Learning off-road driving from simulation

点击翻译按钮获取中文摘要

Driving vehicles off-road in Canadian terrain is a skill of many of our forestry, construction and mining workforce, but it also poses great risk to human and equipment safety due to sliding, skidding and roll-over. CM Labs produces world-leading training and simulation products that allow operators of heavy equipment to face such situations in the safety of a game-like software experience, without personal risk, however, many of the most challenging environments and rapid motions cannot yet be captured by our state-of-the-art models. In collaboration with McGill's Mobile Robotics Lab, this project will leverage recent advances in Reinforcement Learning to produce new off-road driving policies and further improve CM Labs' tire-terrain dynamics models. A key aspect of this work is the ability of a intelligent software to explore driving possibilities only in simulation, for which we will rely on integration of CM Lab's Vortex Studio with the most recent advances in AI. Findings in this approximate domain will then be transferred to small and agile outdoor vehicles built recently at McGill, to validate the models and find differences between our simulators and reality. The results will improve outdoor automation in Canadian spaces, and the ability of CM Labs to train our workforce to safely conduct operations even in Canada's most challenging environments.