Comma.ai launches an $88 universal car interface called Panda
techcrunch.com
techcrunch.com
That's total bullshit. There are so many companies seriously working on autonomous vehicles, such as GM Cruise, Auroa Innovation, Drive.ai, Zoox, Nuro.ai, nuTonomy, Varden Labs, AImotive, Ford, and Nvidia
Edit, because George is on here: George, I want to like you, because I think your tech is really good (the street image segmentation from RGB images is seriously impressive- https://commacoloring.herokuapp.com/) and I think your business model of offering cheap devices and free services to crowd-source data, along with the comma points to incentivize users could give you a big competitive advantage. But the hyperbole makes it difficult to take you seriously.
If my sunday psychology profile is correct, he seems to be passionated and fast to hack around and displeased by big corporations business model + slowness + suboptimality. Thus trying to dig diagonally on his own without the enterprise cruft. So far so good, the issue is that it seems he's lying about his abilities. He managed to do 20% and spins it as 70% claiming he'll be at 100% before everybody else. Without independant testing and legislation of course.
I've a friend that has a comma neo with openpilot and it works enough well. Is not perfect, but they are on something.
The best projects that we know today were made by hackers with 0 enterprise skills.
I cannot compare anything in the casual software world, no matter how complex it is, to driving a car in streets. It's bridging physics and AI. Maybe hotz has a point and other SDV module vendors are as fake but that's irrelevant.
Edit: I'm actually curious now that I think about it. Are there any good examples of hacker projects (or specifically projects that aren't intended to make money or pay people's salaries) that had to deal with any sort of regulatory hurdles?
It worked for Cruise. Got them a billion dollars and a seat on GM's board.
That was surprising, because GM had a good self-driving effort with CMU, and had progressed to driving in Washington DC traffic autonomously.[1] That was way ahead of Cruise, and working in 2013. Then, somehow, GM totally blew it and lost that project, with the people going elsewhere.
[1] https://www.cmu.edu/news/stories/archives/2013/september/sep...
Every Toyota comes equipped with lane following and smart cruise control. They call it Toyota Safety Sense, and even the lowest priced, $18k Corolla has it.
Are you gonna hate this guy just because he has a face and a personal reputation?
Or are you going to watch with bemused interest as he successfully scales the obstacles and actually puts some nice products out there.
he released the hardware plans, specs, and software as opensource. You can go build it, today. And if you have a supported car (currently certain honda civics, and accura ILX's.. soon honda crv, toyota prius, and chevy bolt/volt) you can actually use it, today. There are videos on youtube of people doing exactly that.
Changing your business model, is IMO.. the opposite of folding. it's adapting.
That's a hell of a non-sequitor
Udacity has their own self-driving vehicle. They built it for their Self-Driving Vehicle Car Engineer Nanodegree program; I just started term 3 yesterday, and apparently they're going ahead with the plan for student to run code on it.
https://medium.com/udacity/term-3-in-depth-on-udacitys-self-...
Apparently, they have a test track, and will have a safety driver on-board. For that part of the term (dubbed "Integration"), we'll be broken into teams of 10-20 students, and (somehow) have to coordinate on building the software pieces to run on the vehicle. This will all be done using a custom simulator designed to replicate the test track (similar to what has already been used). Once something is working, the group will then merge the pieces (nodes and modules for ROS and something called AutoWare) all together in a shared github repo. That'll be shipped over to Udacity for testing on "track days", and results sent back to the team for analysis.
I'm not sure how well this whole thing is going to work; it's very much an experiment that I believe is unique for a MOOC and hasn't been tried before (develop software as a group project of international students, to drive a real self-driving vehicle on a test track in Mountain View, CA - yep, that seems reasonable).
I guess we'll see when we get there. The vehicle is named "Carla", and (overall) it is open-sourced. Here are some links to learn more about it:
https://www.udacity.com/self-driving-car
https://github.com/udacity/self-driving-car
https://github.com/CPFL/Autoware
The vehicle itself is apparently some standard vehicle converted by a third-party company (for the life of me, I can't recall the make or model of the vehicle, nor the company that did the conversion!); if you wanted to DIY this yourself at full scale, you'd basically want a vehicle that uses a mostly drive-by-wire control system for the brakes, accelerator, and steering. If that isn't available, then something that uses electro-hydraulic power braking and steering (or electric power-steering) would be the next best thing.
I'd shy away from trying to do this with a standard vehicle; building the actuators you would need to control everything is not a trivial thing to do (and especially to do so safely and reliably).
Honestly, if you really wanted to do this, and you didn't want to kill yourself, others, or damage property (as much?) - start out smaller with an RC car or similar. There are plenty of examples on how to do this on the internet, but my favorite go-to example is this one:
http://blog.davidsingleton.org/nnrccar/
That one occupies a "soft spot" in me because David Singleton was one of the original students in the Stanford-sponsored ML Class of 2011, taught by Andrew Ng. I was a part of that class; it taught us enough to build such a vehicle, and David was inspired to do so and follow thru with it. What used to cost hundreds-of-thousands of dollars (if not millions) back in the 1980s:
https://ftp.utcluj.ro/pub///docs/imaging/Autonomous_driving/...
...could in 2011 be done for much, much less.
Final note: I don't have any association with any of this other than being a student in the Udacity course and interested in learning about this technology.
And yes - we're running student code on the car, around a test track, and we'll send you the results (I think a video feed as well). It's your final project in the Self-Driving Car Nanodegree program.
I can't wait to watch the first time a student's code drives it around the track.
Hindsight is 20/20
Do you also need to be warned that if you create a plane autopilot, the FAA might get involved?
No, from my guess, Hotz copped out when he realized that his prototype, while fun, was no where near production ready and would be completely illegal to ship to any customers.
Hotz never seemed like the guy who would do things similar to what Cruise did. Vogt had already sold two startups prior to Cruise, and he was well versed in the tactics of the startup world. Have great articles put out, polished PR touches, the whole works.
OTOH, Hotz had this 'Hacker' air to whatever he does. From the iPhone, to PlayStation, to even hacking the cars he was building the tech with, he had a very different attitude compared to the let's-be-poised-for-a-buyout vibe that many startups give. Startups like Cruise are scared of breaking rules, worrying about how it might bite them later in the day.
If you ask me, Cruise was never going to make it on its own. The major car companies have decades of supply chain process management and huge amounts of capital behind them. Cruise couldn't never do what it said it would do, and I would imagine, a buy out was always on the founders minds.
Comma.ai is a different beast. They aren't really worried about rules. That's terrible attitude for a 'startup', I agree. People here have valid arguments about thinking about the NHTSA before hand, but I think they just dared to go ahead without giving a crap. It backfired, but I wouldn't attribute his decision to failure in building a working product.
So when George said, "there are only three real competitors: Waymo, Tesla and us." What he really says is that Comma.ai's technology is on par with Waymo and Tesla's. The rest of the companies are either subpar or have no existing product to demo, likely both.
He also famously said "shippability" is the biggest differentiator between Comma One and the rest but decided not to fight the uphill legal battle being that he's taking the Android approach to grab the market share.
So, I kinda buy it.
[1]: http://www.thedrive.com/tech/5772/the-cancelled-comma-one-wo...
Looking at the github repository code, comma is a joke.
Sorry what's a bridge? I don't get the reference.
He may have claimed that but he hasn't really.
The chffr app (video / nav data) + panda dongle (sensor data) is a simple way to let anybody participate in data collection, and for comma.ai to build enterprise value by accruing and organizing a gigantic data set that's critical to them or any third party executing the ML at core of self driving.
Now pair this with a tweet they posted in June:
"Just realized something. You are really going to be happy you have comma points. Get in early... http://getchffr.com/ #ICO"(1)
I think that's an incredibly compelling / fascinating ACTUAL use case for an ICO vs the current litany of pyramiding and outright scams we're seeing.
Take their chffr app + panda dongle and give people a financial INCENTIVE to accrue "points" by running the two data harvesting mechanisms and suddenly I am seeing comma's strategy in a new light.
Sure, you could install Hotz's dongle, give Hotz your data, earn "comma points", feel like you are helping to train self driving cars, and trade these comma points for - what exactly? a comma plushie? - while Hotz turns around and sells your data to insurance companies.
Or you could get a free OBD dongle directly from your current insurance company [1] and watch your rates go up/down depending on how safe of a driver you are.
Ultimately, Comma.ai needs to find an automotive partner. It will be nearly impossible for the company to beat Google, Uber, and (presumably) Apple, who all have strong partnerships with auto OEMs, in quality or quantity of driving data without direct integration and extensive testing on vehicles designed for autonomous driving. Not to mention Tesla, which is an OEM itself, and can design in data collection/processing for self driving vehicles from the ground up.
[1] https://jalopnik.com/progressive-insurances-driver-tracking-...
ICO as a way to convert "Internet Points" into something of material value and reward users directly for contributions is incredibly fascinating to me.
In theory Hotz could (and should) just pay you directly in $USD for your driving data, arbitraging a bit off the top before selling it to insurance companies.
But the genius of him "paying" you in cryptocurrency is that it will get you to give it away for free "points"! And to pay him for the hardware to do so! Because you think that a market might develop around these "points" where they will get bought up by a bigger fool, pumping up their value way beyond their fundamental value to the end consumers of this data - insurance companies.
These "points" have no other value or use, as they would not otherwise function as currency (store of value, medium of exchange, unit of account) unless Hotz decoupled point generation from car data generation and just started minting an arbitrary virtual currency.
Otherwise, it's just sending data to Comma in exchange for points. If you have one master account minting tokens (throughout the life of the currency), it's not really decentralized. They could have a SQL database with the same decentralization and security profile.
My guess is that the only use of blockchains here is to jump on the current ICO hype.
In that location, I'd first though it was one of Google's vehicles But why the Nevada plate? Turns out this is Embark.[1] They even show their AU 0030 license plate in the video.
Embark has been testing in Nevada, where the regulations are looser and there's less traffic. Their demo video is on a a deserted, straight two-lane road, closed for the occasion. Good place to debug.
Embark isn't on the list of CA DMV autonomous vehicle test companies. Also, CA won't allow autonomous vehicle testing of vehicles over 10,000 pounds, figuring that if something goes wrong, it's better to have car accidents than heavy truck accidents.
When Otto did that "autonomous delivery" of Budweiser with a semi, their vehicle was surrounded by chase cars. It was a much more controlled environment than the video indicates. There's too much "fake it until you make it" in this business. Waymo and Volvo are building systems that can handle unexpected trouble. Some of the others, not so much.
How ? What's different in the tech they are building ?
2017: Comma.ai will compete with ScanGauge
hm.
(http://www.businessinsider.com/george-hotz-techcrunch-disrup...)
In addition to this interface, we also give away source code to a driving agent: https://github.com/commaai/openpilot
Looks like this when running (video from a user): https://www.youtube.com/watch?v=-VN1YcC83nA
Is your plan to release pieces of a self driving kit gradually and crowdsource through early adopter kind of customers(or collaborator in a way) ?
I'm sorry but the hardest parts are still missing and reading you're competing with actual companies with serious power (brain and money) it's more than curious to stay polite.
ps: also waymon fleet might be 200, but if they inherited Google past years of data they have tons of terrain and sensors data. I mean wat.
Do you have any idea how ridiculous you sound? Waymo is a serious engineering effort with a lot of talented people working to do it right. There's a lot more to it than just how much data you throw at your NN.[0] Not to mention other companies comply with the law and report all their accidents.[1] Last year, you canceled your product[2] rather than answer a very short and simple safety questionnaire[3]. (Seriously, I've seen SaaS security compliance questionnaires with five times as many questions on them; that letter from the NHTSA was nothing!)
If you're still driving your vehicles around the streets of the SF Bay Area without any permits you are putting a lot of people in danger, including me. Pardon me if I think you're not only full of it, but irresponsible as well.
[0] http://www.explainxkcd.com/wiki/index.php/1838:_Machine_Lear...
[1] https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/testi...
[2] https://www.theverge.com/2016/10/28/13453344/comma-ai-self-d...
[3] https://www.scribd.com/document/329218929/2016-10-27-Special...
I have a lot of respect for your technical chops and the iPhone/PlayStation hacking you did back in the day, but the lofty claims here are hard to back up. At least Tesla has high margin electric car sales to keep them alive while they catch up to their claims.
It's easy to fool the lay public and tech journalists who don't have any technical backgrounds, but there is a world of a difference between doing fundamental AI research and development vs. patching together off the shelf commodity hardware and wiring it all up with some python.
That said, I wish you all the luck, I think a16z is betting on an acquisition like GM did with Cruise when they didn't really have anything that advanced either.
[1] https://electrek.co/2017/06/13/gm-self-driving-chevy-bolt-ev...
The product is aimed at hackers and car enthusiasts. You want to know your top speed, 0 - 60 etc? This might be the device for you.
Afaik there's already products out there that do the same thing. This could be at a lower price point though.
There doesn't seem to be anything new about this product other than marketing and a high price tag. I'd love for that to not be the case, however. Maybe I'm missing some features of it.
It supports 3x CAN, 2x LIN, and 1x GMLAN. It also charges a phone. On the computer side, it has both USB and Wi-Fi.
It uses an STM32F413 for low level stuff and an ESP8266 for Wi-Fi. They are connected over high speed SPI, so the panda is actually capable of dumping the full contents of the busses over Wi-Fi, unlike every other dongle on amazon. ELM327 is weak, panda is strong.
>George Hotz:
None of the < $100 ones can dump all the data, they don't have the bandwidth. Or be reliable enough to send a packet at 100hz for hours without dropping any. Trust me, I wouldn't be building hardware if I didn't have to.
[1] https://www.amazon.com/s/ref=nb_sb_ss_c_1_9?url=search-alias...
Main reason I haven't released it yet is that the connection management, while quite resilient, is so hackish that I am bit embarrassed by it.
The only time I've really used that data was the time some jerk stopped me, cursing at me for driving like a maniac and endangering his children (who he'd been allowing to play in the street). With the logged data, I was able to show that from the time I turned onto his street, I had been going no faster than 10mph below the limit.
The OBD-II error-reading software is very useful. It showed me that the Hyundai mechanics were completely incompetent. After holding my car in the shop for a week, they finally concluded that the problem is exactly what these diagnostics told me they were (plus 10 minutes of googling).
[1] https://play.google.com/store/apps/details?id=org.prowl.torq...
So I said "f-it" and dropped the coin for the original wired Kiwi, and that worked perfectly.
When they announced the Kiwi 3 kickstarter, I jumped on that with funding, and it works great - but with Bluetooth; it's a quality device. You get what you pay for.
It seems the real value for comma.ai is just collecting a huge amount of driver data and then selling it to Autonomous car makers or generating algorithms based or the data themselves. Either way, if that's the case why sell it for $88??? sell it at cost ($5) or even give it away for free and just quickly build a huge user base. Am I way off here?
Plus, psychologically it feels like an investment with ROI in "comma points" for those who do contribute their data, with some vague promise of a reward later when someone (who?) decides to purchase comma points. So the upfront cost is justified.
Service X (where X is an existing hit) for Area Y is getting popular.
The ones I found are either pretty expensive (e.g. https://freematics.com/pages/products/freematics-obd-emulato...) or seem sketchy (e.g. https://de.aliexpress.com/item/OBD-simulator-OBD-development...).
I am also aware of https://github.com/carloop/simulator. However I would prefer a one piece solution. Any tips?
https://en.wikipedia.org/wiki/OBD-II_PIDs
should be integrated with the ones by the given manufacturer on the specific model, and AFAIK all the money is in the software that can read those non-standard PID's.
And this is border-line (or downright just over the border) with hacking and/or reverse engineering the manufacturer's own software.
I only quickly looked over the source code so I am not sure if I missed something. But looking at the top level code, it seems like something anyone can accomplish after taking the Udacity SDC class (minus the CAN interfacing, but that shouldn't be all that hard?)?
Hotz confirms if you read the comments on the article.
Where these devices come in extremely handy, however, is by allowing you to look up the opaque fault codes that your car's computer produces when something goes wrong.
For $10-$20 off eBay or similar online stores, you can get a device that will tell you, for instance, when you have a cylinder misfire without having to go to a mechanic. (You may have to go to a mechanic to get it fixed, but at least you'll have a rough idea what needs fixing.)
Note that Hotz is claiming his readers will do a lot more than the conventional ones you've seen before. Whether or not that's true remains to be seen.
You set up Torque into logging mode, clear your codes, turn on the GPS and mapping, and drive away. When your car does "that thing" take a note of the time, then when you get home, go back thru the logs around that time mark, and take a look at all the sensor data, plus any thrown codes. If you know your car, and understand what is going on, it can help you to narrow down the diagnosis.
While a regular code reader (or getting the codes read) will help you by giving you the codes, being able to get the codes, timestamped in a log along with sensor data can go much farther in telling you where a particular problem lies.
While there are other code readers out there that will do this logging and tracking too, they typically cost hundreds or thousands of dollars - well out of the price range of most home mechanics.
There are many other things that Torque can do or be used for (for instance, virtual gauges are a nice feature - you can set up an Android tablet with a cluster of them giving you much more feedback than you can get from your regular dash instrument cluster, and it's cheaper and easier than setting up a bunch of physical gauges on a stock vehicle).
panda+cabana allows joe user to identify (and report back to the project) what they are seeing on the CAN bus.. so it can be catalogged and added to the database for all the thousands of different vehicle makes/models on the road.
Back in 2007 we read CAN data from a Toyota RAV4 to get wheel encoder, steering angle, acceleration, speed, and braking data [2].
I am surprised Uber/Lyft aren't using something like this. I often see the car location jumping around in SF FiDi.
[1] https://en.wikipedia.org/wiki/Street_canyon [2] http://www.acfr.usyd.edu.au/research/projects/others/darpa.s...
Combined with the phone camera, we can get and keep a really precise location fix.
Nowadays most vehicles have several CAN and LIN buses and it's unlikely that all traffic gets routed to the OBD2 connector, but some stuff is. Again, it depends on the manufacturer.
There are lots of standardized signals defined that a OBD device can request: https://en.wikipedia.org/wiki/OBD-II_PIDs
It is though, far short of everything needed. He's pitching it as perhaps part 1 of a bunch of other stuff you would need. That's aside from the idea that DIY self driving sounds, er, unwise.
I thought hotz finally found his spot and excepted some small nuc like board + a set of cam and sensors ..
I didn't understand what comma is trying to do and how selling 1000 "FitBits for your car" to "car enthusiasts and hackers" for $88 helps to compete with Waymo.
Current generation vehicle electrical architectures are not network safe and rely on physical security.
There is a big trade-off here for the next generation; the electrical architecture will be COMPLETELY locked down, and end users will be locked out of making any updates to modules - something that tinkerers and home-brewers are at odds with. Read only access will still be allowed.
It is exceedingly foolish to believe George Hotz's extremely obvious lies. George Hotz is a smart guy who made some big mistakes, and isn't man enough to admit that he was lying.
Follow us on twitter to see: https://twitter.com/comma_ai
I dare you.
Oh right, you'd never do that, because some tiny part of you knows that George Hotz was lying his ass off about his tech, and that you'd have to be a crazy person to trust such a flagrant liar with your life.
Maybe this Panda thing is engineered really well, and the money and reputation will get him to a point where he can run a nice company. Maybe not and this things get some people killed and he will be thrown in jail for building a recklessly unsafe product, I don't know. I don't own a car right now so all I can do is talk big.
The issue is that he was obviously lying. He claimed he was weeks away from autonomous cars. If this was true, he would've sold the tech for a staggering pile of cash.
Instead, all the evidence looks like he thought he could solve the problem on his own, in record time. Then he made a pretty basic POC (able to handle a small number of very simple conditions). Then he realized he couldn't do it. And so he pretended that he'd solved a problem nobody else had, but wasn't going to share or sell the answer with anybody else because the government is bad.
Unfortunately, some people hate the government enough that they want to believe this. But it's a dumb story and an obvious lie. George Hotz does not have autonomous driving technology. He never did, and he never fucking will. He's a snake oil salesman.
Working, video by a user --> https://www.youtube.com/watch?v=-VN1YcC83nA
Where's the snake oil?