Data-driven Parameter Learning without Groundtruth for Improving Robotic State Estimation

Abstract

Probabilistic state estimation for robotics is a field that has greatly matured in recent years. Yet the performance of modern estimators is still heavily dependent on robot model parameters that can be difficult to determine from first principles. While it may be sufficient to hand tune these parameters, this is often time consuming or impossible due to the sheer number of unknown parameters. In this thesis, we investigate methods for learning parameters based on data, to come up with the parameter values most suitable for the particular robot and sensor. We first develop a novel continuous-time motion prior trained with data to improve an existing continuous-time estimation framework. We then provide a detailed investigation into parameter learning within a Gaussian variational inference setting using Expectation Maximization. We validate our work on various trajectory estimation problems using a 36 km long vehicle dataset collected as part of this work.