Extracting Kinodynamic Constraints from Expert Driver Data for High-Speed, Mobile Robot Navigation in Off-Road Environments
Keywords: field robotics, motion planning, off-road, kinodynamic, constraint, learning
TL;DR: A method for generating kinodynamic constraint profiles from expert-driven trajectories
Abstract: High-speed, off-road driving demands rapid decision-making and adaptability to diverse environments, making it a domain where humans outperform autonomous robots. To enable robots to meet or exceed human driving performance in such environments, motion planners must be able to reason about the kinodynamic limits of their platforms in the context of the observed environment. In this paper, we present a statistics-based method for generating kinodynamic constraint lookup tables from human-expert demonstration. With this method, we aim to expedite in-field development of autonomous robot systems by eliminating lengthy model training and parameter tuning. We present a case where our method was used to investigate the impact of two types of kinodynamic constraints with both physics-based models and human-expert data on planned motions. To analyze the differences, we compare plans generated by a recombinant motion planning search space using the different constraint models. Data for these experiments originates from physical tests on a field robot built from a Polaris RZR: Side-by-Side platform outfitted with sensors and computing for high-speed, off-road navigation. Comparisons of solutions to 4,316 planning problems extracted from a separate set of logs indicate that an average of 28.8% of states in each solution generated by the physics-derived baseline exceeded the data-driven kinodynamic limits from the expert-driven model. Additionally, the velocity limits imposed by the expert-driven model were less conservative in some regions of the curvature-roll-pitch space, leading to 36.6% of solutions exhibiting higher average speeds than the physics-derived baseline.
Submission Number: 17
Loading