There are two different visions of the future of self-driving cars. Both visions agree that a smart car needs to process a massive amount of information in order to make real-time decisions.
One vision is that smart cars will be really smart and will include a lot of edge computing power and AI that will enable a car to make local decisions as the car navigates through traffic. Cars will likely to able to communicate with neighboring cars to coordinate vehicle spacing and stopping during emergencies. This vision requires only minimal demands for external broadband, except for perhaps to periodically update maps and to communicate with things like smart traffic lights.
The other vision of the future is that smart cars will beam massive amounts of data to and from the cloud that includes LiDAR imagery and GPS location information. Big data centers will then coordinate between vehicles. This second vision would require a massively robust broadband network everywhere.
I am surprised by the number of people who foresee the second version, with massive amounts of data transferred to and from the cloud. Here are just some of the reasons why this scenario is hard to imagine coming to fruition:
Volume of Data. The amount of data that would need to be transferred to the cloud is massive. It’s not hard to foresee a car needing to transmit terabytes of data during a trip if all of the decisions are made are made in a data center. Most prognosticators predict 5G as the technology that would support this network. One thing that seems to be ignored in these predictions is that almost no part of our current broadband infrastructure is able to handle this kind of data flow. We wouldn’t only need a massive 5G deployment, but almost every part of the existing fiber backbone network, down to the local level, would need to also be upgraded. It’s easy to fall into the trap that fiber can handle massive amounts of data, but the current electronics are not sized for this kind of data volumes.
Latency. Self-driving cars need to make instantaneous decisions and any delays of data going to and from the cloud will add delays. It’s hard to imagine any external network that can be as fast as a smart car making its own local driving decisions.
Migration Path. Even if the cloud is the ultimate network design, how do you get from here to there? We already have smart cars and they make decisions on-board. As that technology improves it doesn’t make sense that we would still pursue a cloud-based solution unless that solution is superior enough to justify the cost of migrating to the cloud.
Who will Build? Who is going to pay for the needed infrastructure? This means a 5G network built along every road. It means fiber built everywhere to support that network, including a massive beefing up of bandwidth on all existing fiber networks? Even the biggest ISPs don’t have both the financial wherewithal and the desire to tackle this kind of investment.
Who will Pay? And how is this going to get paid for? It’s easy to understand why cellular companies tout this vision as the future since they would be the obvious beneficiary of the revenues from such a network. But is the average family going to be willing to tack on an expensive broadband subscription for every car in the family? And does this mean that those who can’t afford a smart-car broadband connection won’t be able to drive? That’s a whole new definition of a digital divide.
Outages. We are never going to have a network that is redundant down to the street level. So what happens to traffic during inevitable fiber cuts or electronics failures?
Security. It seems sending live traffic data to the cloud creates the most opportunity for hacking to create chaos. The difficulty of hacking a self-contained smart car makes on-board computing sound far safer.
Who Runs the Smart-car Function? What companies actually manage this monstrous network? I’m not very enthused about the idea of having car companies operate the IT functions in a smart-car network. But this sounds like such a lucrative function I can’t foresee them handing this off to somebody else? There are also likely to be many network players involved and getting them all to perfectly coordinate sounds like a massively complex task.
What About Rural America? Already today we can’t figure out how to finance broadband in rural America. Getting broadband along every rural road is going to be equally as expensive as getting it to rural homes. Does this imply a smart-car network that only works in urban areas?
I fully understand why some in the industry are pushing this vision. This makes a lot of money for the wireless carriers and the vendors who support them. But the above list of concerns make it hard for me to picture the cloud vision. Doing this with on-board computers costs only a fraction of the cost of the big-network solution, and my gut says that dollars will drive the decision.
It’s also worth noting that we already have a similar example of this same kind of decision. The whole smart-city effort is now migrating to smart edge devices rather than exchanging massive data with the cloud. As an example, the latest technology for smart traffic control places smart processors at each intersection rather than sending full-time video to the cloud for processing. The electronics at a smart intersection will only communicate with the hub when it has something to report, like an accident or a car that has run a red light. That requires far less data, meaning far less demand for broadband than sending everything to the cloud. It’s hard to think that smart-cars – which will be the biggest source of raw data yet imagined – would not follow this same trend towards smart edge devices.
I was at a conference last week and I talked to three different people who believe that driverless cars are going to need extremely fast broadband connections. They cite industry experts who say that the average car is going to require terabytes per day of downloaded data to be functional and that only extremely fast 5G networks are going to be able to satisfy that need. These folks talk about needing high-bandwidth and very low latency wireless networks that can tell a car when to stop when encountering an obstacle. This vision sees cars as somewhat dumb appliances with a lot of the brains in the cloud. I would guess that wireless companies are hoping for this future.\
But I also have been reading about experts that instead think that cars will become rolling data centers with a huge amount of computing capacity on board. Certainly vehicles will need to communicate with the outside world, but in this vision a self-driving car only needs updates on things like the current location and for road conditions and traffic problems ahead – but not the masses of data anticipated by the first future vision cited above.
For a number of reasons I think the second vision is a lot more likely.
Self-driving cars are almost here now and that means any needed network to support them would need to be in place in the near future. That’s not realistically going to happen. Most projections say that a robust 5G technology is at least a decade away. There are a dozen companies investing huge sums on self-driving car technologies and they are not going to wait that long to even investigate if controlling cars from external sources makes sense. Every company looking into self-driving technology is operating under the assumption that the brains and sensing must be in the cars – and they are the ones that will drive the development and implementation of the new car technology. It’s not practical to think that the car industry can wait for deployment of the needed networks that are not under their control or reasonably available.
Who’s going to make the huge investments needed to build the network necessary to support self-driving cars? The ability to deliver terabytes of data to each car would require much faster data connections than can be delivered using the normal cellular frequencies. Consider how many fast simultaneous data connections would be needed to support all of the cars on a busy multilane highways in a major city. It’s an engineering challenge that would probably require using high frequencies. And that means putting lots of cell sites close to roads – and those cell sites will have to be largely fed by fiber to keep the latency low (wireless backhaul would add significant latency). Such a network nationwide would have to cost hundreds of billions of dollars between the widespread fiber and the huge number of mini-cell sites. I can’t picture who would agree to build such a network. The total annual capital budget for all of the wireless companies combined today is only in the low tens of billion range.
Even if somebody was to build the expensive networks who is going to pay for it? It seems to me like every car would need an expensive monthly broadband subscription, adding significantly to the cost of owning and driving a car. Most households are not going to want a car that comes with the need for an additional $100 – $200 monthly broadband subscription. But my back-of-the envelope tells me that the fees would have to be that large to compensate for such an extensive network that was built mostly to support self-driving cars.
The requirement for huge numbers of cars to download terabytes of data per day is a daunting challenge. The vast majority of the country today doesn’t even have a landline based broadband connection capable of doing that.
There are also practical reasons not to put the brains of a car in the cloud. What happens when there are power outages or cellular outages. I don’t care how well we plan – outages happen. I’d be worried about driving in a car if there was even just a temporary glitch in the network.
There are also issues of physics if this network requires any connections to be made by millimeter wave spectrum, or even spectrum that is just a little lower on the frequency scale. There is a huge engineering challenge to get such signals to track a moving vehicle reliably in real-time. Higher frequencies start having doppler shifts even at walking speeds. Compound this with the requirement to always have true line-of-sight and also the issue of connecting with many cars at the same time on crowded roads. I have learned to never say that something isn’t possible, but this presents some major engineering challenges that are going to take a long time to make work – maybe decades, and maybe never.
Finally are all of the issues having to do with security. I’m personally more worried about cars being hacked if they are getting most of their communications from the cloud. If cars are instead only getting location and other basic information from the outside it would be a lot easier to wall of the communications stream from the operating computing process, and reduce the chances of hacking. It also seems like a risk if cars get most of their brains from the cloud for a terrorist or mischief-maker to disrupt traffic by taking out small cell sites. There would be no way to ever make such devices physically secure.
I certainly can’t say that we’ll never have a time when self-driving cars are directed by a large outdoor cloud, as often envisioned in science fiction movies. But for now the industry is developing cars that are largely self-contained data centers and that fact alone may dictate the future path of the industry. The wireless carriers see a lot of potential revenue from self-driving cars, but I can’t imagine that the car industry is going to wait for them to develop the needed infrastructure.