We’ve all heard the story at some point in the past few years. The tech and auto industries are on the cusp of rolling out vehicles that can drive themselves and will forever transform the way we move. We’ll be able to summon driverless pods that will whisk us to our destinations, making personal vehicles unnecessary and freeing up all the space wasted on parking to be used for parks and public gathering spaces — or so they tell us.
Tech luminaries continually promise driverless vehicles are just a few years away, until that time passes and… they’re still a few years away. They’re definitely coming, they reassure us. But what if, counter to their claims, self-driving vehicles aren’t so close after all?
That’s not to say there won’t be some application of these technologies in the near future. New cars — at least the high-end models — are increasingly equipped with technologies to help with parking, changing lanes, slowing when you get close to another vehicle, automatically starting the wipers when it rains, and more. And transport trucks will likely see a high level of automation in the coming years, particularly for highway driving, with humans taking over to get to and from the highway.
However, when it comes to getting vehicles to drive themselves on regular roads, under all weather conditions, with people running in front of them, cyclists passing by, and other (human-driven) cars all around; well, that’s going to take a bit longer than technologists thought, or at least led us to believe.
The Six Levels of Autonomous Driving
Before looking at the current state of autonomous vehicles and the issues that still need to be overcome for a wide rollout, it’s important to understand the different levels of autonomy. There aren’t simply regular vehicles and autonomous vehicles, but a scale which goes from absolutely no self-driving tech to fully autonomous capabilities where there doesn’t even need to be a steering wheel because a human should never have to take control.
The vast majority of the vehicles on the road today are level 0, meaning they have no ability to control themselves; but some vehicles equipped with the features mentioned above—adaptive cruise control, park assist, lane keep assist—are level 1. They can control speed or steering in limited situations, but that’s the extent of their capabilities.
At level 2, the car is able to drive itself, handling steering and speed, but the driver must still have their hands on the wheel at all times because the system is not able to monitor its surroundings and can’t be fully trusted.
The next three levels are when the autonomous system really starts to take over and vehicles are equipped with a number of sensors to detect their surroundings; the three most common of which are LiDAR, cameras, and radar. Argo AI CEO Bryan Salesky wrote a brief explanation of the capabilities and limitations of each in a recent piece that we’ll return to later.
We use LiDAR sensors, which work well in poor lighting conditions, to grab the three-dimensional geometry of the world around the car, but LiDAR doesn’t provide color or texture, so we use cameras for that. Yet cameras are challenged in poor lighting, and tend to struggle to provide enough focus and resolution at all desired ranges of operation. In contrast, radar, while relatively low resolution, is able to directly detect the velocity of road users even at long distances.
Returning to the classifications, a level 3 vehicle is able to drive itself in designated areas and under limited conditions to the degree that the driver can stop paying attention to their surroundings, but they must still be ready to take over if the vehicle runs into a scenario it can’t navigate, at which point the system will alert the driver to resume control. The main difference between levels 3 and 4 is that, at level 4, the driver no longer has to be in the driver’s seat. If the vehicle runs into a complicated scenario and the driver doesn’t respond to the request to take control, the vehicle will be able to safely get itself off the road until it can get human assistance.
The driverless gold standard is level 5, which is what will be required if Silicon Valley’s transportation revolution has a chance of coming to fruition. Level 5 is the Google car pod without a steering wheel or the moving office where everyone faces each other. In order to be considered level 5, the autonomous system needs to be able to navigate every imaginable driving, road, and weather condition — which is precisely why it’s nowhere near ready. There are many hurdles — both hardware and software — that self-driving cars still need to jump before they can reach level 5 capabilities, and they’re not going to be as easy to work out as the technology’s boosters initially suggested.
Pushing Vehicles Beyond Their Capabilities
In November 2017, Las Vegas launched a self-driving shuttle operating on a 0.6-mile (one-kilometre) loop to demonstrate its embrace of new technologies. However, the shuttle didn’t last an hour before getting in an accident; and while the human driver of the other vehicle was deemed to be at fault, there have been a number of accidents and road violations by driverless vehicles which suggest that the systems haven’t yet reached the level where they can navigate complicated situations to avoid accidents in the same way that humans might.
That’s not to suggest that all companies are putting vehicles in unsafe situations. Indeed, most companies working on self-driving systems appear to be proceeding with caution. After taking a drive in one of GM’s self-driving vehicles in November, WIRED writer Aarian Marshall wrote that “it jolted, disconcertingly, to a stop at even the whisper of a collision,” concluding that “[i]f the Silicon Valley motto is ‘move fast and break things’, Detroit’s seems to be ‘move below the speed limit and ensure you don’t kill anyone’.” I imagine most people will be happy with Detroit’s cautious approach.
If Silicon Valley’s motto is “move fast and break things,” Detroit’s is “move below the speed limit and ensure you don’t kill anyone”
However, that doesn’t mean they’re all putting safety first. Uber and Tesla, in particular, seem to be pushing the boundaries of safety to get their vehicles on the road as quickly as possible. In Tempe, Arizona, a driverless Uber was hit when going through a yellow light, and in Pittsburgh they’ve been in multiple accidents and have been filmed driving the wrong way on city roads. During their time in San Francisco, the vehicles seemed to completely ignore bike lanes, creating a safety risk for cyclists, and ran a number of red lights. Uber initially claimed the issue with red lights in San Francisco was the result of human error, not the autonomous tech, but documents obtained by The New York Times suggest the company wasn’t telling the truth.
Elon Musk has a record of overpromising and underdelivering, and that’s no different when it comes to driverless vehicles. He promised that Tesla’s Autopilot feature would deliver “full autonomy,” even though engineers “did not believe the system was ready to safely control a car” — some of them left the company as a result. Musk’s ignorance of safety is particularly shocking given that, in May 2016, a Tesla Model S using Autopilot was the first autonomous vehicle to be involved in a fatal accident, killing the driver. The National Transportation Safety Board found that the system’s limitations played a role in the accident because it allowed the driver to use it in areas it was not designed for and to keep his hands off the wheel for extended periods of time even though it was only rated for level 2 functionality.
Musk has continually failed to meet his self-imposed timelines for Tesla’s Enhanced Autopilot and Full Self Driving packages, which cost $5,000 and $8,000, respectively, and customers have launched a class-action lawsuit for being misled into buying features that don’t exist. A planned cross-country drive by an autonomous Tesla due to happen by December 2017 has also been indefinitely delayed. GM’s director of autonomous vehicles, Scott Miller, doesn’t believe Tesla vehicles are equipped with sufficient sensors to allow full self-driving capabilities since Musk has rejected LiDAR and only equipped vehicles with cameras and radar. Musk obviously rejects this assertion, but he has admitted that the vehicles may need more computing power.
Elon Musk says self-driving cars will be ready in two years — the same thing he said two years ago
In December 2017, Musk updated his timeline for the arrival of fully self-driving vehicles to two years, and said they’ll be significantly better than humans in three years. Of course, two years ago he also said they’d be ready in two years, so maybe take that prediction with the usual caveat that comes with a Musk timeline: it’s complete bullshit.
The Challenges Facing Driverless Technology
Tesla and Uber aren’t the only companies finding their initial timelines for driverless vehicles weren’t entirely accurate, as Marshall recently chronicled missed deadlines and delays across the industry. Google said we’d all have access by 2017, but that didn’t happen. Ford announced a 2021 deployment, which its new CEO is now also suggesting won’t happen. And Volvo promised to distribute self-driving SUVs to 100 Swedish families in 2017, but now says only 100 people will get them in 2021 and they may only have semi-autonomous features.
GM is seemingly the only traditional automaker continuing to make bold claims — promising an autonomous vehicle without a steering wheel by 2019 at CES — but it remains to seen if the company can really meet its timeline or if it’s just making another overambitious announcement like so many before it. GM’s self-driving vehicles were involved in six accidents in September 2017 alone, so there’s good reason to be skeptical.
In order to achieve level 5 capabilities, self-driving vehicles need to be ready for absolutely anything. Marshall quoted Nutonomy CEO Karl Iagnemma as saying the first 99 percent of autonomous capabilities were “a walk in the park” compared to the final 1 percent — though it seems unlikely the technology that far along.
Describing the state of the sensors on autonomous vehicles, Salesky explained that “[i]ndividual sensors don’t fully reproduce what they capture, so the computer has to combine the inputs from multiple sensors, then sort out the errors and inconsistencies”; and, on top of all that, “significant work” remains to be done “to lower costs, reduce sensor count, and improve range and resolution.” While more companies are getting into the space, the price of equipping a vehicle with LiDAR sensors is still in the five-figure range, according to SFChronicle.
Backing up the earlier suggestion that autonomous vehicles may not be able to handle the situations that some companies are currently putting them in, Salesky also acknowledged that attentive drivers are good at “subconsciously estimating the next few seconds of behavior from other road users” and anticipating what they might do; tasks which algorithms cannot yet perform to nearly the same level.
The autonomous systems also struggle in a variety of driving situations. In busy urban cores, there are a lot of people, vehicles, and sometimes bikes that need to be tracked, and the algorithms simply haven’t reached the level where they’re capabale of handling it, as GM, Uber, and other companies have demonstrated. They also struggle in bad weather conditions. Anyone who’s ever driven in the winter knows how dirty vehicles can get, which makes cameras particularly unreliable, and both rain and snow can mess with the LiDAR lasers. Plus, if there’s snow on the road, extra attention needs to paid to other drivers and to one’s own driving, and autonomous systems are having more trouble being trained in such conditions.
Drivers had to take over from Uber’s self-driving car ten times for every eight miles
Data on the autonomous technologies of various companies also shows that while some are doing better than others, progress may be stalling. Data obtained by Recode in March 2017 showed that a driver had to take over from Uber’s autonomous system “an average of 10 times for every eight miles driven,” and the distance between “bad experiences” — when the car did something it wasn’t supposed to — was actually getting worse. Uber’s tech was judged to be by far the worst in the industry, while Google’s Waymo had the best — a fact Uber has even admitted.
While Waymo does seem to be excelling, new data about its system and its pilot service in the suburbs of Phoenix are illustrative of the current state of self-driving technology.
Leading Tech is Still Far From Perfect
In a breakdown of the data on Waymo’s driverless vehicles at the end of 2017, Tasha Keeney showed that while the company tripled the distance between human interventions from 2015 to 2016, it’s currently stalled at an average of around 5,000 miles (8,050 kilometres). While Keeney acknowledged that the lack of progress may be a result of testing in more difficult situations, she ultimately concluded that “autonomous taxis will not be ready for prime time until the early 2020’s, if then.”
If any company can deliver on such a timeline, it seems that Waymo is best positioned, but even then a lot of progress will have to be made over the next few years. Waymo began running a driverless taxi service in Phoenix, Arizona near the end of 2017 and its limitations are instructive of the state of the technology.
The vehicles involved in the service are only level 4, meaning they’ll be fine in most situations, and if they run into anything they can’t handle they’ll be able to safely pull themselves over since there are no human operators. The service doesn’t cover the whole of the Phoenix metropolitan area, but is limited to the suburb of Chandler, where the vehicles won’t need to worry too much about humans on foot, nor should the roads be too narrow or congested with other vehicles. The weather in Chandler is clear, sunny, and dry — perfect for the current state of autonomous vehicles.
Waymo recently began testing its vehicles in Michigan to give them more experience in challenging conditions, but it’s unlikely they could safely carry passengers there and the company won’t be rolling a driverless taxi service in a Detroit suburb anytime soon. Waymo may be ahead of other companies, but it still has a long way to go.
It can be exciting to indulge the visions of Silicon Valley CEOs, but it’s also important to remain realistic about what they’re able to deliver, especially given their habit of getting ahead of themselves. Driverless vehicle technology is progressing, but not at the speed we were initially led to believe. A lot of the issues that remain to be worked out aren’t so simple, and it will likely take time to develop the capabilities necessary for the vehicles to drive in more difficult conditions.
As companies continue to work out these aspects of the technology, the hype about driverless vehicles may wane, at least for a few years; and while driverless trucks may become more common on highways and autonomous shuttles may find their way into retirement communities, it’s highly unlikely that self-driving cars will become a feature of urban life anytime soon.