DeepSafe: High-definition sensor-real edge-case training frameworks for the next generation of robustly trained AVs

点击翻译按钮获取中文摘要

Autonomous Vehicle (AV) pilot studies have demonstrated that AVs can handle "normal" driving; 99% of our day-to-day driving experience. However it is proving harder than expected to train AVs to deal with the multitudinous unusual events that can happen on the road, known as edge-cases. When an AV fails to understand an edge-case, driving behaviour becomes unreliable and unsafe, with examples include over-braking causing rear-end collisions, lane-keeping failures, and in the worst cases fatal high-speed collisions.It is essential for AVs to be able to handle edge-cases safely and reliably for public and consumer acceptance of AVs on the road, for the technology developers and automotive manufacturers to meet upcoming regulatory standards and so that the automotive industry can realise a return on the considerable investments made to date in AV technology.Training AV perception systems for edge-cases is challenging; the volume of real-life training data is limited. Simulation and synthetic data are widely recognised as being needed but currently available synthetic training data is not sufficiently sensor-real for simulation-trained AI to be successful in improving perception and decision-making on edge cases.DeepSafe brings together leaders in the simulation supply chain to resolve synthetic data issues inhibiting successful training for AV perception using simulation.