Post by Nadim Hossain

We know that neural networks are great at powering Perception for self-driving cars -- helping the robot "see" its environment -- but what if we trained deep learning models for the actual task of driving, instead of just perception? Could an integrated AI system replace connected but separate layers of an autonomy stack? If so, what does this mean for the vehicle's performance, and for the interpretability of the system for safety validation? This is the line of thinking discussed by Uber ATG's Chief Scientist Raquel Urtasun in this fascinating talk from #cvpr2020, "Interpretable Neural Motion Planning". Some promising takeaways for autonomous vehicle developers: - End-to-end planner outperformed a more traditional Motion Planning approach on quantitative measures - Promises to be offer faster reaction times due to shared computation, as well as improved developer productivity - Concerns of interpretability for safety validation and incorporating rules (eg traffic laws) are addressed Congrats to Raquel and team on the recently published papers! The video is a great jumping off point: https://lnkd.in/gqnwj68 #ai #deeplearning #ml #autonomousvehicles #selfdrivingcars #uber #uberatg #productmanagement