Drago Anguelov — Robustness, Safety, and Scalability at Waymo

Rare Examples

Detecting rare examples is crucial for training effective autonomous vehicle models. Dragomir Anguelov highlights the difference between rare and hard examples, emphasizing that not all hard examples are beneficial to label. He suggests building models that estimate the distribution of features to identify truly rare instances 1. Anguelov also discusses the use of active learning and data augmentation to maximize the value of collected data 2.

Our domain is ripe for finding the rare examples. It's one of the main tasks you need to do. Most of the time you drive, it should be boring. And we collect a ton of data, which is great.

--- Dragomir Anguelov

This approach helps in efficiently utilizing the labeling budget and improving model performance.

   

Adversarial Challenges

Addressing adversarial challenges is essential for the robustness of autonomous vehicle systems. Dragomir Anguelov explains that using multiple sensors and redundancy can mitigate adversarial attacks, making the system more robust 3. He also mentions the importance of unsupervised domain adaptation to generalize across different conditions, such as varying weather.

Hybrid stack with multiple different sensors is more robust. Part of the beauty of having active sensors is one of them can fail and they can still fairly independently detect things for you.

--- Dragomir Anguelov

These techniques ensure that the system remains reliable even in challenging scenarios.

   

Scalability

Scalability is a significant concern in deploying machine learning models for autonomous driving. Dragomir Anguelov emphasizes the need for robust systems that can generalize well by incorporating the right structure and inductive biases 4. He also shares his experience with the challenges of transitioning from research to production, highlighting the unexpected issues that arise when scaling up 5.

In a large enough data set, everything that can go wrong does go wrong. And then you find a whole set of yet more rare cases that you need to worry about.

--- Dragomir Anguelov

These insights underline the complexity of creating scalable and reliable autonomous systems.