Game Theoretic Analysis of Road User Safety Scenarios Involving Autonomous Vehicles Department of Information Engineering Umberto Michieli Leonardo Badia 11/09/2018
Ri Rise of Au Autonomo mous Vehicles (AVs) 1
Ri Rise of Au Autonomo mous Vehicles (AVs) 1
Tr Transition to AVs Smooth transition Need to overcome many conflicts 2
PR PRO 1. Accidents ↓ 2. Less stressful time 3. Less road congestion 4. Decreased emissions 5. Eventually faster than human-drivers 3
CO CONS 1. Social acceptance 4
CO CONS 1. Social acceptance 2. Technological issues • snow/rain • yellow-lights • partial occlusion MO MOORE’S ’S LAW 3. Interactions w/ humans • road regulations • AVs are cautious RI RISK SK AVERSE RSE 4
Our Our Appr pproach MODEL SIMULATION VALIDATION Expect Exp cted Con Concl clusion ons: • Game Theory extensions • Accidents ↓ as share of AVs ↑ • Need for new traffic regulations • Need for communication systems 5
Mo Modeling Human-AVs interactions Ga Game Th Theory Statistics (our approach) • Players have different utilities • Distinguishable set of actions • Statistical generality Pr Propose sed mo models: s: 1. Cyclist vs. Vehicle on Zebra Crossing 2. Pedestrian vs. Vehicle 6
1. Cy 1. Cyclist t vs. . Vehicle on Zebra Cr Crossing Nature AV probability p Cyclist à SIMULTANEOUS Yield Walk Cycle AV BAYESIAN GAME Go Stop Go Stop Go Stop COMMON KNOWLEDGE 5 3 -400 15 -500 20 7 10 -500 15 -300 15 & FULL RATIONALITY Human Driver probability (1-p) Cyclist AIMS AI MS: Yield Walk Cycle Human Driver Cyclist vs. AV or human driver Go Stop Go Stop Go Stop Accident rate curve as AVs ↑ 8 6 -400 15 -500 20 15 1 -400 7 -200 7 7
1. Cy 1. Cyclist t vs. . Vehicle on Zebra Cr Crossing Nature AV probability p Cyclist TWO TW O PU PURE NEs Yield Walk Cycle 1. (CY, SG) AV 2. (CC, SS) Go Stop Go Stop Go Stop 5 3 -400 15 -500 20 7 10 -500 15 -300 15 ONE MIXED NE ON Human Driver probability (1-p) If A If AV : (C,S) Cyclist If If h human : yield (p 1 ) or Yield Walk Cycle Human Driver cycle (1-p 1 ), go (p 2 ) or stop (1-p 2 ) Go Stop Go Stop Go Stop p1=93.7% p2=2.7% 8 6 -400 15 -500 20 15 1 -400 7 -200 7 8
1. Cy 1. Cyclist t vs. . Vehicle on Zebra Cr Crossing 0.30 low speed medium speed high speed 0.25 0.20 % of Collisions 0.15 0.10 0.05 10 20 30 40 50 60 70 80 90 % of Autonomous Vehicles 9
1. 1. Cy Cyclist v t vs. V . Vehicle o on Z Zebra Cr Crossing 0.12 low speed medium speed high speed 0.10 % of Fatal Injuries 0.08 0.06 0.04 0.02 10 20 30 40 50 60 70 80 90 % of Autonomous Vehicles 10
2. P . Ped edes estri trian v vs. V . Veh ehicle Pedestrian Out Cross PA PAYOF OFF IS TI TIME: ETA of Vehicle vehicle to make decision (t a -t c ’) Keep Brake 0 (t c -t a ) (t c ’-t a ) (t c -t a ) (t a -t c ) moves at 1.4 .4 m/s lane-width = 3.7 la .75 m 1/ 1/t c 1/t c ’ 1/ 1/t a 1/ t a < t c t c < t a < t c ’ t a > t c ’ NE NE shif ifts: CK CB O 11
2. P . Ped edes estri trian v vs. V . Veh ehicle 63,58% 70% 59,82% 60% 50% 40% 32,23% 30% 21,07% 19,12% 20% 4,19% 10% 0% CROSS-KEEP CROSS-BRAKE OUT AVs AV SIMULATION PARAMETERS: 𝑤 ∼ max( 𝒪 (50,10), 0) km/h } Human Driver AV human-dr hum drivers AVs } 𝑤 ∼ max( 𝒪 (30,10), 0) km/h AV a = 2.5 m/s d ∼ 𝒱 (10,50) m huma uman n dri drivers rs reaction time t r = 1.5 s à t c ’ higher than AVs 12
Co Conclusions à Fu Futu ture W e Work rk 13
Co Conclusions à Fu Futu ture W e Work rk § Game theory useful for human-AV interactions à improve realism 13
Co Conclusions à Fu Futu ture W e Work rk § Game theory useful for human-AV interactions à improve realism § Models are lightweight à embedding into communication systems and traffic simulators 13
Co Conclusions à Fu Futu ture W e Work rk § Game theory useful for human-AV interactions à improve realism § Models are lightweight à embedding into communication systems and traffic simulators § Accident rate ↓, dominance of pedestrians à new regulations needed, then new game analysis 13
Thank you for the attention! Questions?
Recommend
More recommend