Lecture
MIT Self-Driving Cars (2018)
- Autonomous vehicles are projected to transform society by saving lives through the elimination of human error categories (drunk, drugged, distracted, drowsy), while potential risks include job displacement in trucking, algorithmic fatality scenarios, and security vulnerabilities from hacked source code.
- Market adoption timelines are characterized by skepticism regarding short-term predictions, with full driver taxi services in major U.S. cities not expected before 2032 and widespread multi-city operations potentially occurring by 2045, though rapid societal acceptance could accelerate shifts overnight.
- The industry faces a technological divergence where human-centered autonomy (Levels 0-3) is viewed by many experts as irresponsible due to overtrust and human fallibility, whereas full autonomy (Levels 4-5) requires solving perception and control problems that may remain decades away from the necessary near-100% accuracy.
- Sensor technology relies on a combination of cameras for high-resolution data at low cost, LiDAR for accurate depth and 360-degree visibility despite high cost and weather limitations, radar for performance in adverse conditions under 200 meters, and ultrasonic sensors for close-range estimation, with future cost reductions in solid-state LiDAR expected to be critical.
- Commercial deployment milestones from 2017 include Waymo's first public rider and driverless rides in Phoenix, Uber's accumulation of two million autonomous miles, Tesla's fleet of over 300,000 autopilot-equipped vehicles, and Audi's planned end-of-2018 release of an L3 system where the insurance company assumes liability during designated traffic jam pilot use.
- Ethical and operational frameworks mandate that full autonomous systems must provide a "safe harbor" in emergencies, potentially overriding driver intent only during imminent danger, though society is currently considered unready for systems that autonomously prevent drunk driving or deny service.
- Future research directions prioritize deep reinforcement learning for control and planning, deep learning for scene interpretation, and advanced human-robot interaction models utilizing 3D convolutional neural networks to estimate driver cognitive load, eye pose, and state, as solving the human-robot interaction problem is identified as the immediate key challenge.