Tech giants and automakers were falling over each other to announce their self-driving vehicle plans a few years back. All of the FAANG (Facebook, Amazon, Apple, Netflix, and Google (now Alphabet)) minus Netflix, were interested in self-driving cars. As if to not leave Netflix behind, Volvo announced plans to have Netflix in driverless cars. Investors were giddy with possible returns. Social media was buzzing with astonishing claims. Startups sprouted all over the world touting various benefits of an unproven tech. Nearly all of them envisioned driverless cars plying streets by 2020. Uber and Lyft were (and still are) banking on self-driving taxis to finally see a profit.
10 million self-driving cars will be on the road by 2020 — Business Insider headline in 2016
Well, it’s 2020 and no actual driverless car, taxi, or truck is on the road. To be fair, driver-assist tech has advanced. Adaptive cruise control, lane control, collision alert, and Tesla’s autopilot can ensure smooth sailing on the highway. Yet, true autonomous, the one where the car drives itself without a human intervening is not here except in test vehicles. This flies in the face of all the furor of a few years back. Investors are cautious, startups are going belly up, and companies are scaling back ambition.
So, what happened? What hurdles did the driverless car hit? And, is there any hope? Or was it another case of overhype and low returns?
The different types of driverless tech
Driverless, self-driving, or autonomous all refer to a vehicle that is smart enough to drive from one point to another without human assistance. They are classified into six levels from the manual (Level 0) to full automation (Level 5).
A driverless vehicle can be designed in two ways. One is a de novo driverless vehicle—designed from grounds up with the intention of being autonomous. The other is a set of tech to be mounted on a regular vehicle to make it autonomous. The former is more capital intensive. But, it may make sense because the car is designed with the purpose of it being driverless. The vast majority of companies are targeting the latter segment. It makes sense why. The upfront investment is less, time to experimentation is short and once the tech is ready it can be attached to any vehicle (in theory) making it more viable. This includes Waymo, Uber, Lyft, GM’s Cruise, Apple, etc. Cruise has also built its own model of a driverless car with no steering.
The cause of delays
Autonomous cars work on a simple principle — take information from surrounding and make judgment calls — something human drivers do all the time. The problem is putting it into reality. A driverless car would use cameras and radar tech to take in info, then decide what to do based on hard-wired code. Machine learning trains the car on millions of miles of data. Using this plus knowledge of traffic rules, the car should make proper driving decisions. And herein lies the problem, to collect the driving data, an imperfect machine has to be on the road. This puts other pedestrians and vehicles in a training exercise they did not agree to.
Simulations can’t solve autonomous driving because they lack important knowledge about the real…
Large-scale real-world6t7ty data is the only way
An alternative is training simulations. So, engineers develop various models for the car to train on. Do they work as well as real-life learning? Not quite. Driving a car involves a lot of everyday communication — eye contact, or waving someone to cross the road even when the light is green. Also, it is hard to simulate rare traffic conditions like someone dashing across the street not at a crosswalk, or for sudden weather issues.
Driverless cars have been marred by accidents. The most noticeable one was by an experimental Uber vehicle that ran over a pedestrian walking across the street with a bike. The vehicle also had a human behind the wheel. Investigations into the mishap identified numerous issues at the tech level as well as human error. Uber halted their trials to fix the flaws. Accidents have also occurred with Tesla Autopilot. A lot of them have to do with the drivers. Tesla Autopilot is not autonomous and requires the driver to be attentive much like another driver-assist tech. Not doing so can be fatal.
Uber had “inadequate safety culture, exhibited by a lack of risk assessment mechanisms, of oversight of vehicle operators, and of personnel with backgrounds in safety management” — NTSB report on 2017 Uber crash
COVID-19 has added further delays to autonomous car development. Social distancing rules made work hard at offices or in test cars, where 2 drivers are mandated. Shipment delays of necessary equipment didn’t help either. Investment in this tech may also dry out with the delays in development. At this rate, one wonder is we should we write driverless tech off? And is there any real-world benefit of having this tech anyway?
The case for driverless tech
Proponents of driverless tech often tout the tech as safer and greener than human driving. The safety of driverless cars remains to be seen. On the face of it, autonomous cars won’t be drunk, napping, or texting when driving. On the other hand, a true Level 5 autonomous car is still a work in progress. Likewise, the idea that self-driving cars may reduce car ownership or driving may not be true. An increase in tech efficiency is often offset by expanded use. One study showed that given the option of a simulated self-driven car (a chauffeur), families made more trips. Car ownership is another facet. The onset of ride-sharing apps hasn’t reduced car ownership. Although, this may change when autonomous cars are introduced.
So while driverless car’s environmental benefits and safety are debatable, there are tangible benefits. These cars are great for the elderly, the long commuter, and the disabled. Ridesharing companies are banking on this tech to make profits. The trucking industry can also benefit from automation. The first company to break through the driverless tech would reap riches for their investors and owners alike. Innovations in the driverless tech can also distill to other nascent tech areas. But, then again, all this only happens when the tech is available for consumers.
The progress so far
The autonomous driving industry uses two benchmarks to measure progress — total miles driven and manual override per 1000 miles (also referred to as disengagement rate). The former is a proxy for the driving data accumulated that trains the model. The latter measures how good autonomous driving is. If an autonomous vehicle needs regular interruption by humans to avoid traffic incidences, it is not very safe (or very autonomous).
Everyone hates California's self-driving car reports
California's Department of Motor Vehicles released the latest batch of reports from companies that are testing…
At 100% interruptions, the car is basically driven by humans. So, the higher the total miles is and the lower the manual overrides are, the more autonomous a car is. But, not all miles are the same. Driving in rural areas or highways is not the same as crowded cities. So, both benchmarks could be gamed by driving a car on a closed course, not crowded freeways, or deserted rural roads.
Data on the benchmarks are hard to find because only California mandates their reporting. Companies dislike California’s mandate. Their argument is the benchmarks are not an accurate measure. One can always game the system by driving around on a less traversed roads. The benchmarks also put the companies in a bad light if their data is not good. Regardless, one thing is true that these metrics are not good for intercompany comparison.
“Comparing disengagement rates between companies is worse than meaningless: It creates perverse incentives….If I wanted to look even better, I’d do a ton of easy freeway miles in California and do my real testing anywhere else” — Bryant Walker Smith, Associate Professor, USC School of Law and an expert in self-driving cars
Alphabet’s Waymo and GM’s Cruise have impressive data on the metrics available. They clocked close to a million miles in 2019 in CA and had less than 0.1 disengagement rate per 1000 miles. Yet, one cannot base this meager and easy to game data to predict which company would win the autonomous race, if any.
So what does it all mean?
In a nutshell, we won’t get a self-driving car this year. We may not get it in the next few years. But, we are probably closer to getting one than ever before.
We are closer than ever to an autonomous car, and yet so far!
We may never get a true Level 5 autonomous car. Instead, the driver-assist tech would become sufficiently advanced to navigate the highways on its own with little intervention. Road tech may also develop to guide the cars. We have made significant progress from the initial frenzy days. In the end, the initial frenzy can after all be the churn of industry. Excitement is followed by better sense and then real progress.