What Hurdles Do Self-Driving Cars Face As Waymo Gets Ready For Prime Time?
As self-driving cars edge closer to deployment for the general public, a spate of stories from The Information, Motor Trend, and Forbeshave pointed out specific difficulties that these vehicles are experiencing. Waymo receives the most focus since it has the largest fleet of self-driving vehicles, currently 600 and growing, the most miles driven (9 million) and the soonest launch date (2018).
Some of these issues are truly difficult, whereas others seem more surmountable.
Motor Trend highlights the trolley problem as an issue automotive companies must confront.
Will a computer-controlled car willingly sacrifice itself (and its driver) if it is presented with this impossible situation: to run over children mingling around their stalled bus in the middle of a blind corner, or to drive off the adjacent cliff? No executive has yet properly answered that question, saying it’s a theoretical situation. But that is the sort of theory that must be embraced and programmed into every car—and the autonomous cars’ occupants must be willing to accept the uncomfortable outcome generated by a software coder.”
This actually seems incorrect. Millions of human drivers have been licensed to operate motor vehicles without ever receiving instruction on how to approach the trolley problem. This is not an issue standing between self-driving cars and deployment.
The Information quotes an anonymous source reporting that Waymo vehicles have trouble separating individual people in a group. According to the report, Waymo minivans see groups as “one unidentifiable mass.”
It’s hard to believe Waymo’s perception system is quite this limited. Deep neural networks have gotten really good at “instance segmentation.”
It’s possible that problems arise in specific circumstances, but Waymo has an entire facility dedicated to testing special circumstances over and over until they get it right.
If there is a problem, it’s probably with prediction, as the anonymous source states. “They have a hard time predicting where each individual will go.”
Uber’s accident in March in Tempe, Arizona that resulted in the death of Elaine Herzberg stemmed in large part from the failure of Uber’s autonomous driving system to accurately predict her path. Prediction is a hard and critical problem.
Motor Trend tags infrastructure quality as a necessity for self-driving cars. To some extent, this may not actually be such an important factor. Self-driving cars are designed to work in the world as it exists, not as we wish it might be. Just as humans successfully navigate dilapidated infrastructure, so will autonomous vehicles.
That said, better infrastructure is always beneficial. It’s not an accident that autonomous vehicles are being tested in Phoenix, Arizona, a state with some of the best roads in America.
Self-driving car companies will be able to choose where they operate for the foreseeable future. If a state wants to attract autonomous vehicles, paving and painting the roads is a good start.
The Information flags left turns, especially “unprotected” left turns (where there is only a green light, not a green arrow), as a significant problem for Waymo vehicles.
Unprotected left turns are dangerous maneuvers even for human drivers, so it’s not surprising that these challenge self-driving cars.
Just like with human drivers, however, this seems like a practice-makes-perfect situation. The more unprotected left turns that vehicles make, the better their algorithms will become. Waymo has already driven more miles than a human would drive in 600 years.
Even if left turns cause problems, however, there is a surprisingly obvious solution: don’t turn left. UPS famously designs its delivery routes to almost entirely avoid left turns, saving money and time in the process.
According to Motor Trend writer Mark Rechtin, “until the tort lawyers and insurance industry sort out the at-fault versus no-fault circumstances, this technology grinds to a halt with the first fatal accident.”
This seems like a solvable problem. Compared to the technological challenges, which are novel, the legal issues seem routine. There will be lawsuits, and new legal precedents will probably emerge, but it seems unlikely that liability will be a binding constraint on autonomous vehicles. To Rechtin’s point, several collisions have actually already occurred, but progress continues.
Amir Efrati at The Information reports that Waymo vehicles sometimes, “miss turns because they can’t get into a turn lane on time.” This may be a problem for self-driving cars, but it’s also a problem for human drivers, who miss turns all the time because they struggle to merge.
The question is whether this is a safety issue or merely an inconvenience. If autonomous vehicles create unsafe driving conditions because of their struggles to merge, this would be a reason to hold off on launching services to the general public. If a failure to merge merely means that a vehicle has to circle around the block, then we can trust passengers to decide whether they are willing to put up with the annoyance, or they would prefer to drive themselves.
Mike Ramsey from Gartner has written in Forbes that autonomous vehicles are entering the Trough of Disillusionment phase of the Gartner Hype Cycle. Regulation is one of several challenges facing the mobility industry. Ramsey remarks, “Audi said the lack of clarity on regulations is preventing it from selling the A8 equipped with Traffic Jam Assist in the U.S.”
Regulation is a potential dealbreaker for autonomous vehicles. Furthermore, the lack of consistency and clarity in U.S. regulations may hold back even partially autonomous vehicles, such as the A8.
A silver lining is that the patchwork nature of state-by-state regulation in the U.S. means that, in the short-term, autonomous ridesharing services may be able to “shop” for the best regulatory environments. Companies can launch in geographic locations that are friendly to autonomous driving while avoiding jurisdictions with unclear or unfriendly regulations.
In a fascinating scoop, Amir Efrati from The Information reports on a surprising collision involving an autonomous vehicle safety driver. “In [dozing off], the driver inadvertently turned off the self-driving car software by touching the gas pedal, and then failed to take over the wheel.” The vehicle reverted to manual driving mode, but the safety driver remained asleep. The car eventually swiped into the highway median.
This scary and bizarre accident reflects the challenges faced by human safety operators. The contractors perform hour after hour of mind-numbing monitoring of the autonomous driving system, and yet are responsible for snapping to action at a moment’s notice when an emergency arises. Several different companies have highlighted the challenges of asking humans to perform this task.
This type of human factor engineering problem seems trivial at first blush, compared to the technical difficulty of computer vision or motion planning. But consider that the self-driving cars at Uber ATG have been grounded for six months, due in large part to a distracted safety driver.
The Information highlights stopped cars as a challenge for Waymo autonomous vehicles, particularly when a line of cars are stopped waiting to turn into a parking lot, for example.
It’s not clear what the precise challenge is – the statement is that the cars get “confused,” which anthropomorphizes the vehicles but doesn’t shed light on the nature of the problem.
The primary risk here would seem to be rear-ending the line of cars, causing a multi-vehicle pileup. But such reports have not emerged from Arizona. Indeed, the California DMV reports that most autonomous vehicle collisions involve the reverse scenario – human drivers rear-ending self-driving cars.
Motor Trend highlights the challenges posed by inclement weather. This isn’t a problem in Phoenix, where the weather is nearly always hot and sunny, but in locations that experience snow and rain, self-driving cars face meaningful hurdles.
Most self-driving car tests seem to take place in parts of California, Nevada, and Arizona that provide excellent weather. More difficult locations like Pittsburgh and Boston, however, are beginning to host autonomous vehicle testing.
Weather seems like a challenge that engineers will tackle incrementally over time, instead of jumping right into the snow.
Self-driving cars face many challenges, but the good news is that autonomous vehicles are getting better every day. Although some scenarios may cause difficulties today, tomorrow’s autonomous vehicles will overcome them and expand into new areas, which will bring their own set of challenges.
One of the major accomplishments of Waymo and other self-driving car developers is that the vehicles are reaching the point that even if a particular maneuver is still challenging, alternative routes and decisions are available. Just like new drivers, autonomous vehicle developers are likely to limit self-driving cars to the scenarios they can handle, and gradually expand the range of what self-driving cars are permitted to do, as their capabilities improve.
Originally Posted on Forbes at: https://www.forbes.com/sites/davidsilver/2018/10/05/what-hurdles-do-self-driving-cars-face-as-waymo-gets-ready-for-prime-time/#6d3af4ef4d0f