Published on May 8th, 2018 | by Steve Hanley0
Autonomous Vehicles: False Positives, 3D Maps, & Networks
May 8th, 2018 by Steve Hanley
Everybody knows autonomous vehicles are coming, we just don’t know when. Some are already here, operating in limited geographical areas. Others are years away. Regulators are trying to be supportive of the technology while protecting the public interest. New technology always comes with risks, but if the course of human history tells us anything, it is that you can’t stand in the way of progress.
Uber And False Positives
One of the biggest hurdles for software engineers working on autonomous driving systems is creating algorithms to deal with what are known as “false positives.” Look at it this way. You are driving down the highway when a sheet of snow and ice falls off the tractor trailer in front of you. (Those of you living in warm climates will have to take our word for it that such things sometimes happen.) Do you a.) brake hard, causing the cars behind to slam into you, b.) slow and assess the situation, c.) do nothing, or d.) change lanes abruptly, mash the throttle, and speed ahead to avoid the falling snow?
Those are decisions autonomous driving systems must make thousands of times a second. The software that controls them can be set to be more or less sensitive to unexpected inputs, what engineers call “false positives.” They don’t want their cars to go into full panic braking mode every time a plastic bag or a piece of paper blows across the road ahead. On the other hand, they should be able to detect stopped emergency vehicles that are actual fire engines.
An Uber self-driving car struck and killed a pedestrian on Tucson, Arizona, in March. A preliminary report by Uber says the car’s computer did detect the person crossing the road but took no evasive action because it treated that information as a false positive, according to a report by The Information. In Uber’s case, the algorithm erred on the side of ignoring the data and deciding it was getting a false positive.
It is cold comfort to the victim of her family, but this one of the inevitable bumps in the road (no pun intended) on the way to fully autonomous vehicles. If self-driving cars really can prevent 50% of fatal accidents, that means nearly 20,000 Americans will die in automobile collisions involving autonomous cars each year. Nobody thinks twice about 35,000 to 40,000 traffic deaths a year today because we are used to humans driving around and killing each other.
The question is, how many deaths are we willing to tolerate when machines are doing the driving? Perhaps this is an area in which manufacturers would welcome some carefully crafted regulations, if only to give them legal cover for those inevitable cases when their self-driving cars inflict grievous bodily harm on people.
Autonomous Driving Without 3D Maps
To date, companies like Waymo, Uber, Cruise Automation, and many others approach autonomous driving by making a digital imprint of a given area known as a 3D map. In it, every speed limit, stop sign, curb, lane marking, and traffic light is embedded in a neural map that the computer can compare with the data it is receiving from onboard sensors. Only by comparing one with the other can it know when and how it is safe to proceed.
As you might imagine, that much data requires terabytes of memory and incredible processing power. Just last week, Elon Musk said Tesla owners may need to upgrade the computers in their cars in order to enable full Level 5 autonomous driving. Tesla may be ahead of those other companies when it comes to being able to decode the environment surrounding its cars in real time without a complete 3D map, but it still has a long way to go before its cars learn to avoid traffic barriers and stopped emergency vehicles.
Researchers at MIT Computer Science and Artificial Intelligence Laboratory believe they have the answer — a much simplified system that combines GPS data with input from Lidar and infrared sensors to accurately “see” the road ahead out to a distance of 100 feet. The MIT research was supported in part by the National Science Foundation and the Toyota Research Initiative. And yes, 100 feet is not that far, but it’s farther than Wilbur and Orville’s first flight. Patience, grasshopper.
“The reason this kind of ‘map-less’ approach hasn’t really been done before is because it is generally much harder to reach the same accuracy and reliability as with detailed maps,” says CSAIL graduate student Teddy Ort, who was a lead author on a related paper. “A system like this that can navigate just with on-board sensors shows the potential of self-driving cars being able to actually handle roads beyond the small number that tech companies have mapped.” The research report will be presented at the International Conference on Robotics and Automation in Brisbane, Australia, later this month, according to a report by Science Daily.
The system is called MapLite and uses sensors for all aspects of navigation. The GPS data is only used to get a rough estimate of the car’s location. The system first sets both a final destination and what researchers call a “local navigation goal,” which has to be within view of the car. Its perception sensors then generate a path to get to that point using Lidar to estimate the location of the road’s edges.
MapLite can operate without physical road markings because it makes the basic assumption that the road will be relatively more flat than surrounding areas. “Our minimalist approach to mapping enables autonomous driving on country roads using local appearance and semantic features such as the presence of a parking spot or a side road,” says MIT professor Daniela Rus.
“At the end of the day we want to be able to ask the car questions like ‘how many roads are merging at this intersection?'” says Teddy Ort. “By using modeling techniques, if the system doesn’t work or is involved in an accident, we can better understand why.”
“I imagine that the self-driving cars of the future will always make some use of 3D maps in urban areas but when called upon to take a trip off the beaten path, these vehicles will need to be as good as humans at driving on unfamiliar roads they have never seen before. We hope our work is a step in that direction.”
Autonomous Ride-Hailing In Texas
Drive.ai, a California start-up with roots in the Artificial Intelligence Lab at Stanford, is partnering with the Hall Group and the city of Frisco, Texas to bring autonomous ride-hailing to the 10,000 members of Hall Group’s commercial and residential communities in the area. According to TechCrunch, the service will use designated pick up and drop off points. That means it will only need a limited amount of onboard data storage, as its operating area will be geofenced and well documented digitally.
“Self-driving cars are here, and can improve the way we live right now,” says Sameep Tandon, co-founder and CEO of Drive.ai. “Our technology is safe, smart, and adaptive, and we are ready to work with governments and businesses to solve their transportation needs. Working with the City of Frisco and Frisco TMA, this pilot program will take people to the places they want to go and transform the way they experience transportation.”
Before the service begins in July, the partners will conduct an extensive public awareness campaign to educate city residents to the fact that autonomous cars are operating in the area. The cars themselves have external billboards to communicate with pedestrians and other drivers.
Autonomous ride-hailing is seen by many as the antidote to urban congestion and that may be true. But let’s hope the future brings us self-driving cars that are not as homely as a 1959 East German Trabant. The orange and blue color scheme does nothing the conceal the unrelenting ugliness of these vehicles.