Self-Driving Vehicle Verification Towards a Benchmark Nima Roohi, Ramneet Kaur, James Weimer, Oleg Sokolsky, Insup Lee University of Pennsylvania July 13, 2018
Winter of AI for Autonomous Vehicles • July 13, 2018 2
Winter of AI for Autonomous Vehicles • July 13, 2018 3
Outline • System Specification – Finite set of cars – Only move forward – Driving on a straight road • Five Verification Challenges – (Robust) Safety – (Robust) Liveness – Responsibility Sensitive Safety • Current State of the Automatic Formal Verification Tools – Unable to specify any of these problems July 13, 2018 4
Position and Delay Signals • July 13, 2018 5
Minimum Longitudinal Distance Signal • July 13, 2018 6
Dangerous Longitudinal Situation Signal • July 13, 2018 7
Blame Time Signal • July 13, 2018 8
Policy • July 13, 2018 9
Outline • System Specification – Finite set of cars – Only move forward – Driving on a straight road • Verification Challenges – (Robust) Safety – (Robust) Liveness – Responsibility Sensitive Safety • Current State of the Automatic Formal Verification Tools – Unable to specify any of these problems July 13, 2018 10
Safety • July 13, 2018 11
Robustly Safe and Robustly Unsafe • July 13, 2018 12
(Robust) Liveness and Responsibility Sensitive Safety • (robust) Liveness: – Something good will eventually happen • RSS: – In case of an accident there will be a responsible party July 13, 2018 13
Outline • System Specification – Finite set of cars – Only move forward – Driving on a straight road • Verification Challenges – (Robust) Safety – (Robust) Liveness – Responsibility Sensitive Safety • Current State of the Automatic Formal Verification Tools – Unable to specify any of these problems July 13, 2018 14
Tools • We cannot even specify these challenges using these tools – marks when a tool fails – White cells do not guarantee a feature support July 13, 2018 15
A Few Differences • Delay is made explicit in dynamics and policy • Computation of blame time signal is slightly changed • Policy is (slightly) changed July 13, 2018 16
Recommend
More recommend
