Just like teenagers, self-driving cars need practice to really learn to drive

A self-driving car, of unknown level of education. Image: Grendelkhan/Flickr/creative commons.

What do self-driving cars and teenage drivers have in common?

Experience. Or, more accurately, a lack of experience.

Teenage drivers – novice drivers of any age, actually – begin with little knowledge of how to actually operate a car’s controls, and how to handle various quirks of the rules of the road. In North America, their first step in learning typically consists of fundamental instruction conveyed by a teacher. With classroom education, novice drivers are, in effect, programmed with knowledge of traffic laws and other basics. They then learn to operate a motor vehicle by applying that programming and progressively encountering a vast range of possibilities on actual roadways. Along the way, feedback they receive – from others in the vehicle as well as the actual experience of driving – helps them determine how best to react and function safely.

The same is true for autonomous vehicles. They are first programmed with basic knowledge. Red means stop; green means go, and so on. Then, through a form of artificial intelligence known as machine learning, self-driving autos draw from both accumulated experiences and continual feedback to detect patterns, adapt to circumstances, make decisions and improve performance.

For both humans and machines, more driving will ideally lead to better driving. And in each case, establishing mastery takes a long time. Especially as each learns to address the unique situations that are hard to anticipate without experience – a falling tree, a flash flood, a ball bouncing into the street, or some other sudden event. Testing, in both controlled and actual environments, is critical to building know-how. The more miles that driverless cars travel, the more quickly their safety improves. And improved safety performance will influence public acceptance of self-driving car deployment – an area in which I specialise.

Starting with basic skills

Experience, of course, must be built upon a foundation of rudimentary abilities – starting with vision. Meeting that essential requirement is straightforward for most humans, even those who may require the aid of glasses or contact lenses. For driverless cars, however, the ability to see is an immensely complex process involving multiple sensors and other technological elements:

  • radar, which uses radio waves to measure distances between the car and obstacles around it;
  • LIDAR, which uses laser sensors to build a 360-degree image of the car’s surroundings;
  • cameras, to detect people, lights, signs and other objects;
  • satellites, to enable GPS, global positioning systems that can pinpoint locations;
  • digital maps, which help to determine and modify routes the car will take;
  • a computer, which processes all the information, recognising objects, analysing the driving situation and determining actions based on what the car sees.

How a driverless car ‘sees’ the road.

All of these elements work together to help the car know where it is at all times, and where everything else is in relation to it. Despite the precision of these systems, however, they’re not perfect. The computer can know which pictures and sensory inputs deserve its attention, and how to correctly respond, but experience only comes from traveling a lot of miles.

The learning that is occurring by autonomous cars currently being tested on public roads feeds back into central systems that make all of a company’s cars better drivers. But even adding up all the on-road miles currently being driven by all autonomous vehicles in the U.S. doesn’t get close to the number of miles driven by humans every single day.

Dangerous after dark

Seeing at night is more challenging than during the daytime – for self-driving cars as well as for human drivers. Contrast is reduced in dark conditions, and objects – whether animate or inanimate – are more difficult to distinguish from their surroundings. In that regard, a human’s eyes and a driverless car’s cameras suffer the same impairment – unlike radar and LIDAR, which don’t need sunlight, streetlights or other lighting.

This was a factor in March in Arizona, when a pedestrian pushing her bicycle across the street at night was struck and killed by a self-driving Uber vehicle. Emergency braking, disabled at the time of the crash, was one issue. The car’s sensors were another issue, having identified the pedestrian as a vehicle first, and then as a bicycle. That’s an important distinction, because a self-driving car’s judgments and actions rely upon accurate identifications. For instance, it would expect another vehicle to move more quickly out of its path than a person walking.


Try and try again

To become better drivers, self-driving cars need not only more and better technological tools, but also something far more fundamental: practice. Just like human drivers, robot drivers won’t get better at dealing with darkness, fog and slippery road conditions without experience.

Testing on controlled roads is a first step to broad deployment of driverless vehicles on public streets. The Texas Automated Vehicle Proving Grounds Partnership, involving the Texas A&M Transportation Institute, University of Texas at Austin, and Southwest Research Institute in San Antonio, Texas, operates a group of closed-course test sites.

Self-driving cars also need to experience real-world conditions, so the Partnership includes seven urban regions in Texas where equipment can be tested on public roads. And, in a separate venture in July, self-driving startup Drive.ai began testing its own vehicles on limited routes in Frisco, north of Dallas.

These testing efforts are essential to ensuring that self-driving technologies are as foolproof as possible before their widespread introduction on public roadways. In other words, the technology needs time to learn. Think of it as driver education for driverless cars.

People learn by doing, and they learn best by doing repeatedly. Whether the pursuit involves a musical instrument, an athletic activity or operating a motor vehicle, individuals build proficiency through practice.

The ConversationSelf-driving cars, as researchers are finding, are no different from teens who need to build up experience before becoming reliably safe drivers. But at least the cars won’t have to learn every single thing for themselves – instead, they’ll talk to each other and share a pool of experience.

Johanna Zmud, Senior Research Scientist, Texas A&M Transportation Institute, Texas A&M University .

This article was originally published on The Conversation. Read the original article.

 
 
 
 

What's actually in the UK government’s bailout package for Transport for London?

Wood Green Underground station, north London. Image: Getty.

On 14 May, hours before London’s transport authority ran out of money, the British government agreed to a financial rescue package. Many details of that bailout – its size, the fact it was roughly two-thirds cash and one-third loan, many conditions attached – have been known about for weeks. 

But the information was filtered through spokespeople, because the exact terms of the deal had not been published. This was clearly a source of frustration for London’s mayor Sadiq Khan, who stood to take the political heat for some of the ensuing cuts (to free travel for the old or young, say), but had no way of backing up his contention that the British government made him do it.

That changed Tuesday when Transport for London published this month's board papers, which include a copy of the letter in which transport secretary Grant Shapps sets out the exact terms of the bailout deal. You can read the whole thing here, if you’re so minded, but here are the three big things revealed in the new disclosure.

Firstly, there’s some flexibility in the size of the deal. The bailout was reported to be worth £1.6 billion, significantly less than the £1.9 billion that TfL wanted. In his letter, Shapps spells it out: “To the extent that the actual funding shortfall is greater or lesser than £1.6bn then the amount of Extraordinary Grant and TfL borrowing will increase pro rata, up to a maximum of £1.9bn in aggregate or reduce pro rata accordingly”. 

To put that in English, London’s transport network will not be grinding to a halt because the government didn’t believe TfL about how much money it would need. Up to a point, the money will be available without further negotiations.

The second big takeaway from these board papers is that negotiations will be going on anyway. This bail out is meant to keep TfL rolling until 17 October; but because the agency gets around three-quarters of its revenues from fares, and because the pandemic means fares are likely to be depressed for the foreseeable future, it’s not clear what is meant to happen after that. Social distancing, the board papers note, means that the network will only be able to handle 13 to 20% of normal passenger numbers, even when every service is running.


Shapps’ letter doesn’t answer this question, but it does at least give a sense of when an answer may be forthcoming. It promises “an immediate and broad ranging government-led review of TfL’s future financial position and future financial structure”, which will publish detailed recommendations by the end of August. That will take in fares, operating efficiencies, capital expenditure, “the current fiscal devolution arrangements” – basically, everything. 

The third thing we leaned from that letter is that, to the first approximation, every change to London’s transport policy that is now being rushed through was an explicit condition of this deal. Segregated cycle lanes, pavement extensions and road closures? All in there. So are the suspension of free travel for people under 18, or free peak-hours travel for those over 60. So are increases in the level of the congestion charge.

Many of these changes may be unpopular, but we now know they are not being embraced by London’s mayor entirely on their own merit: They’re being pushed by the Department of Transport as a condition of receiving the bailout. No wonder Khan was miffed that the latter hadn’t been published.

Jonn Elledge was founding editor of CityMetric. He is on Twitter as @jonnelledge and on Facebook as JonnElledgeWrites.