Abstract
This paper describes continuous-space methodologies to estimate the collisionprobability, Euclidean distance and gradient between an ellipsoidal robot modeland an environment surface modeled as a set of Gaussian distributions.Continuous-space collision probability estimation is critical foruncertainty-aware motion planning. Most collision detection and avoidanceapproaches assume the robot is modeled as a sphere, but ellipsoidalrepresentations provide tighter approximations and enable navigation incluttered and narrow spaces. State-of-the-art methods derive the Euclideandistance and gradient by processing raw point clouds, which is computationallyexpensive for large workspaces. Recent advances in Gaussian surface modeling(e.g. mixture models, splatting) enable compressed and high-fidelity surfacerepresentations. Few methods exist to estimate continuous-space occupancy fromsuch models. They require Gaussians to model free space and are unable toestimate the collision probability, Euclidean distance and gradient for anellipsoidal robot. The proposed methods bridge this gap by extending prior workin ellipsoid-to-ellipsoid Euclidean distance and collision probabilityestimation to Gaussian surface models. A geometric blending approach is alsoproposed to improve collision probability estimation. The approaches areevaluated with numerical 2D and 3D experiments using real-world point clouddata. Methods for efficient calculation of these quantities are demonstrated toexecute within a few microseconds per ellipsoid pair using a single-thread onlow-power CPUs of modern embedded computers