Differentiable Wheel Terrain Interaction Model and Pose Predictions on Uneven Terrain

Arun Kumar Singh, Amith Manoharan, Aditya Sharma, KAUSTAB PAL, Himani Belsare, Madhava Krishna

Keywords: Wheel-terrain interaction, uneven terrains, implicit differentiation

TL;DR: A differentiable wheel-terrain interaction and 6dof pose prediction on uneven terrain and its use in gradient-based planning

Abstract:

Off-road navigation of autonomous vehicles requires the use of $6dof$ models for planning safe paths that satisfy stability constraints. The existing approaches pose prediction approaches mostly rely on training neural networks from vehicle motion data. However, these approaches are data intensive, struggle to generalize to novel scenes and require navigating the vehicle over potentially dangerous terrains for data collection.

In this paper, we present a model-based approach that only requires an elevation map of the terrain (such as pointcloud obtained from a LiDAR). We formulate the wheel-terrain interaction and the resulting $6dof$ pose prediction of the vehicle as a non-linear least squares (NLS) problem. Importantly, we can leverage implicit differentiation rules to compute the gradient of the predicted pose with respect to the input parameters. We also briefly discuss how such differentiable models can be leveraged for gradient-based planning over uneven terrains.

Submission Number: 2