into the formulation. There are also computationally expensive methods such as the Finite Element Method (FEM) and Discrete Element Method (DEM) which are more flexible in terms of their range of applicability but come with the drawbacks of many parameters that can be difficult to tune and the aforementioned computational cost. In our work, we aim to create a real-time wheel-soil model that utilises both the well established semi-empirical as well as a machine learning algorithm trained using DEM simulation results to compute the interaction forces. Our goal is to build a terramechanics model that is efficient enough to run in real time but that has the wide range of applicability that a DEM model has. The model works by training a neural network to augment the results of the semi-empirical model by learning the difference between the interaction forces that the model predicts and the ground truth forces - which for our purposes is taken the be the forces recorded in a position-based particle DEM simulation. This is done by first performing tests on the DEM particles to establish an equivalency between the DEM parameters and the coefficients that the semi-empirical model uses to define the soil type. Then, by running DEM simulations of various dynamic maneuvers that are difficult for the semi-empirical model to accurately capture and recording the state of the wheel, the DEM forces and the forces that the semi-empirical model would predict at each timestep, we generate training data for the machine learning algorithm. To do this we use the wheel's state variables as inputs and the difference between the DEM and the interaction force prediction of the semi-empirical model. Thus, the neural network learns to add/subtract from the semi-empirical model predictions in order to bring its outputs in line with what we expect based on DEM simulation results. We will present results for the hybrid model results for the wheel-soil interaction forces in the context of a single wheel testbed.