On the role of AI in autonomous driving: prospects and challenges

Transcription

1 On the role of AI in autonomous driving: prospects and challenges April 20, 2018 PhD Outreach Scientist

2 1.3 million deaths annually Road injury is among the major causes of death 90% of accidents are caused by human failure $871 Billion economic damage in the US annually General AImotive 2

3 SAE levels of autonomy SAE Name Description Actuation Monitoring Fallback performance System capability 0 No Automation Warning signals n/a 1 Driver Assistance Active driver assistance (steering or acceleration) Some driving modes 2 Partial Automation 3 Conditional Automation 4 High Automation 5 Full Automation Active complex driver assistance (steering and acceleration) Limited perception and decision making during the dynamic driving task, active human supervision Limited perception and decision making during the dynamic driving task, passive human supervision Can handle all aspects of dynamic driving task under all conditions Some driving modes Some driving modes Some driving modes All driving modes

4 Self-driving software: a high-level architecture Raw Sensor Data Low-level Fusion Recognition Engine Image Space Object Detection Environment Model Data Sensors Camera Radar Ultrasonic LIDAR Cross- Calibration Object Classification Environment Abstraction Detection Output High-level Fusion Instance Association Motion Engine Model Space Object Tracking Behavior Prediction Decision of Action Trajectory Control Engine Message Generation Actuation DBW safety GPS CAN IMU Location Engine Image/Map Space Trajectory Planning Raw Sensor Data Mapping Localization Location + Egomotion Egomotion 4

5 Leveraging AI: Recognition Classical approach CAR NOT CAR Heuristic features Binary classification Support Vector Machines Not scaling well Zheng, Y. and Blasch, E., 2016, May. Multispectral image fusion for vehicle identification and threat analysis AI-driven approach Artificial Neural Networks Learned semantic features Multi-class classification Integrated networks 5

6 Leveraging AI: Localization Classical approach Mur-Artal, R., Montiel, J.M.M. and Tardos, J.D., ORB- SLAM: a versatile and accurate monocular SLAM system. Dense point clouds Low-level features High mapping efforts Non-robust and non-deterministic Source: Velodyne AI-driven approach MAP DATA LOCALIZATION & MOTION SENSOR DATA Semantic feature-based Less data, more intelligence 6

7 Leveraging AI: Planning & Control Classical approach FINITE STATE REPRESENTATION FINITE TABLES CLASSICAL CONTROL Formal logic Hard-coded rules Problem: bending the rules Deterministic transition AI-driven approach Reward-based Reinforcement learning Black box solutions Data <-> scaling 7

8 Simulation in training and testing Real world testing systems are much more expensive Not all scenarios are safe to test / can be recreated In simulation: Variables not accessible in real world testing can be observed Models can be modified and parameters can be changed The time scale of the system may be extended or shortened Multiple instances can be scaled to run simultaneously

9 The world is not black & white MODULARITY AI-AIDED DEVELOPMENT VIRTUAL TESTING FIELD TESTING 9

10 Technology meets safety. And regulations Vienna Convention on Road Traffic 1968 ISO/WD PAS Road vehicles -- Safety of the intended functionality 10

11 Takeaway: the 3 pillars of self-driving technology Algorithms Development tools Integration in an independent self-driving software for sensor handling, perception, localization and decision making. Customized software tools for data handling, training, testing, verification and maintenance. Hardware Low-level optimization of algorithms, rethinking chip architectures for boosting common operations for high performance and low power. General AImotive 11

12 About us COMPANY Founded in 2015 by CEO Laszlo Kishonti Incorporated in Germany, headquartered in Budapest GLOBAL GROWTH 4 offices across 3 continents: Budapest, Helsinki, Mountain View and Tokyo 4 licensed testing locations: Hungary, Finland, California, Nevada Self-driving test car from 4 manufacturers: Toyota, Volvo, Citroen, SAIC TEAM & EXPERTISE 180+ passionate team, including 140+ highly skilled engineers In-house developed AI algorithms and framework 10+ years in high-performance embedded programming STRONG TRACTION Strong track record of delivering customer projects since 2015 Running up to 7 projects simultaneously, across 3 continents (Europe, USA, Asia) INVESTORS Robert Bosch VC, Nvidia, Inventure VC, Draper Associates, B Capital, Prime Ventures, Cisco Investments, Samsung Catalyst Fund, Day One Capital, Private investors 12

13 We Have Strong Traction in Automotive Industry Further Partners 13

14 Contact Budapest Office Szepvolgyi ut Budapest, Hungary 1907 Colony St, Mountain View CA, USA Helsinki Office Lapinlahdenkatu 18, Building 15C Helsinki, Finland