Chapter
The Challenges of Autonomous Driving Validation
1:54:48 - 2:06:53 (12:04)
Validating autonomous driving is harder than proving the software works since it needs to encounter problems in an infinite dimensional space as fast as possible. Getting relative safety of autonomous driving compared to human driving is a challenge since most safety cases are driven by physical operational driving data, including cases where the driver intervened.
Clips
Pre-training can provide a large boost in reducing data needs for models, as there are many signals that carry over across different domains and sensors.
1:54:48 - 1:56:20 (01:31)
Summary
Pre-training can provide a large boost in reducing data needs for models, as there are many signals that carry over across different domains and sensors. However, current models, such as behavior prediction for other vehicles on the road, still require continuous addition of data.
ChapterThe Challenges of Autonomous Driving Validation
Episode#241 – Boris Sofman: Waymo, Cozmo, Self-Driving Cars, and the Future of Robotics
PodcastLex Fridman Podcast
Analyzing traffic data proves to be challenging due to the sparsity of events that happen on the freeway.
1:56:20 - 1:58:32 (02:12)
Summary
Analyzing traffic data proves to be challenging due to the sparsity of events that happen on the freeway. Methods such as data augmentation and simulation are used to overcome this challenge.
ChapterThe Challenges of Autonomous Driving Validation
Episode#241 – Boris Sofman: Waymo, Cozmo, Self-Driving Cars, and the Future of Robotics
PodcastLex Fridman Podcast
The evaluation problem is the hardest part of autonomous driving, not the autonomy problem.
1:58:32 - 2:02:50 (04:17)
Summary
The evaluation problem is the hardest part of autonomous driving, not the autonomy problem. Evaluating final end safety and perfecting it can be a long path from a great demo, even when it looks like it's driverless quality.
ChapterThe Challenges of Autonomous Driving Validation
Episode#241 – Boris Sofman: Waymo, Cozmo, Self-Driving Cars, and the Future of Robotics
PodcastLex Fridman Podcast
The development of autonomous driving technology poses unique challenges compared to the development of aerospace technology, as there are significantly less opportunities to test the technology and prove its reliability.
2:02:50 - 2:06:53 (04:02)
Summary
The development of autonomous driving technology poses unique challenges compared to the development of aerospace technology, as there are significantly less opportunities to test the technology and prove its reliability. The Turing test may not be the most effective method for determining the readiness of autonomous driving technology, as the real challenges of the technology may not arise often enough to effectively reproduce and test.