Self-Driving Cars: State of the Art (2019)

Introductory lecture of the MIT Self-Driving Cars series (6.S094) with an overview of the autonomous vehicle industry in 2018 and looking forward to 2019, including Waymo, Tesla, Cruise, Ford, GM, and out-of-the-box ideas of boring tunnels, flying cars, connected vehicles, and more. This covers the state of the art in terms of industry developments and not the perception and planning algorithm development. The latter will be covered in detail in future lectures. 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
1:53 - 2018 in review
4:49 - Fatalities
8:29 - Taxi services
10:54 - Predictions
16:55 - Human-centered autonomy
19:42 - Levels of autonomy and proliferation strategies
24:48 - Out-of-the-box ideas
27:28 - Who will be first?
29:26 - Historical context
31:05 - Underlying beliefs of the industry and public
32:32 - Driving is hard
35:32 - Humans are amazing
37:10 - Humans and automation don't mix well?
41:55 - Two approaches: Lidar vs Vision
49:54 - In the meantime… data
52:49 - The road ahead

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

[lexfridman]

Vehicle autonomy is one of the most challenging and exciting problems in AI, and I believe it will be one of the defining problems in the field for the 21st century. As researchers and engineers we have the opportunity to create intelligent systems that save lives, increase access, productivity, and reduce the stress of traffic!

[TomHarrisonJr]

I have been following autonomous transport and heard a lot of talks, read a lot of papers, and drive a lot of miles (in my Tesla w/AP). Your presentation here is the most obviously correct and balanced overview, and the most cogent I have seen. Above all your passion for the topic is infectious. As as software engineer for 30+ years, now working in ML, I have been strongly considering a move into this space. What I saw here makes that even more appealing. Awsome.

[claudiomaiasantos]

I would love to see an update to this video reagrding the latest tesla conference and introduction of their new computer, machine learning and neural system.

[Carmen-nnhi]

This is why I love Linkedin, it sends me to things like awesome lectures! Kudos Lex, I didn't realize you were such a good teacher.

[ericruthenberg]

I enjoyed your technical description of the proliferation of cameras on cars. You said the amount of cameras was "humongous". Very precise. Thanks

[PoeRacing]

I really applaud the discussion regarding the nuance involved in comparing the relative safety of autonomous technology in development and manually driven cars. That it not often emphasized in these discussions. I am convinced that autonomous cars will be VASTLY safer than manual cars but there are still many engineering challenges to tackle. Managing the public perception of the risk involved in this development process is crazy important.

[DucNguyen-cqiw]

2021, I have been watching this lecture series during the Covid-19 lockdown. A lot of interesting stuff. Thank you for making it online available.

[markuslissheim]

I don’t know if it was clear that Lidar can not see through objects so they know if the car needs to stop or not. There would be a big difference between a paper bag flying over the road and a tire. The lidar would avoid both objects by maneuvering the car around it but vision and radar would be able learn what the difference of the two objects are.

[b-soaring]

Lex, a reasonable question, to which the answer is simply a year (not an essay) is "When will you be able to buy, or hail, an autonomous vehicle that will pick you up at home, you take a nap, wake up and you're at work". The answer can be (and often is) obfuscated with "it depends where I live" or "it depends on the law" or "it depends on my hair color", all of which are frankly irrelevant to the numerical answer which summarizes the perception of the person answering (assuming they know where they live). If you ask this same question over a period of years (as I have, to hundreds of people) you learn three things: (1) expectation is ahead of actuality, (2) expectation is sliding backwards faster than real time and (3) we're already past dates commonly predicted in 2015. FYI you can't collect these answers in a public forum without a nudge effect on other respondents. My colleagues who (like me) work in the industry are incapable of answering the question without the preceding essay explaining how some not-quite-autonomous solution will be great and how huge questions arise like whether to kill a crack dealer or get to work on time. But the best approach is discard the preceding hedges and caveats and just record the date. I'd be genuinely interested to know your answer.

[eddies]

Are you aware of comma ai? My prius is level 2 autonomous.

[Tadams]

Make Art. Art is never obsolete; art only creates more creatives.

[jstov]

much respect for posting your lectures on YouTube.

[ML-phbo]

I can’t help but think we should work toward having people live within walking and biking distance of their homes & use virtual presence to interact richly with others without shipping ourselves. Autonomous vehicles will be useful but we’re ignoring the broader issues.

[RikiB]

Why wasnt boston dynamics brought up during the motion comparison with humans (37:00)? Those robots falling over seem many years old and he calls them "state of the art"?

[DugOrion]

Lex - thank you for the great video. It really helps when thinking about the future.

[valdisgerasymiak]

Great video, looking forward to see updates in 2020)
I wish in my university would come people from Waymo, Audi, Voyage, Cruise, NVIDIA to read lections...

[winomaster]

So, how do the northern locations deal with snow.

When drivers can not see the lines, often they form two lanes where three lanes were before. How do self driving cars handle that.

[Noble]

I suspect given the vast network of the toll road systems they could be retrofitted with a closed loop Lane that will only accommodate autonomous vehicles.

[QuicksilverSG]

Here's an idea: Solve the autonomous drone delivery challenge first. Then after inner-city delivery traffic is substantially reduced, determine where autonomous piloting technology can be most productively introduced.

[pkla]

3 fatalities in a billion miles is an insanely amazing number.