Udacity plans to build its own open-source self-driving car
techcrunch.com
techcrunch.com
So, this is super cool. I'm surprised it's Udacity doing it; on a couple of fronts:
It seems, on the surface, outside of their core competency...but thinking about it, it does make some sense, if they really have enough paying students (and maybe sponsoring organizations) to make it work. I mean, schools that teach auto repair don't work on fake cars. Why would a school teaching self-driving limit themselves to simulated cars?
But, it's also surprising that Udacity has the funds to make it work. The cost of building self-driving car technology must have come way down just in the past few years. And/or Udacity must be making a lot more money than I would have guessed based on how crowded their market is.
Regardless, it's super cool!
The DARPA Grand Challenge in 2005 and Urban Challenge in 2007 are what started it all.
Now that Google, Uber and others are racing to commercialize, all the government has to do is not overregulate. No need to "push forward an agenda".
(FWIW DARPA, then known as ARPA, did the same thing the internet a few decades ago.)
I'm more surprised/disappointed that someone on HN wasn't aware of those initiatives. It shows why people are for defunding government research. Though I imagine DARPA being part of the military, wouldn't have been in that boat, but my point remains the same to NASA, NSF, NOAA etc.
You've made an assumption that I'm unaware of them, which is incorrect.
I just think they're very small in the grand scheme of things. The DARPA challenges are a blip on the radar of the federal budget, as it relates to highways, auto safety, fuel efficiency, and a variety of other areas where self-driving cars will have a tremendous impact.
Seems like creating free and open source simulator would be of more value than trying to get students to build hardware based vehicles?
DARPA, YC, and other have taken this approach for various reasons in some projects - and it seems if the intent is to teach, learn, share, etc. - it'd be a better investment.
How so? We have students all around the world (almost every country is represented in our DAU) who will never have access to a car outfitted with hardware and sensors (easy $125k), never mind the costs needed to get a permit to get on the road ($50k!). Being able to contribute code and see the results run in real environments (ask Sebastian what he thinks about simulation!) could be a huge advantage to students around the world in their quest to jump into this industry and get credibility.
Speaking of credibility, we want to prove to the world we really know what we're doing, and that our curriculum is truly legit.
tl;dr: We hope that by open sourcing our car that we can give opportunities to students around the world who otherwise wouldn't have it.
Thanks, appreciate you addressing the questions.
If you wanted to ping Sebastian and let him know about the AMA, I'm sure there would be a lot of interest beyond just me getting an answer to my questions from him; I truly am curious to hear his take, since as you say, the barrier to rendering the software in hardware would likely be beyond the reach of an individual student or even likely a group of students in closes proximity to each other.
I did simulations before doing anything in practice. Benefits were many because there is nothing like the real thing.
Edit: Removed presumptuous comment about myself
We're taking a lot of learnings from our Georgia Tech OMS program to find amazing students from around the world.
The response from Udacity suggests not.
My example might be a bit contrived, but I think there are going to be many valid (and far better!) questions in this discipline that should be asked and considered, and I think your graduates need to be equipped to do so.
We were strictly obeying the course boundaries, and had a terrible time getting through narrow gates where DARPA's waypoint file had a narrow width designed to guide us through the gates. If you look at videos of our runs, you can see the vehicle backing up and trying to get through a narrow obstacle. It's trying to get past a real-world obstacle on one side and a GPS limit on the other, which has narrowed the allowed path to where it can't quite fit.
So for the second run, I put in a patch to add 1 meter to DARPA's lane width. But I forgot to push it out to the vehicle, and we botched the second test run. It was in place for the third test run, though.
My friend owns a Model S and on the way to lunch it got confused on off ramps and parts of the road that were not well defined by lines. It basically shut off autopilot and required my friend to take the wheel.
In the 2004 Grand Challenge, the bounds were much wider and most vehicles screwed up. As it turned out, you could almost drive the 2005 Grand Challenge by staying centered in the DARPA-defined path. But we didn't know that in advance.
Driving with purely cameras isn't too far away.
See Comma's research: http://comma.ai/research.html
Arguably the biggest difference between the 2004 challenge (best distance was 7 miles) and the 2005 challenge (5 vehicles finished the race) was simply experience and refinement of existing technology. Stanford and CMU (top finishers) completed the race with very different approaches.
The Velodyne lidar was a prototype in the 2005 challenge and the first commercial version was extremely valuable to the teams that used it during the Urban Challenge. Google relied heavily on it and similar sensing technologies while developing their fleet, and many other groups going after full autonomy have also relied on lidar, so it certainly continues to be important. There's lots of different opinions on the future value of lidar vs vision. Camera quality and processing power + the effectiveness of CNNs are pushing a lot of people towards vision as a primary sensor.
There's also been a lot of progress for driving in urban environments in modeling, mapping, and prediction. A lot of this comes from collecting lots of data and building maps and behavioral models for objects the vehicle will interact with (cars, people...) Obviously for that the advancements in ML and deep learning don't hurt.
Consultant at Bain or McKinsey
Developer at a Big 4 tech company
Investment banker on Wall street
Other highly competitive jobs
(in a situation where experience alone wouldn't have earned the position)
For example since I attended a mediocre state school these companies did not come to interview at my school, and never granted interviews to those who directly applied.
A few years later I ended up getting hired by these same companies based on the merit of my actual experience.
However you see the big difference here? I couldn't fairly say that the big 4 hire from mediocre state schools (or from Udacity), just because someone ended up getting a job there later.
We do indeed have people who got jobs fresh out of graduating from our program. Obviously it's hard for me to discount all their prior experience and say it was entirely up to us. What I can confirm is that within months of graduating from our program they were able to get these jobs. You can read about some of these students here: https://www.udacity.com/success
The purpose of top degrees is signaling. Udacity, by its very mission (trying to democratize education), cannot offer that signal.
Of course, people with Udacity degrees likely already do get top jobs. They're not using the signal of the Udacity degree to get them though.
This is obviously so far outside of Udacity's core wheelhouse that I have to assume it's simply an ego project for the founder. Unfortunately, before you start pursuing unrelated ego projects, your company should have several billion in cash in the bank.
I cannot possibly see how this ends up working out well for Udacity. Developing self-driving cars is very expensive and not their core expertise at all.
I get quite frustrated with articles and headlines and even analysts who don't understand the quite fundamental difference.
Uber, who lives almost entirely in that space, will be in for a rude awakening IMO. Unless they already know they are ten years behind Google, and plan to use instrumented human semi-operated vehicles for learning, and cynically market it as "self driving" when it isn't. That'd work, too.
"Kickstarter - Self-driving car - release in 2019! 4795% funded! Backers get one for just $10,000!"
"Stock tip - self-driving car startup - invest now to get pre-market shares in this company! Top Secret!"
That sort of thing. When the TV morning shows start pumping self-driving cars every day.
oops.. Ex-Googler Sebastian Thrun says the going rate for self-driving talent is $10 million per person http://www.recode.net/2016/9/17/12943214/sebastian-thrun-sel...
Google's program is a tech demo for PR purposes. Everyone else is trying to actually put systems in cars.
Thrun guesses that Lidar will ultimately be unnessecary, and I think that by the time the hard AI problems are solved, he'll be proven right. Humans navigate with a pair of eyeballs and not much else, after all. But while extraneous sensors can always be removed, not having enough could hamper progress.
Only if by "not much else", you are referring to a ridiculously performant image processor - the visual cortex does an amazing job! My guess is it will be many decades before we can get similar performance in hardware/software.
Behind that is computer vision, revolutionized by deep learning. It can identify street signs, makes and models of vehicles and other road objects, it can keep a car centred in a lane when there's limited information such as snowy conditions, or with chaotic visual information such as a sun dappled country road covered in leaves.
Google has patented a method for interpreting the hand signals of police officers. They can read cyclist signals, and interpret the body language of pedestrians to interpret whether they intend to cross the road or not to avoid false braking incidents.
And Nvidia has demoed a way to build 3d point cloud fields (like Lidar) using off the shelf cameras. Robust computer vision is heavy on compute, but they're doing it.
But behind all this is the higher level reasoning needed to deal with tens of thousands of edge cases- this is the hard part.
Google made a special project out of programming their cars to effectively assert themselves at four-way stop signs, which is one problem amongst many.
No single edge case is unsolveable, but taking them all on, and getting the whole rube goldberg machine to six sigma reliability is an epic slog. This is the part that's going to take another 10 or 20 years, though we'll probably see L4 applied in a constrained capacity fairly soon.
That may be true, but we are still ways from catching up to human performance on things like "where is the edge of this dirt road"
There are so many other important problems in the world that need solving
If self driving cars become mass marketed, then that's reduce death/injuries, less energy consumption/pollution from more efficient driving, more time saved from efficient driving, possible reduction of manufactured cars from a sharing economy standpoint.
I used ros a bit for drone simulations 5 or so years ago, things might have changed these days.
This is the story:
Online correspondence school announces it's making a self-driving car, and issuing "nanodegrees" of dubious reputability.
I'm sorry, but what does a correspondence school have to do with self-driving cars? How does this promote the core business? How does even relate to the core business? Does Udacity have any technical talent that would even be relevant to this? (I'm betting they don't have any computer vision experts.)
This strikes me as a company with too much money, and not enough supervision.
This is not going to end well, for anyone.
Sebastian Thrun is a cofounder of Udacity, and one of the world's top experts in self-driving cars. Sebastian Thru won 1st place in the DARPA Grand Challenge in which self-driving cars raced across the Mojave desert. He is also a VP at Google, where he has worked on Google's self-driving car technology. One of the first courses at Udacity was called "Artificial Intelligence for Robotics."