Technical and Legal Challenges for Urban Autonomous Driving Seung-Woo Seo, Prof. Vehicle Intelligence Lab. Seoul National University sseo@snu.ac.kr
I. Main Challenges for Urban Autonomous Driving I. Dilemma in Autonomous Driving II. Approach to Human like Driving I. Intention Aware Decision Making II. Imitation Learning III. Autonomous Driving Research in SNU I. Demonstration of SNUver IV. Conclusion 2
Challenges for Urban Autonomous Driving
Considerations for Urban Autonomous Driving Moving & static objects Pedestrians Other vehicles Traffic light & signs Unforeseen events Crossing intersection Turning Lane changes Parking Entering and exiting drop off stations Etc.
First Self-driving in City Road in Korea(2017. 6. 22)
Yeouido Area in Seoul
Demonstration at Yeouido Area in Seoul Driving course on Yeuido 2 3 1 7 4 5 6 7
Dilemma in Autonomous Driving In urban environments, dilemma situations frequently occur Crossing a double-yellow line to pass by an illegally parked car Lane-change in heavy traffic Decisions at a yellow traffic light 8
Dilemma in Autonomous Driving 3 Different Aspects I. Legal aspect II. Interactivity aspect III.Technology aspect 9
Legal Aspect Crossing a double-yellow line to pass by an illegally parked car Crossing a double-yellow line illegal & socially compliant decision VS. Waiting until an illegally parked car leaves legal & impractical decision
AV violating the traffic law
Interactivity Aspect Interactive driving (ex. Lane cut in) 12
Dilemma in Autonomous Driving 3 Aspects I. Legal aspect EX) Crossing a double yellow line to pass an illegally parked car II. Interactivity aspect EX) Lane change in heavy traffic unsignalized intersection III.Technology aspect Human-Like Driving 13
Approach to Human-Like Driving
TASK 1. LANE CHANGE IN HEAVY TRAFFIC TASK 2. INTERSECTION TASK N. HIGHWAY Policy Optimization Policy Optimization Policy Optimization Single Task Policy 1 Single Task Policy 2 Single Task Policy N 15
Model for Decision Making X t 1 A X t A The state space S is a joint space : Ego-vehicle s state space,, : Other vehicles state space,, : Other vehicles driving intention Θ, Ot 1 R O t R The action space A : A =.,.,. Yt 1 t 1 Y t t The reward model Very high penalty when vehicle is predicted to collide. Very high reward when vehicle arrives at its goal. Low penalty when vehicle moves at each step Passing through intersection as fast as possible without any collision 17
Experimental Environment SNU Campus road Total length : ~4km 기숙사삼거리 국제대학 원 Start 행정대학 원 대운동장 자동 화 시스 템 연구 소 Goal 18 18
Imitation Learning Learning from Expert Drivers Expert drivers understand human interactions on the road and comply with mutually accepted rules, which are learned from countless experience Behavior Cloning Inverse Reinforcement Learning, Learning, Learning, Technique Technique Policy Derivation Mapping from states to actions (Supervised Learning) Reconstruct reward function Brenna D. Argall, at el. A survey of robot learning from demonstration, Robotics and Autonomous Systems 57 (2009): 469 483 19
Imitation Learning Driving dilemma in single lane road Crossing a double-yellow line to pass by an illegally parked car Demonstration of expert drivers Sang Hyun Lee and Seung Woo Seo, A Learning Based Framework for Handling Dilemmas in Urban Automated Driving, IEEE International Conference on Robotics and Automation(ICRA), 2017 20
Imitation Learning Experimental Environments SNU Campus road Total length : ~4km 21
Autonomous Driving Research in SNU
SNUver SNU Automated Drive [November 19, 2013] Grand Prize in unmanned self driving car contest [June 22, 2017] Automated Driving in Urban Environments [November 4, 2015] Driverless taxi on SNU Campus [November 15, 2016] Door to Door Automated Driving on SNU Campus 23
SNUver 1 (2015)
SNUver 2 (2016)
SNUvi (2017)
Discussed several key issues related to dilemma in urban autonomous driving Briefly introduced our learning-based approaches to human-like driving There still remain many challenges that make the urban autonomous driving very hard Future Work 27