Artificial intelligence is on everybody’s lips these days.Vehicle automation, the merging of AI and automotive is the fodder of ginormous tech events like SXSW, Web Summit, Collision and CES. It’s hard not to get dazzled by the endorphin driven vibes and almost deterministic technological anticipations encircling these places. A nice contrast are smaller, more industry-specific conferences because you typically get the latest of the greatest but minus the hype. Last week I got to attend Technology Future’s communications conference here in Austin for the third year in a row. A technology valuation conference with experts from telecommunications and industries working at the bleeding edge of the Internet of Things, smart city development and vehicle automation, I learned about the current status of enabling technologies such as spectrum policies, 5th generation wireless systems (5G) and other crucial issues.
Professionals doing the nitty-gritty of implementing these technologies tend to have a no-nonsense approach to these issues. They’re knowledgeable and imaginative to envision interesting scenarios, yet while keeping an eye on plausibility. Several of the TFI presenters cautioned against the notion that AI is about to take over our jobs and that we will see a mass market of self-driving cars on the roads anytime soon.
I’ll leave the verdict to the experts, but will reflect on some of the bottlenecks I think lie ahead. More than anything, I think the market for fully autonomous vehicles depends on popular opinion, regulatory institutions and future market demands. Because ultimately, it’s not the technology itself, but our willingness to make compromises with it that determines our path forward.
1.Learning from events that have not yet happened:
A convolutional neural network misclassifying a cat as a dog is one thing. 4000 lbs of metal rolling down the highway at 100 mph while “learning on the job” is a real liability.
SAE International defines six levels of automation for automakers (actually zero automation plus 5 levels). It’s mainly the promises of the third and fourth level that gets people excited for the future. Semi-autonomous level 3 cars are on the road today, and they can drive themselves. Except for when they can’t. Like when, on a windy day, a piece of plastic flies in front of the car that the deep learning system has never seen before and now has to figure out on the fly that it doesn’t need to come to a screeching halt and cause a mass pile-up. Well ok, so let’s say the neural network was trained to recognize flying plastic, but might miss that cigarette butt someone threw out the window. Or a sock. The point is, computer vision runs on convolutional neural networks (CNNs), and they are typically data hungry. An image can say more than a thousand words to us humans, but to a computer, an image is just differently weighted pixels of varying colors and shades. AI does not have the instincts to react efficiently in unknown situations and needs to base its decision on statistical occurrences of real or simulated events already learned, so it needs lots of interaction with this data before it can classify correctly on its own.
At some point the unsupervised learning process will be developed enough for the car to know what to do in a split second. Car manufacturers are putting Level 4 in pilot projects already, and this will help it pick up and process obscure pieces of data. It’s even possible that we might have a Level 4 car on the road next year “doing narrowly defined fleet applications” (whatever “narrowly defined” means) if GM gets its way with the transportation authorities. Bottom line, I think the long-tail problem with statistical outliers required for 4th and 5th level automation could put the end goal a bit further out than many think. At least for most practical purposes. A convolutional neural network misclassifying a cat as a dog is one thing. 4000 lbs of metal rolling down the highway at 100 mph while “learning on the job” is a real liability. Yet, I suspect we might see limited versions of level 4 and 5 robo-taxis at lower speed limits on predictable routes.
2. Network connectivity and Latency:
In a panel, founder and president of iGR Iain Gillott suggested that the intermittence that might occur with the LTE network is akin to having a surgeon decide to take a long break in the midst of a complicated surgery.
Most engineers working with vehicle automation would probably claim that a car’s safety should not rely on the data is receives from its communications network. Sensors, cameras, LIDAR, software and edge computing capabilities is necessary to be fully “automonous”. It should not have to depend on a wireless network that can crash, slow down or be compromised. Nonetheless, most discussions on vehicle automation address connectivity between cars and cars to infrastructure. The federal government is an active stakeholder in this vehicle-to-vehicle (V2V) discussion. Helping deploy vehicle-to-vehicle infrastructure, a U.S. Department of Transportation proposal mandates that car manufacturers have to install technology to make such communication possible (there are rumors this might falter under Trump). For two decades the FCC has made available dedicated short range communication spectrum (DSRC) to ensure low latency communication between vehicles. This regulation is not without contention, though. The telecommunications industry is pushing for FCC to release the spectrum waves and claims the DSRC system will be obsolete when 5G rolls out.
The questions remain: when will 5G be ubiquitous, what will it cost and will it be distributed evenly so that self-driving vehicles won’t run into uncovered areas? High speed broadband is already severely unevenly distributed and rural areas could become the losers.
The TFI speakers spoke about the various hurdles of rolling out 5G. Like, the cost (over a trillion dollars if I remember correctly). Or the number of cell towers that would be needed. There was mention that 5G would need to use both optical fiber and radio spectrum. Installing fiber and cell towers in sparsely populated areas is expensive, so there is very little commercial incentive to do so. How do we overcome that? LTE apparently isn’t the most reliable solution either. In a panel, founder and president of iGR Iain Gillott suggested that the intermittence that might occur with the LTE network is akin to having a surgeon decide to take a long break in the midst of a complicated surgery. Not ideal for any situation where connectivity is critical!
Might we realistically depend on a combination of many solutions for providing a network that serves both a dramatic increase in dataflow and the latency needed? This article quotes Brian Daugherty, CTO of the Motor and Equipment Manufacturers Association (MEMA): “You are going to see a knitting together of these technologies (DSRC and 5G), so that the low-latency things – the collision avoidance – will be handled by V2V and the active sensors on the vehicle. Slightly longer range things, like traffic congestion, other things, will be handled by the LTE and 5G systems of the future…..The DSRC systems will probably use LTE and 5G to do updates to the systems, potentially for security certificate management and updates. I think we will see a lot of synergies between the technologies.”
“After talking with most of the major autonomous car makers, including Tesla, BMW, GM, and Toyota, I realized that decision makers in large automotive companies don’t have a magic solution any more than startup software developers do.” ~Zach Aysan
A 5G driven V2V infrastructure could leave you vulnerable to data theft and cybercrimes. You might WannaCry or have a Meltdown if your computer gets hacked and your private information stolen or ransomed. A hacked car, however, is also a lethal weapon. Data scientist and cyber security expert Zach Aysan writes in his piece for the Weekly Standard, Terrorists Could Use Teslas to Kill Us: “After talking with most of the major autonomous car makers, including Tesla, BMW, GM, and Toyota, I realized that decision makers in large automotive companies don’t have a magic solution any more than startup software developers do.”
Suggestions to improve security by using blockchain for the data exchange between connected vehicles/infrastructure are interesting. Blockchain is the verification technology that makes cryptocurrency like Bitcoin possible, but it can be used in other applications where verification and strong encryption is needed. In fact, last week I received an email offering me to post the infographic you can see at the bottom. Blockchain secured driver data is an interesting idea, and traffic security is definitely one of the better use cases I have heard for blockchain technology. But experts are skeptical. Is it feasible to expect that self-driving cars everywhere be built around blockchain technology? And what about network connectivity and latency issues? Mining blocks for other cryptographic purposes than currency might require less computing power than Bicoin since the units of transaction can represent anything, e.g. distances driven by a car. But its dependence on connectivity and the data flows that require split second judgements seems to raise some questions. One benefit could be that your mobility remains private. Which brings me to the next point.
4. The market and young people’s adoption
Several recent surveys indicate that Generation Z is in fact much more excited about driving than we often give them credit for.
I’m finally back in my own camp here. Here’s the most important contingency: You can have all the automated technology in the world, but if people don’t want it, it’s not going to make waves. And people do care about privacy issues, which is eliminated with self-driving technology.
I do actually think we are all warming up to autonomous driving experiences. The benefits are just too many. At least until something awful happens. And Millennials and Generation Z do indeed seem more positive to self-driving cars than older generations. (I bet this number might even be higher when age-related ailments impair freedom-loving Boomers from keeping their drivers licenses). However, we shouldn’t push assumptions around Gen Z’s driving aversiveness too far. Several recent surveys indicate that Generation Z is in fact much more excited about driving than we often give them credit for. ERIE Insurance recently did a survey which found 89 percent of 14- to 17-year-olds who don’t yet have a driver’s license plan to get one, and of these, 96 percent are excited about learning to drive.
It’s hard to know for sure, but just because they won’t to trade cell phones for cars, it doesn’t mean they ditch cars. They want both! Growing up with Teslas all around, they might associate cars with sophisticated computing rather than obsolete, gas-guzzling technologies that gave cars such a bad rap with Millennials.
In conclusion, are we discussing a false dichotomy?
Knowing that we already have self-driving (autopilot) cars on the road, perhaps it’s rather a matter of gradual change where full automation remains an elusive, but perhaps not that important end goal? Perhaps instead of making road trips in a fully automated vehicle, you’ll enjoy having a car that not only self-parks, but self-valets farther away so you don’t have to look for parking or walk through the lot in the rain? Perhaps you can order a local self-driven shuttle within the perimeters of pre-qualified locations? And when driving on the highway, perhaps you’ll remain behind the steering wheel only to intervene if a Moravec’s paradox situation strikes, but still have freedom to entertain your screen addiction during less eventful stretches? What do you think?
ProvidedAuto Accessories Garage