Lecture
MIT Self-Driving Cars (2018)
- Safety Potential: Autonomous vehicles aim to eliminate the "four D's" of human error (drunk, drugged, distracted, drowsy driving) which contribute to approximately 1.3 million global annual traffic fatalities.
- Economic Impact: Shared mobility models enabled by autonomy could reduce the cost of transportation by one to two orders of magnitude by eliminating individual car ownership.
- Employment Concerns: Significant societal fear exists regarding job displacement in the transportation and trucking sectors due to AI integration.
- Ethical Liability: A core philosophical challenge involves accountability for algorithmic decisions resulting in human fatalities, particularly when systems operate as "black boxes" regarding ethical grounding.
- Security Risks: The reliance on millions of lines of source code creates vulnerabilities where external hackers could manipulate vehicle behavior to endanger human life.
- Adoption Forecast: Rodney Brooks predicts the earliest viable driverless taxi service in a major U.S. city will not occur until 2032, with multi-city availability by 2045.
- Tech Adoption Trend: Historical data indicates that the time required for technology to achieve 0-100% societal adoption has decreased significantly in the 21st century.
- SAE Levels: The SAE J3016 taxonomy defines levels 0–2 as increasing automation, level 3 as conditional automation (constrained geographically), and levels 4–5 as high/full automation.
- Human-Centered Autonomy (A1): This approach includes levels 0–3, requiring human intervention upon system failure; liability remains with the human driver.
- Full Autonomy (A2): This approach encompasses levels 4–5, where the AI is fully responsible; systems must be capable of bringing the vehicle to a "safe harbor" stop without human intervention.
- Takeover Constraints: Current human-centered systems provide no guaranteed time for human takeover; reaction windows can range from zero to several hundred milliseconds depending on the scenario.
- Teleoperation Nuance: Remote human control does not constitute full autonomy; as long as a human is in the loop, the system remains human-centered.
- Data Scale: MIT researchers have collected over 300,000 miles and 5 billion video frames from 21 instrumented Tesla Autopilot vehicles to analyze driver behavior.
- Autopilot Usage: Data indicates that 33% of miles driven by instrumented Teslas are performed using Autopilot, contradicting aviation-based fears of immediate user disengagement.
- Driver Attention: Analysis of 5 billion frames revealed that driver glance patterns (looking at road, cluster, or phone) do not significantly change between manual driving and Autopilot usage in the dataset.
- System Failures: The research highlights 8,000 recorded manual-to-machine transfer of control events, often driven by driver discomfort or desire to deviate from the path.
- Sensor Fusion vs. LiDAR: A fusion of low-cost sensors (cameras, radar, ultrasonic) can currently match LiDAR performance, provided LiDAR costs decrease and ranges increase in the future.
- Radar Characteristics: Radar offers low resolution but high reliability in adverse weather (rain, snow) and is the industry standard for speed detection.
- LiDAR Characteristics: LiDAR provides high-resolution 3D mapping and works well in darkness but fails in snow/fog and remains prohibitively expensive for mass adoption.
- Camera Characteristics: Cameras offer the highest resolution, longest range, and lowest cost, making them ideal for deep learning, but struggle with depth estimation, night vision, and adverse weather.
- Waymo Milestones: Waymo achieved full autonomy (no safety driver) for public riders in Phoenix in November 2017, accumulating four million autonomous miles by that date.
- Tesla Autopilot Scale: Tesla reported over 1 billion miles driven via Autopilot by late 2017, utilizing a custom deep learning stack on 8 cameras with no reliance on LiDAR.
- Audi L3 Definition: The Audi A8 Traffic Jam Pilot (late 2018) is classified as Level 3 only if operated under specific constraints (bump-to-bump traffic <60 km/h), where the system assumes liability.
- Visual Odometry: End-to-end deep learning methods (e.g., DeepVO using CNNs and RNNs) now outperform traditional SLAM approaches for camera-based vehicle localization.
- Audio Perception: Machine learning can detect road surface conditions (e.g., wet vs. dry) by analyzing tire noise spectrograms.
- Driver State Monitoring: Key metrics for driver state include head pose, blink rate, cognitive load, and binary classification of "looking on-road" vs. "off-road."
- Emotion Detection: Counterintuitive findings suggest that smiling drivers may report lower satisfaction with voice-based navigation, while stoic expressions can correlate with high satisfaction.
- Behavioral Challenges: Full autonomy systems must learn to "assert" themselves in complex environments like busy crosswalks and unsignalized intersections where polite stopping is socially counterproductive.
- Future Roadmap: The course prioritizes solving perception and control problems for 99% accuracy via deep learning, acknowledging that near-perfect full autonomy may be decades away.