Use of AI to quantify the "safety of autonomous vehicles"
summary
Autonomous vehicle safety requires demonstrating roughly 200–400 million miles between fatal crashes so physical testing alone is infeasible; the claim is that systems need 99.99999% reliability. Simulation combines cloud Monte Carlo sampling with a PCR-style amplification of rare hazardous scenarios (SMOTE-like oversampling proves inadequate) and reduces false negatives in experiments (MNIST class-skew tests and sunny→dusk image shifts show dangerous bias). A focused sampler generates many hazard-rich runs so validated controllers achieve controllability and exposure estimates without billions of real miles.
AI-written summary · about
chapters (5)
- 0:00Start
- 8:00Example City Driving Active Accident Avoidance
- 10:22Example City Driving Hazard Analytics
- 19:34Example Hazardous Freeway Driving Following Vehicle Making Maneuver
- 20:40How do we Model? Leverage Distributions
related talks
08mHuman–AI Collab for Reliable Automotive Control Systems
57mWriting Safety Critical Automotive Software for High Perf AI Hardware
18mAutomated Risk Evaluation and Mitigation for AI Systems
02mAI-defined vehicles and autonomy
32mIntelligent Intersections
41mNot Only Cars: “AI, Please Test My App”
52mPhysical AI: Crafting Resilient Systems with Emotional Intelligence
59mOn Human-robot Interactions via Multimodal Intent Modeling and Reachability-based Safety Assurance