Topic: Autonomous driving
Uber’s autonomous cars drove 20,354 miles and had to be taken over at every mile, according to documents – Recode (Mar 16, 2017)
One of the great things about autonomous driving technology is that regulators require companies to keep track of how those cars are performing, and in the case of California that data is published annually, providing a great insight into how each company’s technology is advancing. However, occasionally, internal documents on testing emerge that provide lots of detail too, and such documents have apparently been leaked to Recode (and BuzzFeed). There’s lots of interesting data here, and it suggests progress is being made and Uber is driving lots of miles in its various cars. It’s worth comparing some of the numbers here for Uber with those reported by other companies in California by way of putting them in context: Uber says its rate of disengagements per mile is 0.8, for example, whereas Waymo’s cars in California are now at a rate of 0.2 per 1000 miles, or some 4,000 times better. Now, Waymo’s cars have been driving in the state for much longer than Uber’s, but that’s still a massive discrepancy in performance. And it’s also worse than Tesla’s rate of 182 disengagements over 550 miles driven in 2016. So it appears Uber has a long way still to go in autonomous driving, and it’s therefore remarkable that it’s already using these cars to ferry real passengers around Pittsburgh.
This feels like something of a slime ball move on Uber’s part on two fronts: firstly, trying to move the court case with Waymo out of open court and behind closed doors; and secondly, essentially trying to push the case off its back and onto Levandowski’s. I had said previously that the course was going to be fascinating for the details it would provide about how Uber developed technologies and how it would defend against what look like fairly solid allegations, but if it gets its wish here we won’t get to see any of that. But I think it’s the attempt to make this a case about an employee rather than the company that seems particularly sleazy – if the allegations are indeed true, then Uber and not Levandowski benefited the most, and making this seem like a dispute between an employee and former employer feels like a total misrepresentation.
via USA Today
Nvidia Partners with Bosch and Truck Maker Around Autonomy (Mar 16, 2017)
Here are two partnership announcements from Nvidia, the first a deal with automotive component maker Bosch to incorporate Nvidia chips in self-driving solutions, and the second with truck maker PACCAR for self-driving truck technology. Nvidia continues to be one of the biggest names in autonomous driving, and certainly one of the most successful chip vendors (hence Intel’s Mobileye deal). These deals come on top of lots of existing ones, but trucks are a particularly interesting area – it feels like that’s a segment of the market that could actually see real-world adoption of autonomy much sooner than cars.
For all the hyberbolic references to monopolies that sometimes afflict the tech industry, here’s a case where one company really does have what appears to be a monopoly, and on a critical component for autonomous vehicles: LIDAR. LIDAR is the same visual radar technology at the heart of the Waymo-Uber lawsuit, because they’re two of only a very small number of companies currently attempting to make their own, while everyone else buys them from Velodyne at $30-40,000 a pop. The global market for LIDAR is currently in the thousands, and the company expects to ship around ten thousand this year, but it and others would obviously have to ramp to tens of millions a year to supply the global automotive industry in the longer term. And those prices will come down massively – Waymo has supposedly reduced the cost dramatically for its own units.
via The Information
It’s striking to me that this piece never mentions Nvidia once, even though that chip company has been making much of the noise in the automotive space over the past year, especially when it comes to autonomous driving. That prominence is clearly a driver for this deal, with Intel signing some of its own deals but not getting nearly the buzz Nvidia has been. Mobileye, meanwhile, has been striking deals left and right with a variety of players (though it recently ended its relationship with Tesla). Just in the last three months, it’s announced partnerships with HERE and BMW over mapping, but it’s also in many of other car manufacturers’ existing cars and their autonomous plans. Given Intel’s ongoing struggles in the mobile space, its recent loss of the Microsoft server business to Qualcomm and the ARM architecture, and the ongoing stagnation in PCs, it needs some new drivers of growth, and in-car technology could provide that. Mobileye is also attractive as a business – it’s growing fast (almost 50% year on year in 2016) and profitable, with fairly high margins. So this isn’t an acquisition that will take years to contribute to Intel’s business, although its overall scale is still small. But it’s integration opportunity and the eventual opportunity to sell a joined up solution of chips and sensors which Intel will be focusing on here.
California DMV: Humans soon no longer required in self-driving cars – San Francisco Chronicle (Mar 10, 2017)
Michigan’s autonomous driving laws already allow testing of cars without drivers, and given that these two states are home to much of the testing going on, California clearly feels it needs to keep up. Those Michigan laws assume that carmakers are going to comply with all applicable regulations, and therefore require that any testing is done by or in partnership with those carmakers, while the proposed California law has no such restrictions (logical given the biggest local testers are tech companies and now legacy automakers). In both cases, the states are deferring somewhat to the National Highway Traffic Safety Administration to set the overall rules and to some extent approve cars for autonomous driving without a driver. This Chronicle piece quotes a spokesperson from Consumer Watchdog, which has been particularly harsh (perhaps deservedly so) on Uber/Otto, but also seems to be one of the main organizations demanding tougher regulation of autonomous driving in general in California. What’s interesting is that there are so few voices on the other side of this rapid push towards autonomous driving.
Alphabet’s Waymo filed an injunction against Uber for allegedly stealing intellectual property – Recode (Mar 10, 2017)
The fact that Waymo is suing Uber isn’t new, but this new step of filing for an injunction is, and that’s important because it could speed things up considerably. Judging the case in full could have taken months if not years, but a request for an injunction will involve convincing a judge in a much shorter space of time that there’s enough merit to the case for him or her to intervene in the near term. So we’ll know rather sooner how solid Waymo’s case here is, and will likely also get additional details from both sides about exactly what’s been going on. Importantly, we’ll get more from Uber than its brief initial statement about the accusations being baseless, which will be intriguing because from where I sit the forensic evidence looks fairly compelling. As I’ve said before, though, the toughest aspect of this for Waymo and its lawyers is proving that Levandowski actually used the files he downloaded rather than simply his memories of work he’d previously done.
California and Michigan have to be the two states where the most testing of autonomous vehicle technology is being done, with the former home to most of the tech companies in the space and the latter the home of several legacy automakers. The FT is here citing data from the California DMV, which you can see in its raw form here. What’s fascinating is the mix of companies here, as I’ve said before – there are several traditional carmakers (VW, Mercedes, Nissan, BMW, Honda, Ford, and Subaru), several big names from the tech world (Waymo, Tesla, Uber, Baidu, Faraday Future, and Cruise [now part of GM]), and a variety of other smaller companies. But Waymo has by far the largest number of cars and miles driven (and most accidents). But the California DMV is certainly the source of some of the most interesting data on self-driving testing anywhere in the world right now.
via Financial Times
Pay-as-you-drive insurance isn’t a brand new concept – indeed, I remember a colleague writing a report on this about five years ago when I was at Ovum. The basic concept is that the insurance company finds a way to measure actual driving behavior and then offers lower rates to those drivers who drive most safely. There are a number of pilots and active programs underway already, and this Tesla program just takes it a step further by focusing on drivers who turn on the Autopilot feature. Outside of this program, Root measures actual driving behavior through an app, but with Autopilot-enabled Teslas, there’s apparently no such hurdle to overcome. That’s great validation for Tesla (especially given the recent worries over its latest software), and also for autonomous driving technology as a whole – a key argument made by essentially all of its proponents is that it will be safer than human drivers. I’ll be curious to see if this program eventually gets expanded to cover other ADAS systems (since Autopilot is technically ADAS rather than autonomous technology), and whether Root’s data backs up Tesla’s claims about safety over time.
Didi has opened a self-driving lab in the U.S. with famed Jeep hacker Charlie Miller – Recode (Mar 8, 2017)
This seems like a smart move – even though there are lots of talented engineers in China, the nexus for development of autonomous driving today has to be either Silicon Valley or Detroit, so putting a base of operations in the former makes a lot of sense. There’s no evidence here that Didi is otherwise expanding into the US (after all, its new partner Uber is dominant here and that likely wouldn’t go down well), but that’s not to say Didi won’t try to hire from the other companies in the area. It’s already hired Charlie Miller, who came from Uber itself and was best known for having hacked a connected Jeep while it was driving a while back. The competitive intensity in this market, especially over hiring, is only likely to ramp up over time and things will get increasingly nasty as a result (and we’ve already got two lawsuits underway).
Self-driving cars are watching us and recording our data whether or not we’re watching the road — Quartz (Mar 7, 2017)
This article is part good reporting, part opinion, and comes with a clear point of view (which I’d articulate as “carmakers are collecting too much data on us and our driving behavior with insufficient transparency and opt-outs”). But the reporting is well worth reading whether or not you agree with that point of view: the piece does a good job of spelling out all the data that’s being collected by various automakers old and new, and what it’s being used for. And indeed, this data is critical for developing both ADAS and autonomous driving systems, because it’s only by measuring real-world human driver behavior at massive scale that cars can be taught both how to drive like human beings (which is important for trust and comfort) and how to drive better than human beings (which is important for safety). The legacy carmakers obviously have a big advantage here because they have many more cars on the road and hitting the road each year than newcomers like Tesla, let alone non-carmakers like Uber and Google. But it’s how that data is collected and used that makes all the difference here – putting advanced sensors in cars is critical to getting the rich data needed, but it also raises big privacy concerns which I suspect we’re going to hear a lot more about in the coming years.
It’s impossible to imagine any major car manufacturer putting out an ADAS system or autonomous driving technology that was as unready (and as apparently unsafe) as Tesla’s Autopilot software currently appears to be – it would be catastrophic for their brands and reputations. That’s probably the single biggest difference between Tesla and the major legacy automakers at this point, and it’s simultaneously what allows Tesla to move so much faster and what may end up causing major image, safety, and regulatory problems for the company as well. Moving fast and breaking things may be a fine motto for a social network, but it’s clearly not the right approach for a car. The very fact that the current feature set is said to be in beta feels like completely the wrong model for this environment. Tesla seems to be being helped by the fact that many of its drivers are early adopter types and eager to test even technology that isn’t completely ready, but I’m guessing they will feel differently if they or family members are hurt or killed in an accident because of this faulty steering and other erratic behavior. Tesla really ought to pull these updates and roll cars back to previous versions until it fixes the problems.
Toyota’s approach to autonomous driving strikes me as exactly the right one – as this article briefly explains, it’s approaching the problem from two different perspectives, one of which is about improving existing ADAS systems within the cars we’re driving today and in the near future, with the other being focused on Level 4 and 5 autonomy. I continue to be very skeptical that any car company is going to work its way incrementally and smoothly through the levels from 2 to 3 to 4, and believe much more strongly that we’re going to see a Big Bang shift from Level 2 to Level 4, which means that transition is likely to take quite some time. That doesn’t mean things like cruise control, self parking and so on aren’t going to get a lot smarter in the meantime, and that’s a good thing, but it does mean that true autonomy is both a long way away and likely to arrive all at once rather than incrementally. And of course because companies like Toyota have tens of millions of cars on the road already, they’re able to capture lots of data that will help with both the incremental ADAS and eventually autonomous technologies.
via The Verge
Well, well: a rare case of Uber caving to regulators and doing as they ask. It had seemed as though Uber had given up on San Francisco and California in general when it moved its self-driving Volvos to Arizona late last year, but it now appears that it is actually going to go through the steps necessary to gain DMV approval for testing self-driving cars in California after all. This all feels like a totally unnecessary rigamarole for all concerned – Uber has likely gained nothing and lost quite a bit of trust as a result of all this, and now it’s back where it started.
Uber’s self-driving unit quietly bought firm with tech at heart of Alphabet lawsuit – Reuters (Mar 1, 2017)
This is an interesting angle on the Uber-Waymo lawsuit over the alleged stealing of LiDAR technology by Anthony Levandowski – it appears Levandowski’s Otto acquired a company which specialized in LiDAR technology before it was itself acquired by Uber, providing an alternative theory for how the company was apparently able to get up to speed so quickly on the technology. One of Waymo’s key arguments in its suit was that Levandowski appeared to make unreasonably rapid progress on LiDAR following Otto’s founding, and that the only explanation was theft of ideas, designs and so on from Waymo. As an interesting side note, see also this newly-released October 2016 interview with Anthony Levandowski from Forbes, in which he somewhat bizarrely volunteers the information that he didn’t steal any IP from Google when he left. He also talks through his long history with autonomous driving technology, which raises a key point here: clearly Levandowski learned a lot about this technology over the years, and taking that knowledge with him to a new employer clearly isn’t stealing. So how does Waymo prove in court that Otto/Uber used the documents he allegedly downloaded rather than his personal knowledge (or technology from somewhere completely different) in designing LiDAR systems? If you know the best way to build a LiDAR system because you’ve done it before, are you obligated to act as if you have no idea how to do it when you move to a new employer? I’m not a lawyer, but I think some of these questions are fascinating, and are likely to be critical in this case.
This piece is a good reminder of three things: not all testing of autonomous vehicles is being done in California (or even the US), not all testing is being done by tech companies and startups, and countries, states, and cities are competing to be friendly to this testing. Old established carmakers are a long way down this road too – something that was borne out to me by conversations I had with a lot of them at the Detroit Auto Show in January – and they’re testing in their home markets as well as others. And cities like London are competing to be attractive to this testing, because it brings economic activity as well as a reputation for being friendly to technology in general. I learned to drive in central London, and wouldn’t really wish that on anyone, human or machine, but it sounds like the testing is mostly taking place in some of the less busy parts of the city, which makes a lot of sense.
During the very brief period when Uber’s self-driving cars were operating in San Francisco, one of them ran a red light. However, the company at the time engaged in some audacious spin and claimed the car was being driven by a human at the time and that the incident just highlighted the benefits of autonomy. Now, however, the Times is reporting that the car was supposed to be driving itself at that time and the human driver merely failed to intervene in a timely fashion. If validation were needed that the California DMV made the right decision when it stopped Uber from testing its cars without a license, here it is. But this is also yet another case of Uber acting like the rules that apply to everyone else don’t apply to it, and outright lying when it gets caught. And that in turn makes it very hard to believe it when it claims it’s in compliance with rules, and it only has itself to blame. What a terrible few weeks for Uber, pretty much all of its own making.
Waymo Sues Uber over Stealing of Confidential Information (Feb 23, 2017)
Alphabet autonomous driving subsidiary Waymo is suing Uber and its Otto subsidiary over alleged stealing of confidential information by Anthony Levandowski, who was one of the early executives at Waymo and subsequently left abruptly in early 2016 and immediately unveiled a self-driving truck company, Otto. That company, in turn, was acquired just a few months later by Uber. Waymo has done some fairly detailed investigate work that’s outlined in the complaint, and discovered that six weeks before Levandowski’s resignation, he downloaded lots of files from Waymo’s servers, and it argues that these in turn informed Otto’s (now Uber’s) LiDAR designs. As this blog post from Waymo says, fierce competition in autonomous driving technology is a good thing – it’s pushing the market forward rapidly and leading to some great innovations that should benefit consumers. But there are obviously lines companies shouldn’t cross as they compete, and this would be one of those, if it’s proven to be true. This is the second lawsuit in recent weeks involving employees moving between autonomous driving companies – the first involved Tesla and a startup. In both cases, the allegation is in part about stealing proprietary information. Given that Uber is already dealing with the fallout from a sexual harassment and discrimination blowup in the past week and still reeling from the #deleteUber campaign, this is terrible timing, but may also be a sign that the company’s aggressive stance on competition is hurting it in more ways than one.
A good reminder that even when an announcement is made, it often takes weeks if not months for it to actually take effect – Uber announced its move from San Francisco to the Phoenix area in December, but only now is it launching self-driving rides for paying customers in Tempe, a Phoenix suburb. In addition, we still have the disingenuous claims from the governor of Arizona that California was somehow not “open to business” for self-driving cars, despite being the home of the biggest trials in the country. The reality is that Uber wouldn’t comply with applicable regulation and made the decision to leave the San Francisco area rather than comply as others have done. For now, that must feel like good news for Uber – it gets to test its cars without the scrutiny or reporting requirements which would have been imposed in San Francisco. But whether this ends up being a good thing for the drivers and pedestrians of Arizona remains to be seen.
I’ve seen this announcement referred to as being about crowdsourcing in at least one place, and that’s exactly the wrong word to use, because this isn’t about a crowd of people at all, but about real-time data from vehicles. In contrast to crowd-sourced map data, which can easily be manipulated for humorous or nefarious ends, this is a closed-loop system in which anonymized data from BMW cars will help update HERE’s increasingly detailed and real-time maps. And that kind of up-to-the-minute map data will be critical for autonomous driving in future – it’s no good knowing what the road looked like six months ago (or even yesterday) if there’s construction, an accident, or a roadblock today. Putting this technology into one manufacturer’s new cars by itself isn’t going to generate that much data – there simply aren’t enough brand-new BMWs to be useful. But if HERE strikes similar partnerships with other carmakers, then over time it could end up with some of the best real-time map data out there. It’s a little hard to tell from HERE’s release, but the BMW/Mobileye release certainly suggests that the latter will also get to aggregate and use the data. This announcement also highlights the fact that, no matter how clever the technology from Silicon Valley startups, the companies with by far the most and best data will be the car companies and those that partner with them.