Drago Anguelov (Waymo) - MIT Self-Driving Cars

Drago Anguelov is a Principal Scientist at Waymo, developing and applying machine learning methods for autonomous vehicle perception and, more generally, in computer vision and robotics. For more lecture videos on deep learning, reinforcement learning (RL), artificial intelligence (AI & AGI), and podcast conversations, visit our website or follow TensorFlow code tutorials on our GitHub repo.

INFO:

OUTLINE:
0:00 - Introduction
0:47 - Background
1:31 - Waymo story (2009 to today)
4:31 - Long tail of events
8:55 - Perception, prediction, and planning
14:54 - Machine learning at scale
26:43 - Addressing the limits of machine learning
29:38 - Large-scale testing
50:51 - Scaling to dozens and hundreds of cities
54:35 - Q&A

CONNECT:
- If you enjoyed this video, please subscribe to this channel.

[lexfridman]

Waymo vehicles have driven over 10 million miles autonomously, which is an incredible accomplishment. I'm excited to see what Drago and the rest of the Waymo team do in 2019.

[AR-iutf]

Thank you very much Alex. This lecture was very informative. Unlike most of the talks on self-driving cars where speaker spends a lot of time on listing challenges and peripherally on how to solve them Drago does the reverse. He shared a lot more with us within the bounds he can divulge which was extremely useful.
Thanks again to you and Drago

[TylerKoz]

Lex, this is a great service that you are providing to the world. Thank you.

[zakridouh]

These are the type of videos we want, keep it up and thanks :)

[ant]

Thanks for this. You're really helping people like me get their lives together with these kinds of videos.

[clauswirnsperger]

Interesting talk. Shows how complex the problem of autonomous driving is. Thanks Lex for the upload!

[TomHarrisonJr]

Exceptionally clear, concise and understandable presentation. Thanks!

[shivkuma]

1. How does this simulation + real world data approach generalize to other cities than Phoenix?


2. It appears that Tesla could also do a combination of simulation + real world data... No? What is the incremental value of simulation when you have a billion miles of real-world data vs 10MM miles ?


3. What is the value of the specific package of sensors vs the data set (simulation + real world miles) in terms of effectiveness -- are there metrics of effectiveness -- is it a edge-case driven issue irrespective of metrics?

[ritteradam]

The end of the talk about smart agents and learning aggressive / non-agressive behaviour was very interesting, as right now a known flaw with Waymo cars is that they can't take left turns in high traffic. It would have been great if Drago talked more about that specific problem.

[KamenRangelov]

Браво Драго ! Ще пусна видеото на сина ми с надеждата да се вдъхнови и да учи повече :-)

[leixun]

*My takeaways:*
1. AutoML for self-driving cars 22:08

[ebzdyberybencjusz]

17:50 so it's basically the same as Tesla, except Tesla has hundrends of thousands of cars collecting data for them. Waymo had to pay for their fleet of vehicles and has to pay their safety drivers, whereas Tesla has customers who do exactly the same not even knowing they train Tesla's NN for free.

[cafeliu]

Thanks a lot for sharing.
I have a practical question: As models grow larger it's harder for learners to follow up with new research. How should we become useful to ai companys without remaking the wheel?

[MichaelDeeringMHC]

If there are a bunch of waymo cars in the scene with non-waymo cars, will the waymo cars be communicating between themselves and collectively planning actions?

[scenFor]

Love Anguelov's machine ready definition of reasoning: "Check and enforce consistency of your beliefs. And look for explanations of the world that are consistent." Drago Anguelov

Here is one number for The Machine to reason through; #ConcentrationCamps #EndGlobalApartheid
Why there is so little data from flying over imaginary lines? And why the military spends so much more maintaining Exclusive Economic Zones than on preventing road accidents.

[carvalhoribeiro]

Awesome science experiments and ideas. Thank you so much Lex and Drago for lectures like this

[movaxh]

Talk is nice, but pretty general. The good pieces that are important, is usage of all kind of types of data, automation of training, testing at scale, automated machine learning, and good emulated (but simplified) models of other agents to test edge cases rapidly. Rest is just a lot of hard work, tuning and incremental improvements, that are hard to distill without revealing all kind of secret techniques used. To a big extent the perception modeling is solved, and the rest is just doing more complex modeling and prediction of other agents and special situations involving other agents. Unfortunately a lot of this is kind of circular, so again it is just grown over time to be better and better.

[jialuhan]

Highly agree with the Classical vs ML system analogy at 15:51.

[jacobdavidcunningham]

51:53 kind of wondering what's the reason that road was built like that(the squiggle I think in California)

[lennartlut]

Thank you for sharing these ideas, very helpful!