Lecture
Self-Driving Cars: State of the Art (2019)
- The primary mission of autonomous vehicles (AV) is to improve mobility access for underserved populations, increase travel efficiency, and, most critically, save lives by preventing traffic fatalities.
- Approximately one person dies in an auto crash globally every 23 seconds, serving as the driving force behind the industry's safety goals.
- In 2018, Waymo achieved a milestone of 10 million miles driven autonomously, marking a significant step in full autonomous deployment.
- Tesla reached 1 billion miles driven using its semi-autonomous Autopilot system, relying primarily on computer vision and neural networks for decision-making.
- Two major fatalities occurred in 2018: a pedestrian death caused by an Uber vehicle in Tempe, Arizona, and a fatality involving a Tesla Autopilot vehicle in Mountain View, California.
- Critics argue that public attention disproportionately focuses on rare AV fatalities compared to the consistent 80–100 million miles driven per fatality in manual vehicles.
- While Tesla's 1 billion miles include 3 fatalities, this crude metric suggests a three-fold safety advantage over manual driving, though experts caution this comparison is flawed due to differing road types, driver demographics, and vehicle conditions.
- Public deployment of AVs in 2018 remained largely experimental, with services operating at small scales in constrained environments (e.g., Waymo in Phoenix, Voyage in Florida, Cruise in San Francisco) and almost always requiring a safety driver.
- Nuro launched zero-occupancy autonomous delivery services, demonstrating a path for goods transport that does not involve human passengers.
- A "scale" threshold of 10,000 vehicles is identified as the tipping point for moving from prototype demonstrations to meaningful societal impact; New York City currently employs approximately 46,000 Uber drivers.
- Industry predictions for full autonomy vary wildly, ranging from Elon Musk's 2019 target to Rodney Brooks's prediction of widespread bans on manual driving only after 2045.
- Experts warn that early AVs will likely be slower and less aggressive than human drivers, potentially reducing immediate efficiency gains and requiring a superior "human experience" to drive adoption.
- Level 2 autonomy retains human liability and responsibility, whereas Level 3 and above shift liability to the vehicle, requiring the system to guarantee safety without human intervention.
- True "full autonomy" (Level 4/5) is defined by the inability to rely on teleoperation or "safe harbor" fallbacks that require human action; the vehicle must handle all scenarios independently.
- Deployment strategies include last-mile delivery, highway trucking platooning, personalized public transport on fixed routes, and closed-community operations with constrained geofences.
- Future infrastructure concepts include connected vehicle communication to eliminate traffic lights and underground tunnel networks to transform cars into high-speed public transit.
- Historical DARPA challenges (2005 and 2007) were initially mistaken for solving the core AV problem, but the industry realized urban driving involves complex non-verbal communication and unpredictable human behavior that desert racing did not.
- MIT research instrumented 22 Teslas over two years, finding that drivers remained vigilant during 26,000 instances of control transfer, challenging the assumption that humans automatically lose vigilance in semi-autonomous systems.
- The industry is divided into two competing technical approaches: Vision-based systems (Tesla) using deep learning and cameras versus LiDAR-based systems (Waymo) using high-precision sensors and maps.
- Vision-based approaches benefit from massive data availability, low cost, and high resolution but struggle with explainability, edge cases, and performance in extreme weather without vast datasets.
- LiDAR-based approaches offer high accuracy, reliability, and explainability but are expensive, have lower data scalability, and do not easily integrate with deep learning pipelines.
- Sensor fusion combines Radar (cheap, speed detection, low resolution), Ultrasonic (proximity, low cost), Cameras (rich texture, high resolution, weather-sensitive), and LiDAR (precise depth, expensive) to create a robust perception suite.
- Tesla utilizes a camera-radar-ultrasonic stack, while Waymo employs LiDAR as a primary fail-safe, though both companies increasingly integrate multiple sensor types.
- The transition from "toy" AI benchmarks (like ImageNet) to real-world AV deployment represents the defining challenge for 21st-century artificial intelligence.
- Engineers must solve complex human-robot interaction problems, including natural language communication, driver attention monitoring, and game-theoretic negotiation with pedestrians.
- The speaker emphasizes that solving the engineering problems of autonomy is insufficient without addressing the psychological and sociological aspects of how humans interact with and trust these systems.