Learning to Drive with a Touch of Human Knowledge

Abstract: Self-driving vehicles will bring us safer, cleaner, and more convenient transportation. To make this dream come true, we need our autonomous system to perceive, plan, and execute effectively in unstructured environments and have guaranteed safety. While machine learning has significantly enhanced autonomous capabilities, we are still missing key ingredients to achieve the desired goal. In this talk, I will present our approach towards solving autonomous driving. The core idea is to systematically integrate learning methods with structured models and human priors of the world. The effectiveness of our integrated approach has been demonstrated at the full spectrum of self-driving tasks, including localization, perception, planning and simulation, and our developed algorithms have been deployed in real-world production systems. Finally, I will give a brief personal outlook on open research topics towards realistically solving self-driving.

Bio: Shenlong Wang is a PhD student at the University of Toronto under the supervision of Raquel Urtasun. He is also a Senior Research Scientist at the Uber Advanced Technology Group. Shenlong's research interests span the spectrum from computer vision, robotics and machine learning. His recent work involves developing robust algorithms for self-driving and making autonomous vehicles more reliable and scalable. His research has resulted in over 30 papers at top conferences including over 10 oral and spotlight presentations. He was selected as the recipient of the Facebook, Adobe and RBC Fellowships in 2017.