The mood in the AV space is slowly moving to the realization that, hey, full Level 5 autonomy is hard, really hard, and maybe, just maybe, it’s not a reasonable or even desirable destination to target. Apple, for example, just revealed that it’s dramatically scaling back their plans for their somehow still-upcoming car, plans that once included no steering wheel or pedals and provisions for sleep. And our pals over at Tesla seem to now be backtracking on their plan to use a camera-only system, called Tesla Vision, for their automated driving systems and are now looking to add radar back into the mix. Argo AI went out of business, at least in part because Ford realized Level 4 was going to take longer to achieve than originally thought. That last one is especially surprising because out of all these autonomous driving startups, Argo AI had its shit together much more than many others. Everyone seems to be wising up. At least, I hope they are. The Apple announcement is especially interesting, because, at least according to the Bloomberg report and via their anonymous source, we can see a pattern of thought that I think shows a realistic state of the current technology and also how many questions there are still left to be answered. I say this based on these quotes from the Bloomberg story: These quotes are significant for a number of reasons: first, if Apple – a company I think we can all agree is generally extremely capable technically (who among us doesn’t use a Pippin or Newton on a daily basis?) – says full autonomy is beyond our current abilities, then I think that carries some weight. But even more interesting are the next parts, where the anonymous Apple person states that the plan is for “full autonomous capabilities on highways,” so, a sort of Level 4 system, where the automation is restricted to certain areas and situations. …the company is now planning a less-ambitious design that will include a steering wheel and pedals and only support full autonomous capabilities on highways…Apple plans to develop a vehicle that lets drivers conduct other tasks – say, watch a movie or play a game – on a freeway and be alerted with ample time to switch over to manual control if they reach city streets or encounter inclement weather.”
That’s fine, but then the quote goes on to describe how the driver can do other things, pay no attention to driving, u ntil they need to be “alerted with ample time” to take over, and cite conditions as common and capricious as weather as a possible reason why the driver would need to take over. What this statement tells me is that Apple has a hell of a lot of work to do. Really, everyone in this space does, because so far I have yet to encounter a viable and plausible explanation from any AV maker regarding how a handoff from a completely disengaged driver to take over a car traveling at highway speeds would work. Weather changes, like the quote from the Apple insider mentioned, can be sudden. What if the driver is asleep? Or drunk or wearing headphones not connected to the car’s audio system, or crying or laughing hysterically or moments away from orgasm or helping a child or pet or any number of things a human can be doing? How should the car handle a handoff where the driver doesn’t respond? And that’s just for planned handoffs, where the car knows it wants to pass control. What about when the car miscalculates something, or a sensor fails or it encounters some bit of input that confuses it or the sun emerges from behind a tree to blind a camera or any number of other issues? Limiting the operating domain can help, but we’re still just talking about a normal highway here, at presumably highway speeds. Even if we abandon full Level 5 go-anywhere-do-anything autonomy as a pipe dream, there’s still so much more to figure out. Now, I have no idea who this quote came from at Apple. I have no idea if it’s an engineer or an executive or the person who decided to not include a Micro SD slot on the iPhone. What I do know is that Apple’s new revised plans – from the little I can surmise from that quote – sound naïve. Again, it’s not alone. Tesla potentially adding radar back to their suite of hardware after declaring that all they needed are cameras and that their cars were hardware-ready for full autonomy was also naïve, especially considering Tesla made the same claim in 2016 and had to upgrade hardware once already. I hope all of these companies finally admitting that full, independent Level 5 autonomy may be out of reach for the near future will finally turn their attention to all of the other, non-technical problems that need solving, especially the issues of how to deal with uncooperative handoffs and failure management for when things inevitably go wrong. The whole industry needs to get over their idiotic pride and work together on the key problems of getting disabled AVs out of active traffic lanes and a common, expected procedure for dealing with a handoff that never happens. This is the unsexy stuff, and it likely will require regulations and perhaps infrastructural changes and hardware, but there’s really no way around it if we want to find a non-magic Level 5 system to actually work in our reality. Want special highways where you can comfortably and reliably peace out while your car drives? Then we’re going to have to get everyone to cooperate. And there’s hardly anything the American automotive landscape is worse at than standards and cooperation. We don’t even have a standard EV charger port, for fudge’s sake. (This is also why a country with more centralized control like China could very easily beat us to having semi-automated highway travel.) I feel like some companies get it; I recently spoke with Volvo’s autonomy partner and was impressed with their approach (full interview coming soon) because it reverses the usual paradigm: it’s not pitched as a labor and attention-saving system, but a safety backup. Where most Level 2 semi-automated systems are designed so the driver monitors the car and is ready to take over, Volvo’s setup keeps the driver in control, with the computers watching, ready to take over if needed. This feels like a more realistic and usable approach to maximizing the safety benefits with the current level of technology while avoiding the usual Level 2 vigilance problems. Overall, I’m hopeful that these recent developments actually are signaling that we may finally be breaking out of the Unlimited Level 5 pipe dreams, the unwavering pursuit of which only impairs our ability to implement less comprehensive but still extremely useful systems that could enhance safety and give at least the most important automated driving benefits. Remember, the best is the enemy of the good, even the good enough, especially if the best is some magical bullshit that’s probably not even worth banging our heads against the wall to achieve. If you remember the old analogy “speed kills”, you’ll start to understand why highway driving is more difficult right now. The faster you go, the less time you have to react. In addition, a sensor that works well in an urban scenario may be unsuited for the highway because of it’s range (the other way you can buy more time to react). The further ahead you can “see”, the more time you have to react. Current sensors can’t see far enough ahead to safely react or allow enough time to notify the driver to take over (Level 3). This is one reason Musk was so full of it when he said Tesla didn’t need LiDAR (or even RADAR) for their system. Typical vehicle RADAR have a range of around 100 meters. Vehicle LiDAR will have a range of 200-300 meters. 300 meters triples the amount of time a system has to react which is crucial to highway safety. LiDAR also gives you a 3D map over a vertical and horizontal field of view that the current RADAR systems cannot match. Mercedes is currently selling vehicles with Level 3 autonomy in Germany, but they are limited to 60 km/h and below. They are equipped with LiDAR, but they are short range. The fidelity and awareness they are getting from the LiDAR increases system environmental awareness, but range is not sufficient for highway driving. Mercedes’ Level 3 system is coming to the US next year. Like Jason has been saying, autonomous driving is not an easy problem to solve, but focusing on safety rather than hype is the correct path forward. Unfortunately, not every player in the space is taking that approach. Hope my post has given you a bit of a new perspective on ADAS and where we need to go. https://i.imgur.com/KSqjkuE.jpg The current state of affairs is fairly scary. When I started riding many years ago, another rider noted that you’ll all of sudden notice how many people aren’t paying attention while they drive. These days, it’s the majority of drivers (not the people here, you guys know I don’t mean you!), and as I get older and my reflexes slow, it gets more and more concerning. I like the idea of a more predictable technological solution to this, but of course, it’ll be dependent on the tech working (and being kept up). However, it’s also a little unsettling, as it means the driving skill of most drivers will atrophy even further as they become dependent on it. For me, it’s heartening to hear that perhaps we’re going to drop pursuit of the theoretical perfect in favor of the pragmatic good. Same thing, same story. The drivers actively got worse and worse. Every time I think ‘maybe I’ll get another bike’ I see 5 examples of why I stopped. Not, like, giving it a quick poke on its mount to select a different GPS route or skip through the ads in a podcast, no. He had his left hand on the steering wheel, his phone in his right hand, and was hunting head-down through the apps on his home screen. I never did ascertain whether it was a manual due to the heavily-tinted windows, but it was obvious that buddy was not paying any attention whatsoever to the road. I think we’re at least 25 years away from full autonomy, and that’s just for the flying cars (which I hear are just 10 years away ;)). One question I thought of last week on a 7hr+ trip back from Broken Hill, does any of the current semi-autonomous systems detect potholes or road kill and react? With the amount of freshly-killed kangaroo carcasses and balljoint-busting potholes I had to avoid on the mostly arrow-straight drive home, I can’t imagine how stinky or damaged my vehicle would have been after 700+ km if I let the computer drive and hit everything on the way through! [is this shit decaf, or what?] https://www.bloomberg.com/news/features/2022-10-06/even-after-100-billion-self-driving-cars-are-going-nowhere It will require dramatically more processing power (quantum computing?) and access to more information such as real-time weather data from automated weather stations placed along the highway every mile or so (maybe even more frequently), and most importantly, all of the cars on the highway must be communicating with each other to allow traffic issues and other potential problems to be predicted and passed on to all the other nearby vehicles. If all of the cars know what all of the other cars are doing/likely to do, it reduces the potential failure nodes in the whole system, which then reduces the the number of scenarios where the human needs to take over. None of that is remotely ready and won’t be for decades. What convinced me we wouldn’t have full autonomy anytime soon was a small wind vortex at the end of a long bridge over a deep gorge a few years back: I kinda/sorta saw a bit of dust swirling just in time to slightly swerve against it or I might have sideswiped the car next to me. As I was driving an 8400lb work van, it might not have worked out well for the car. Or me.
