One Weird Thing Everyone Gets Wrong About Autonomous Cars

What do your phone and computer have in common with driverless cars?

Your phone, your computer, and your driverless car all need networks to provide them with services. Your phone and computer need networks to unlock their value, and your driverless car does as well.

I’ve written before to talk about how roads are an untapped goldmine, and how the current state of the art for roadbuilding systemically precludes improvement. But I haven’t really said much about driverless cars except they aren’t to be feared.

Driverless cars will be to mobile IOT what smartphones were to mobile web & apps. The economics that driverless cars unlock will be on the scale of the economics the Internet unlocked.

But there is one weird flaw in the way most people think about autonomous cars.

Most people conceptualize driverless cars as being independent, in the same way that a person driving a car around is independent. We talk about “smart” vehicles that can “see” obstacles and “drive themselves”. But those are not accurate representations, just anthropomorphic simplifications.

If your phone loses service while driving, you can keep driving. But if your autonomous car loses service, how long can it keep driving? Moving the burden of driving from the person to a machine creates system requirements that a human doesn’t have.

There are two models of autonomous cars currently popular, both stemming from the Carnegie Mellon model that reached common popularity in the mid-2000’s thanks to DARPA’s Grand Challenge in 2004. (We’ll talk more about the influence of the DARPA Grand Challenge on the trajectory of driverless in a later post.)

These two popular models I’ll call “urban” and “highway”. Most autonomous car developers are using a mixture of the two.

The urban model is to create a large database of images of the local environment, and manually (or semi-manually) map these images into machine-readable localization maps. The autonomous car has a multi-sensor component that creates a real-time machine representation of the surrounding world. The vehicle’s GPS system is error-corrected and used to call the existing machine model from the database. The real-time model is compared to the database model to provide inputs to the autonomous car to improve its localization. These features are intended to be used in complex urban environments whose visual appearance changes regularly, and have location-specific quirks.

Over the last few years, Google’s autonomous vehicle program, focused on the urban model, has logged millions of vehicle miles, set new records in the capabilities of autonomous vehicles, demonstrated just how much safer driverless is, and… can go 25 mph in a tiny patch of California where the weather is perfect all the time and Google provides the wifi.

These golf carts cost more than ten times an average automobile. Ferrari pricing, and I mean nice Ferrari. Don’t ask about snow performance. Or highways. Or a brake pedal. But don’t worry, there’s a button you can slap if you get scared.

The highway model focuses primarily on identifying lane-markings to determine road position and a multi-sensor component to detect obstacles. GPS is used for navigation but there is little-to-no intricate pre-mapping and real-time comparison as in the urban system. Highway driving assist features are intended to be used in controlled highway environments.

Tesla’s autopilot mode was the coolest thing anyone’s seen yet, and appears to be the most versatile implementation of autonomous that’s on the market. Yet it can’t see a trailer sitting across the road. Elon has lamented the need for better road markings. Latex with glitter in it is the solution? That’s like writing a check at the grocery store. Get a debit card, grandma!

Seriously though, road markings get snow on them. They get weird glares and reflections when it rains. Leaves. Dead animals, popped tires, cardboard boxes. Snowplow blades and hot tires can peel them off, and they just plain old get worn out. There’s no road markings when the road is being refinished. Markings can be milled off, moved, reapplied, and sometimes during big projects, several conflicting marks exist on the pavement at the same time, where the only cue to tell which is the right one to use is based on the apparent age of the marks. Some roads don’t ever get lane markings. Lane markings sometimes just kind of… end. We can certainly do better than lane markings.

Both models for autonomous rely on the vehicle generating real-time identification and notification of environmental hazards. Both models provide real-time identification of presence and positional information about nearby vehicles. The real difference is in whether or not you pre-map to improve your knowledge about your vehicle location and its surrounding environment. Having the localized maps improves your ability to operate in extremely complex urban environments, but limits the scalability of your application and increases your back-end labor to deliver driverless in your coverage area.

Both of these models also have fundamental limitations that preclude them from accomplishing level 4 autonomy, and level 4 autonomy is like, super important for us as a society. I define a level 4 autonomous vehicle as one you can put a drunk, a baby, a grandmother, or a drunk baby grandmother into, and have it arrive at their destination safely at five-nines or better.

(Interesting side note: Drivers on average have an “incident” ever 1m miles or so, which means drivers on average operate with 5–9’s uptime currently. As long as we can meet current performance in an autonomous model, we’re fine. Exceeding current performance is great, but not necessary.)

Level 4 autonomous will drop the cost of transportation by an order of magnitude, enable enormous lifestyle improvements, and provide transportation to the poor, invalid, elderly, disabled, addicted, and other people who for one reason or another are unable to provide their own transportation currently.

As nice as it would be to extend mobility to these groups, it is wildly irresponsible try to do so using a model that might require them to provide inputs to the vehicle (which they aren’t capable of doing or legally allowed to do), or where the vehicle might simply stop operating because it lost its signal, leaving the occupants stranded and helpless in what could be a life-threatening situation for them.

Does anyone want kids to die of heat stroke because their driverless car lost its signal on the way to school on a hot day? I’m guessing probably not.

The problem is, neither popular model implements all of the services required to support a fully functional level 4 autonomous vehicle because a level 4 autonomous vehicle has network dependencies that the urban and highway models do not account for.

(In many ways, relating back to the DARPA Grand Challenge, the current models are workaround kludges that assume certain network dependencies simply won’t be available. But like I said, that’s another post.)

Driverless cars need networks just the same as phones and computers. Devices that run apps that provide services require networks.

The urban model’s need for a network is obvious — it has to have a data connection to query the machine representation database to compare to its own sensor fusion view based on the measured GPS location.

The highway model is designed to avoid the network requirement, but in doing so limits the scope of its application. But Tesla does rely on a network connection to update the vehicle’s maps database for navigation, even if the local environment scanning doesn’t rely on a manually-updated pre-mapping database like Google does.

Meanwhile, electric vehicles also need a power network for charging, whereas a gas vehicle requires a fuel station network, and all kinds of rubber tire vehicles need a road network to drive on.

Vehicles require networks, and the more advanced the vehicle, the more advanced the network it requires to support its operation.

According to the upcoming USDOT standard expected for publication in 2020, new vehicles will also require DSRC equipment to support Vehicle-to-Vehicle, Vehicle-to-Infrastructure, and other Vehicle-to-Anything services. This is only the tip of the iceberg.

These are some of the services that level 4 autonomous vehicles will require while it is mobile:

• A navigable road (lest we forget)
• >1Gbps wireless uplink @ 99.999% uptime & the low end-to-end latency
• GPS localization improvement to <1' via differential GPS base stations
• DSRC messaging
• Mapping database for trip navigation
• Machine-translated representations of the environment
• Real-time identification and avoidance of environmental hazards

Right now, nearly everyone has access to navigable roads and mapping databases. The roads may need service but they are definitely there and currently support 30k users per mile per day on average. Most of us have mapping on our phones or on our dash nav systems.

The urban model creates machine representations of the environment and compares them to a database, while the highway model often skips the comparison step. Both models rely on the vehicle generating real-time identification and avoidance of environmental hazards, including nearby vehicles.

However, only a few areas have extremely high speed wireless, and only a few have GPS localization. DSRC messaging is limited to pilot implementations in select cities.

Groups developing autonomous vehicles face network and service availability limitations for testing their implementations. If your vehicle requires an extremely high speed, low latency, high uptime wireless connection, and a GPS base station, what percentage of the US road network has those services available? The answer is… basically none. Where are you going to develop and test your vehicle?

The USDOT published an RFP for autonomous vehicle proving grounds to attempt to address this problem, but with six weeks to respond, I’m not sure that many even took note of the RFP while it was live. If you read the RFP, there was no actual technical criteria to qualify as a “proving ground”, so it’s also unclear if the solicitation actually resulted in any responses that could provide for the needs identified in the last paragraph.

To the USDOT’s point, there are autonomous test facilities that could provide these network services to further the development of autonomous vehicles. This is a blessing for many OEMs, who have extremely limited geophysical area to test their vehicles in, and extremely limited environments for their vehicles to experience. But a handful of controlled test locations with artificial environments providing non-standard network services does not result in a widespread deployment of autonomous vehicles.

Bill Ford declared in February 2016 that Ford needs a “real world, real weather, real traffic” test environment at the Kinetic forum in Kansas City. It’s safe to say that all the companies developing level 4 autonomous vehicles need the same. But the need goes beyond testing, autonomous vehicles need these network services for deployment too.

On top of the lack of network support for autonomous services across the American roadway infrastructure network, we have roads that are literally falling apart. For both physical (roads & power) and data network services, the lower Quality of Service a network can provide, the more service level management has to be offset to the device. This increases the bulk of the device and its cost for a comparable service quality versus improved network QOS support.

Automotive OEMs know that if roads are in bad shape, cars need larger, more expensive suspension and cabin isolation systems. This increases cost and displaces space that can be used for some other feature. In the same way, the more services the autonomous vehicle must provide itself, the bulkier the system will be and the costlier the vehicle.

The road and data networks that support our existing cars are in extreme distress, and the networks that we need to reach level 4 autonomy don’t even exist yet. Current wireless services are not the solution, and existing roads limit the autonomous vehicle’s capabilities while increasing their costs.

Musk has described a Tesla as “a smartphone you sit in”. IBM says driverless cars are “data centers on wheels”. The current car has two to four dozen computers in it. Suppliers for OEMs want to be able to pull data wirelessly to get real-time performance data for their components, but how? Are you going to pay a cellular bill for your car, which mostly benefits the OEM suppliers? No. Or are they going to pay for a cell connection for every vehicle their components are in? No.

Imagine if you had a data center with two to four dozen computers at your home, all trying to connect over a 4G link. It’s outrageous. And yet, for some reason people assume connectivity will somehow take care of itself when we talk about autonomous cars.

Nobody assumes a network is unnecessary in any data-intensive other market, yet in driverless, it’s taken for granted that nobody really needs a network to deliver Level 4. If they admit they need a network, it’s a sideways admission that minimizes the need, “oh, sure, I guess we do use a network, but what’s available is fine.” That might work now (barely), where there are zero users of autonomous, but I promise, that won’t work when significant numbers of connected and autonomous vehicles start to come online.

Finally, let me ask you this: What will you do when your car is driving you? Twiddle your thumbs and stare out the window? Unlikely. You’ll want to watch Netflix, use Skype, play AR/VR games, and all kinds of other bandwidth-intensive applications.

This simple question, “what will you do”, is by itself more than enough to demonstrate that autonomous cars need their own network.

The assumption that driverless cars don’t need a network doesn’t hold, and it’s outright silly to think relying on existing networks is a reasonable situation!

If we’re taking autonomous vehicles seriously, when is it finally going to be time to start talking about the network these vehicles need?

Are you ready to talk about it? I know I am!