I really enjoyed this conversation with Chris. Here's the high-level outline: 0:00 - Introduction 1:09 - CMU and the DARPA Challenges 5:12 - Red Whittaker and Leadership 12:54 - Hardest part about autonomous driving 14:00 - LIDAR 16:42 - Sensor suite cost 19:19 - Level 2 autonomy and the human factor 24:28 - Level 4 autonomy and demonstrating safety 32:00 - Public perception 34:40 - Deployment at scale 38:25 - Most impactful breakthroughs 39:50 - Pedestrian and cyclists 42:25 - Aurora path forward
they should do the next darpa grand challenge at LAX. damn near impossible to drive through departure/arrival without almost running someone over or getting rear ended by all those shuttle buses.
I was already wondering when Lex is going to interview Chris Urmson. I don't know whether Urmson talked about Aurora's recent breaking up with VW. It'd be interesting to know more about the background.
Ha! The intro had me laughin'. I wasn't completely sure you even saw my comments until now. Now that is funny. Maybe next time Lex, we can go for an even more subtle change, and you can spell out your first name as well. I must say, this is a strange back-and-forth, but I'm into it. Keep it up.
In 2015 in Germany 3475 people died due to an accident, 2049 of them died due to the misbehavior of car drivers. A large group of the car drivers, does not drink and drive, does not race on the streets, drives carefully, puts the smartphone away and never tired, has a safe and modern car etc etc, i.e. does the uttermost possible to drive safely. You cannot convince these people that autonomous driving is safer than they used to drive until you get under the number of non-selfinflicted deaths per year.
Off topic, but I want to share a random observation I found interesting. The sigmoid curves often used as activation functions in some types of neural networks closely resemble a cumulative normal distribution curve. Sorry if it's very basic. But I was intrigued by the potential usefulness. I absolutely love these videos!
Suggested question for anyone you interview about self-driving cars: "How good does it have to be?" We allow 15 or 16 year old boys to drive and they're far more dangerous than the average driver. Does a self-driving car have to be better than average? Much better than average? Or just better than teenage boys? If you require self-driving cars to be much better than average, a lot more people will die in the time period between when they're as good as average drivers and the time when they reach whatever threshold you set. This is even more true if self-driving cars are primarily used by teens and other drivers who are more dangerous than average.
Chris is a great leader and visionary, but the generic talk on autonomous was a bit disappointing, specially talking about the obvious rather than challenges facing us on Level 3-5.
Thanks Lex for another great podcast! Why infrastructure is not important as a breakthrough? As China is pulling massive infrastructure on a massive scale to support Autonomous age, I think that breakthrough is equally important and desirable for new coming technologies.
A simple solution to the "people walking in front of autonomous cars" problem, is to snap a photo and report to them to the police. They are identified through facial recognition and sent a ticket. Word gets out that you will get a ticket for walking in front of cars, most people stop doing it.
"When we trust the system we can do even more secondary activity, like smartphones" - I see that "thing" almost everyday in another cars around without any self driving vehicles or autonomy etc...
Very interesting point around 26 minutes about the driver trusting the car after 30 days. Kahneman and Tversky wrote about humans' inability to handle small probabilities. science.sciencemag.org/content/185/4157/1124
Great interview, like your style. Though I thought you were a little soft on the question of proving Level 4/5 capability. All regulatory bodies and insurance companies will require a proof of safety with x sigma better than humans over Billions of miles including all possible corner cases. Last I read it’s 10-15 Billion miles. Tesla is the only Company capable to produce those miles, so I agree with Elon: It’s game, set, and match for Tesla. Why did you not press him on that? Politeness?
it won't be fsd as in you can just hop in a car and go to sleep now, it's full self driving as in it has all the features, such as highway, parking lot, city streets, dirt roads etc. doesn't mean it will be completely ready to go yet, though it's mostly there, also its extremely likely some features will face delays and not come out soon.
I've had a Tesla Model 3 for a year. I've been operating various levels of autopilots for 60 years from aftermarket cruise controls on a '59 Volvo, boats and aircraft up to 435,000 lb. Lockheed L 1011 autoland airliners. I'm also an (old) EE but servo theory hasn't changed much. I think Tesla is going to succeed close to their timetable. For over six months my car is properly dealing with pedestrians 2 wheel vehicles and even cones and barrels. My exposure is of course minuscule but Tesla's fleet is not. Their ability to run software in a background or phantom mode on all the in service vehicles is ace in the hole.
I like how down to earth he is when talking about the actual product vs marketing approaches. This touches a very sensitive problem because if entrepreneurs aren't focused on this, despite the obvious repetitive (and misguided) marketing that's gonna keep selling "self driving vehicle" from level 2 to 5, the public perception needs to follow the actual real world levels, so regulatory agencies gotta stamp the SAE or similar the sooner the better, because the headlines focus only on the money.
I would like more discussions around how humans will interact with the L4-5 autonomous driver in future, asking it to take it easy since you just have av full cup of coffe in your hand, or ask it to drive faster maybe even break some rule bacause I'm late or my wife is giving birth in the car 😀, Will vendor x be faster from a-b on average, or will vendor y always be the kind one that let's the others pass a tight passage first? will humans be able to teach the autonomous driver or give it som helpful tips by speaking to it? like "there is a nasty damage in the road here so hold a bit to the left to avoid unecessary damage or better comfort" etc. I assume that ai-acceleration on other areas like speech recognition and incorporation of some Google assitant v5.0 will also happen in the coming 10 years... Also interesting to elaborate around what more intelligent systems that continously learn how how to please it's owner/customer but still be taking responsibillity for it's actions.
10 years ago I thought Urmson and Google were well on their way to autonomous driving. They both have lost their way. Listening to his response to Musk's lidar disdain convinced me that Musk has the correct path. Video can impute distance, speed and identify objects and driving surfaces by pixel analysis sufficient to steer cars. Lidar cannot read signs or identify tail end threats like video can. Ergo, you don't need lidar for level 4 and 5. You don't need maps and you don't need GPS. They can help in contaminated roads but Tesla has a GPS correction plan for that also. At some point contaminated roads are unsafe for any vehicular traffic. Tesla's vast 300,000 + autopilot fleet will be able to verify the software orders of magnitude faster than all others and clean up the tail ends much quicker. I understand all of Urmson's concerns about level 2 overconfidence. Tesla is blowing through that at lightning speed and will be past it within a year or less.
Lee Oldershaw I fully agree. Also, level 2 autonomy is clearly the stepping stone to level 5. Brushing off level 2 is stupid-you need that to acquire the training data for full autonomy.
I'm sorry, but I am very skeptical about the real world application of this technology. I live in a medium sized city that just launched an LRT system. Even though they are on rails, they still have "drivers". In less than 2 months of operation, there have been about 10 collisions with other vehicles.
The disappointing thing about this interview is that Urmson doesn't acknowledge how Tesla is pushing autonomy forward. Why not admit that Lidar was necessary in 2007 but not as important today, since vision is much better? Maybe because it doesn't fit the Aurora business model. You don't have to be a Tesla fan to acknowledge that they are the main reason the auto industry is so focused on autonomy.
Any self driving person on the program after the 4/22 Tesla investor day needs to be asked the question of Tesla's approach and to defend its use of Lidar, or the entire interview is meaningless.
The entire industry, apart from Tesla, is using Lidar and considers it to be vital part of a self driving solution. The fact Elon, has decided that it is not needed, and has thus forced his entire company to go that way proves nothing. At the moment the Tesla self driving capabilities are hugely lacking compared to some of their direct competitors, until this changes, not using Lidar is just as good as an Elon promise gets, and thats not very much.
I really enjoyed this conversation with Chris. Here's the high-level outline:
0:00 - Introduction
1:09 - CMU and the DARPA Challenges
5:12 - Red Whittaker and Leadership
12:54 - Hardest part about autonomous driving
14:00 - LIDAR
16:42 - Sensor suite cost
19:19 - Level 2 autonomy and the human factor
24:28 - Level 4 autonomy and demonstrating safety
32:00 - Public perception
34:40 - Deployment at scale
38:25 - Most impactful breakthroughs
39:50 - Pedestrian and cyclists
42:25 - Aurora path forward
Lex Fridman, you rock! I like your channel and especially these high-level outlines make your videos great!
Lex just knocking out podcasts with all the AI greats. Thanks as always Lex!
they should do the next darpa grand challenge at LAX. damn near impossible to drive through departure/arrival without almost running someone over or getting rear ended by all those shuttle buses.
I was already wondering when Lex is going to interview Chris Urmson. I don't know whether Urmson talked about Aurora's recent breaking up with VW. It'd be interesting to know more about the background.
Lex, would be great to have George Hotz on your podcast!
Thanks for doing this, Lex!!!
Great conversation! Very entertaining and good insights!
Imho This guy is still in his darpa days..
Ha! The intro had me laughin'. I wasn't completely sure you even saw my comments until now. Now that is funny. Maybe next time Lex, we can go for an even more subtle change, and you can spell out your first name as well. I must say, this is a strange back-and-forth, but I'm into it. Keep it up.
In 2015 in Germany 3475 people died due to an accident, 2049 of them died due to the misbehavior of car drivers. A large group of the car drivers, does not drink and drive, does not race on the streets, drives carefully, puts the smartphone away and never tired, has a safe and modern car etc etc, i.e. does the uttermost possible to drive safely. You cannot convince these people that autonomous driving is safer than they used to drive until you get under the number of non-selfinflicted deaths per year.
Off topic, but I want to share a random observation I found interesting. The sigmoid curves often used as activation functions in some types of neural networks closely resemble a cumulative normal distribution curve. Sorry if it's very basic. But I was intrigued by the potential usefulness. I absolutely love these videos!
WOW! Chris Pratt is a genius!!!
Fantastic interview, loved the insights. Great to get a differing view (to Elon's) of how LIDR sits in the technology mix.
Suggested question for anyone you interview about self-driving cars: "How good does it have to be?"
We allow 15 or 16 year old boys to drive and they're far more dangerous than the average driver. Does a self-driving car have to be better than average? Much better than average? Or just better than teenage boys?
If you require self-driving cars to be much better than average, a lot more people will die in the time period between when they're as good as average drivers and the time when they reach whatever threshold you set. This is even more true if self-driving cars are primarily used by teens and other drivers who are more dangerous than average.
Chris sounds almost exactly like sam harris
Change my mind
Yes he does. Even some of his speech patterns. 100 percent agree.
A mix of Sam Harris and Seth Macfarlane
Chris is a great leader and visionary, but the generic talk on autonomous was a bit disappointing, specially talking about the obvious rather than challenges facing us on Level 3-5.
Thanks Lex for another great podcast! Why infrastructure is not important as a breakthrough? As China is pulling massive infrastructure on a massive scale to support Autonomous age, I think that breakthrough is equally important and desirable for new coming technologies.
Great voice tho
A simple solution to the "people walking in front of autonomous cars" problem, is to snap a photo and report to them to the police. They are identified through facial recognition and sent a ticket. Word gets out that you will get a ticket for walking in front of cars, most people stop doing it.
Boy does he sound like Sam Harris
100% my thoughts exactly! 😆
"When we trust the system we can do even more secondary activity, like smartphones" - I see that "thing" almost everyday in another cars around without any self driving vehicles or autonomy etc...
He has Avengers logo!
Very interesting point around 26 minutes about the driver trusting the car after 30 days. Kahneman and Tversky wrote about humans' inability to handle small probabilities.
science.sciencemag.org/content/185/4157/1124
What happens when bugs smushed on the sensors and lidar?
the bug takes over
Great interview, like your style. Though I thought you were a little soft on the question of proving Level 4/5 capability. All regulatory bodies and insurance companies will require a proof of safety with x sigma better than humans over Billions of miles including all possible corner cases. Last I read it’s 10-15 Billion miles. Tesla is the only Company capable to produce those miles, so I agree with Elon: It’s game, set, and match for Tesla. Why did you not press him on that? Politeness?
Tesla FSD next year. Will that not happen? If the legislation will not go along, would this be the hinder?
it won't be fsd as in you can just hop in a car and go to sleep now, it's full self driving as in it has all the features, such as highway, parking lot, city streets, dirt roads etc.
doesn't mean it will be completely ready to go yet, though it's mostly there, also its extremely likely some features will face delays and not come out soon.
Jeremiah John: Are you an employee and have that kind of information? If so, I trust your words on this. Otherwise, we’ll see.
@@mrm.5787 not an employee, just researched tesla quite a bit, time will tell.
I've had a Tesla Model 3 for a year. I've been operating various levels of autopilots for 60 years from aftermarket cruise controls on a '59 Volvo, boats and aircraft up to 435,000 lb. Lockheed L 1011 autoland airliners. I'm also an (old) EE but servo theory hasn't changed much. I think Tesla is going to succeed close to their timetable. For over six months my car is properly dealing with pedestrians 2 wheel vehicles and even cones and barrels. My exposure is of course minuscule but Tesla's fleet is not. Their ability to run software in a background or phantom mode on all the in service vehicles is ace in the hole.
I like how down to earth he is when talking about the actual product vs marketing approaches. This touches a very sensitive problem because if entrepreneurs aren't focused on this, despite the obvious repetitive (and misguided) marketing that's gonna keep selling "self driving vehicle" from level 2 to 5, the public perception needs to follow the actual real world levels, so regulatory agencies gotta stamp the SAE or similar the sooner the better, because the headlines focus only on the money.
I would like more discussions around how humans will interact with the L4-5 autonomous driver in future, asking it to take it easy since you just have av full cup of coffe in your hand, or ask it to drive faster maybe even break some rule bacause I'm late or my wife is giving birth in the car 😀, Will vendor x be faster from a-b on average, or will vendor y always be the kind one that let's the others pass a tight passage first? will humans be able to teach the autonomous driver or give it som helpful tips by speaking to it? like "there is a nasty damage in the road here so hold a bit to the left to avoid unecessary damage or better comfort" etc. I assume that ai-acceleration on other areas like speech recognition and incorporation of some Google assitant v5.0 will also happen in the coming 10 years... Also interesting to elaborate around what more intelligent systems that continously learn how how to please it's owner/customer but still be taking responsibillity for it's actions.
10 years ago I thought Urmson and Google were well on their way to autonomous driving. They both have lost their way. Listening to his response to Musk's lidar disdain convinced me that Musk has the correct path. Video can impute distance, speed and identify objects and driving surfaces by pixel analysis sufficient to steer cars. Lidar cannot read signs or identify tail end threats like video can. Ergo, you don't need lidar for level 4 and 5. You don't need maps and you don't need GPS. They can help in contaminated roads but Tesla has a GPS correction plan for that also. At some point contaminated roads are unsafe for any vehicular traffic. Tesla's vast 300,000 + autopilot fleet will be able to verify the software orders of magnitude faster than all others and clean up the tail ends much quicker. I understand all of Urmson's concerns about level 2 overconfidence. Tesla is blowing through that at lightning speed and will be past it within a year or less.
Lee Oldershaw I fully agree. Also, level 2 autonomy is clearly the stepping stone to level 5. Brushing off level 2 is stupid-you need that to acquire the training data for full autonomy.
Completely agree. Urmson is not on the path to full autonomy anytime soon.
I'm sorry, but I am very skeptical about the real world application of this technology.
I live in a medium sized city that just launched an LRT system. Even though they are on rails, they still have "drivers". In less than 2 months of operation, there have been about 10 collisions with other vehicles.
The disappointing thing about this interview is that Urmson doesn't acknowledge how Tesla is pushing autonomy forward. Why not admit that Lidar was necessary in 2007 but not as important today, since vision is much better? Maybe because it doesn't fit the Aurora business model. You don't have to be a Tesla fan to acknowledge that they are the main reason the auto industry is so focused on autonomy.
the whole podcast wasn't that good to me, extremely light on details and mostly just discussion of ideas.
Any self driving person on the program after the 4/22 Tesla investor day needs to be asked the question of Tesla's approach and to defend its use of Lidar, or the entire interview is meaningless.
Pride and nostalgia. While I respect the contributions this guy has made, he is so far behind right now.
The entire industry, apart from Tesla, is using Lidar and considers it to be vital part of a self driving solution.
The fact Elon, has decided that it is not needed, and has thus forced his entire company to go that way proves nothing.
At the moment the Tesla self driving capabilities are hugely lacking compared to some of their direct competitors, until this changes, not using Lidar is just as good as an Elon promise gets, and thats not very much.