[자율주행 스터디] Introduction to Self-Driving cars - Week 3-2

Week 3 - Safety Assurance for Autonomous Vehicles

Course info.

• https://www.coursera.org/learn/intro-self-driving-cars?
  • UNIVERSITY OF TORONTO

---

Lesson 2 : Industry Methods for Safety Assurance and Testing

Contents

• Industry perspectives on self driving safety
• Approaches to demonstrating autonomy safety

---

Waymo Safety Perspective

• Safety Level

  • Behavioral Safety
  • Functional Safety
  • Crash Safety
  • Operational Safety
  • Non-collision Safety

• Safety Processes

  • Identify hazard scenarios & potential mitigations
  • Use hazard assessment methods to define safety requirements
    • Preliminary analysis
    • Fault tree
    • Design Failure Modes & Effect Analysis
  • Conduct extensive testing to make sure safety requirements are met

• Levels of testing to ensure safety

  • Simulation testing
    • Test rigorously with simulation, thousands of variations, fuzzing of neighbouring vehicles
  • Closed-course testing
    • Follow 28 core + 19 additional scenario competencies on private test tracks
    • Focus on four most common crashes
      • Rear-end intersection, road depature, lane change
  • Real-world driving
    • Start with smaller fleet, expand steadily
    • Already testing thousands of vehicles, with more on the way

---

GM Safety Perspectives

• Safety Processes

  • Deductive Analysis
    • Fault tree analysis
  • Inductive Analysis
    • Design & Process FMEA (Failure Mode and Effects Analysis)
  • Exploratory Analysis
    • HAZOP : Hazard & Operability Study

• Safety Thresholds

  • All GM vehicles are equipped with two key safety thresholds
    • Fail safes
      • There is redundant functionality (second controllers, backup systems etc) such that even if primary systems fail, the vehicle can stop normally
    • SOTIF
      • All gritical functionailities are evaluated for unpredictable scenarios

• Testing

  • 'Performance testing' at different levels
  • 'Requirements validation' of components, levels
  • 'Fault injection testing' of safety critical funcionality
  • 'Intrusive testing' such as electromagnetic interference, etc
  • 'Durability testing' and 'simulation based testing'

---

Analytical vs Data Driven : Definitions

• Analytical Safety
  • Ensuring the system works in theory and meets safety requirements found by hazard assessment
• Data driven safety
  • Safety guarantee due to the fact that the system has performed autonomously without fail on the roads for a very large number of kms.

---

Are autonomous cars safer?

• Driving is still dangerous!
• Car accidents are amostly caused due to human error (NHTSA Report, 2015)
• In US, on average
  • 1 fatal collision per 146 million km
  • 1 injury collision per 2.1 million km
  • ~ 1 collision per 400,000 km
• Consider California disengagement rates
  • In 2017, Waymo had
    • Driven 563,000 km autonomously in California
    • 63 disengagements
    • 1 disengagement every 9,000 km
  • In 2017, GM had
    • Driven 210,000 km autonomously in California
    • 105 disengagements
    • 1 disengagement evert 2,000 km

---

The Dilemma

• Question
  • How many miles (years) would autonomous vehicles have to be driven to demonstrate with 95% confidence their failure rate within 20% of the true rate of 1 fatality per 140 million km?
• Answer
  • ~ 400 years, with a fleet of 100 vehicles travelling all the time (total ~8 billion miles)