Lex Fridman — MIT researcher and lecturer teaching the 6.S094 Deep Learning for Self-Driving Cars course, focused on human-robot interaction in vehicles.
The gist
This is an MIT lecture on autonomous vehicles delivered by Lex Fridman as part of the Deep Learning for Self-Driving Cars course. He contrasts the utopian promise of saving lives and eliminating car ownership against dystopian fears of job loss, ethical dilemmas, and security risks. Fridman argues that the standard SAE levels of autonomy are useless for engineering, proposing instead just two paths: human-centered autonomy and full autonomy. Drawing on MIT's instrumented fleet of 25 vehicles and over 300,000 miles of Tesla Autopilot data, he claims drivers do not become dangerously disengaged, challenging decades of automation literature. He surveys sensors (radar, lidar, ultrasonic, camera), industry players, and where AI/deep learning can help with localization, perception, planning, and driver-state monitoring.
Big reveals
- Cites Rodney Brooks' prediction that a fully driverless taxi service in a major US city won't arrive before 2032, fourteen years out.
- Waymo achieved a fully autonomous trip in Phoenix with no safety driver in the car, calling it an incredible accomplishment.
- MIT instrumented 25 vehicles (21 Tesla Autopilot) collecting over 300,000 miles and 5 billion video frames of driver behavior.
- Confirms Elon Musk's claim that about 33% of miles in the data are driven autonomously in Autopilot, a remarkable adoption rate.
- Glance/attention allocation does not change significantly between Autopilot and manual driving, contradicting expectations of disengagement.
- In October 2016 Tesla dropped Mobileye and rebuilt its perception/control system from scratch using deep learning on NVIDIA DRIVE PX 2 and 8 cameras.
- Audi's A8 traffic jam pilot is positioned as the first L3 system where Audi accepts liability when the car is in control.
Things worth remembering
- 1.3 million people die in automobile crashes globally each year, with 35,000-40,000 in the United States.
- NHTSA's 'four Ds of human folly' are drunk, drugged, distracted, and drowsy driving.
- Technology adoption rates have accelerated over the 20th century, with each new technology going from 0 to 100% adoption faster than the last.
- Radar uses electromagnetic waves and ultrasonic uses sound waves, both calculating distance from wave bounce-back timing.
- Camera sensors have the longest range and best resolution but fail in darkness, snow, fog, and rain.
- Uber had driven about 2 million autonomous miles by December 2017; Tesla had over 1 billion miles in Autopilot.
- The 'lizard-owl effect' describes how some drivers move their head a lot (owls) while most move only their eyes (lizards) to allocate attention.
- In the data, smiling was the feature most associated with frustrated voice-based navigation, showing emotion in cars is context-dependent.
- Audi's traffic jam pilot is designed only for bumper-to-bumper traffic under 60 kilometers per hour.