Lecture
Self-Driving Cars: State of the Art (2019)
- The autonomous vehicle landscape is projected to evolve with multiple industry leaders and diverse perspectives by 2019, while Waymo achieved 10 million autonomous miles in October 2018 and Tesla recorded 1 billion miles via its semi-autonomous Autopilot system.
- Tesla's 1 billion miles are described as the largest neural network deployment impacting human life, with a calculated safety ratio of one fatality per billion miles, though the speaker notes this comparison to manual driving statistics (one fatality per 80 to 100 million miles) is flawed due to differences in vehicle types, road conditions, and driver demographics.
- Near-term adoption (next 10 to 15 years) is expected to be driven by the creation of a "fun" human experience rather than just safety or speed, with fully autonomous vehicles initially operating slower than humans due to cautious algorithms until a majority of the fleet is autonomous.
- Deployment timelines vary significantly across the industry, ranging from Elon Musk's 2019 prediction to Rodney Brooks's view that bans on manual driving may not occur until the 2030s in significant regions and 2045 or beyond in the majority of US cities, with major automakers targeting 2020 to 2021 for specific milestones.
- The path to full autonomy (Level 4/5) involves specific proliferation scenarios such as last-mile delivery, zero-occupancy taxi services, and truck platooning, alongside the expectation that shared autonomy will utilize human-in-the-loop teleoperation for highway driving.
- Technical challenges persist regarding subtle non-verbal communication, "game-theoretic" interactions, and the necessity for systems to solve perception, control, and planning problems perfectly without relying on human vigilance, which a pilot study of 22 Teslas over two years suggested remained high during 26,000 control transfer moments.
- Sensor strategy debates center on the trade-offs between cameras (cheap, high resolution, data-intensive), LiDAR (consistent, expensive), and radar (weather-tolerant, low resolution), with sensor fusion aiming to match LiDAR capabilities while determining the ultimate fail-safe mechanism in semi-autonomous versus fully autonomous modes.
- Industry predictions suggest the driving task is difficult to solve due to the complexity of visual perception and physical control, where algorithms risk unintended consequences if formalized to maximize reward functions without human intervention.
- Public sentiment indicates 57% of respondents believe Tesla will first deploy 10,000 fully autonomous cars without a safety driver, while 8% predict no entity will achieve this milestone without a safety driver in the next 50 years.
- Future possibilities include connected vehicle communication to optimize traffic flow without lights, autonomous vehicles functioning as mini-trains in tunnels, and flying cars becoming accessible to non-pilots by 2050 according to some projections.
- The overarching goal of deep learning research is to save lives and prevent injuries or fatalities, with the speaker anticipating that AI will transition from toy benchmarks like ImageNet to real-world applications that significantly impact society.