Everything that still baffles robobomili, starting with gulls

Watching a parking lot on your own - how to watch the magic. But if you look behind the scenes, the situation there will not be so rosy. The car confuses snowflakes with obstacles, loses road markings and skips parked cars.

Engineers seek to drive cars to better people, and save millions of lives annually leaving us. The human factor imposes 94% of annual victims.on US roads, according to a report from the US National Statistical and Analytical Center. Robomobils promise to prevent most of them. Even the opportunities that are already available today, such as a warning about losing a strip (available on many cars), can reduce road mortality by 86% - this is how the Insurance Institute for Highway Safety, IIHS assesses the situation.

However, we are still quite far from robomobiles, despite all the advertising. Technology assistance to drivers capable of taxing, braking and driving in the stream (under the supervision of a person) is already entering the market, with Tesla as a leader. However, “it is important to note that these machines are not able to drive safely on their own,” says David Zooby, IIHS principal investigator. "Ready autonomous car that can go anywhere at any time, while you can not buy at your local auto show, and it will not appear there for quite some time." IIHS experienced five leading systems on tracks with hills and turns. None of them ran into obstacles, but almost all of them missed marks one or several times, approached or crossed the markings or turned off while driving.


Incidents during the test.
Above: bump test. Below - cornering tests.
Crimson color - the intersection of the markup, blue - hitting the markup, dark cherry - off the system.

Their achievements are still impressive, and machine learning continues to make a revolution in the field of the possible. Uber and Waymo deliver passengers in robotic vehicles (if there is a human driver following the ride) in different cities, from Pittsburgh to Phoenix. However, the first truly autonomous machines can appear in nursing homes, corporate campuses, and gated communities: in a controlled environment that a computer can easily mark out. “I urge any automaker to launch its mobile in difficult urban conditions, without a driver, in any weather conditions,” says Ryan Chin, co-founder and director of Optimus Ride, working on several campuses and closed areas on a pilot version of its autonomous driving technology. “The industry is not ready for this yet. Even the best systems. ”

What confuses robobom today? Raindrops, obstacles, electrical tape and even gulls - all this causes problems in the work of algorithms. We have collected the most interesting obstacles for robots.

Modified stop signs

Computer scientists have slightly changed the stop signs to see if small changes can confuse robotic cameras in cases where the human driver does not even notice them. Pseudographs made the algorithms confuse stop signs with speed limit signs in two cases out of three, and randomly sticking adhesive tape, which the researchers called “abstract art sticker attack”, in 100% of cases led to a wrong categorization of the sign.

Falling snowflakes

Snowflakes and raindrops are notorious for scattering sensor signals. They can create the illusion that the car is surrounded by obstacles. Algorithms are increasingly coping with the compilation of three-dimensional high-precision environment maps using lasers, and are already able to distinguish between H ₂ O and solid objects, but winter remains one of the biggest problems of robots. Snow blurs the place where, according to computers, the road begins, and changes the tire grip. "In many cold regions, it will take much longer to introduce autonomous cars than some people claim," says Sam Abouelsamid of Navigant research. "In Toronto, in 2020, there will be no rob-mobiles in the winter."

Seagulls

Birds can also confuse computers. In Boston, NuTonomy had to reprogram machines to cope with stubborn gulls. “Local calm seagulls stop cars just by standing on the roadway — NuTonomy's quiet electric cars don't frighten them. The engineers had to make the cars quietly approaching the birds in order to scare them away, ” writes Bloomberg.

Foam

Researchers at the University of South Carolina disoriented Tesla S, covering the obstacles with sound-absorbing foam so that ultrasonic sensors cannot see them. Tesla can also fool a 40-dollar set of Arduino and an ultrasonic transmitter that generates sound waves - the car will not be able to find free parking space or will miss real close obstacles.

Cars moving off the highway

Machines are guided by other machines. At high speeds on the highway, it works well, but this technique can lead to unexpected exits from the road, when the car suddenly starts to go after another, moving sideways. “When a car moves too slowly to keep track of the markup, the active systems for keeping the car in a row use the car going ahead as a guide,” writes IIHS. “If the driving car is heading for the ramp, then the car driving behind it can follow.”

The hills

The IIHS testers, driving over the hills of Central Virginia, found that even advanced driver assistance systems could miss the markup when entering the top of the hill. Not seeing the road ahead, the car began to wiggle left and right to find the center of the row, and to frighten drivers who did not receive warnings about the need to regain control.

Bridges

Bridges are black boxes for robots, as they say Electronic Component News. The bridges lack many clues that are surrounded by ordinary roads, as a result of which the sensors may not cope with keeping the car on the lane. The magazine compared the situation with when “you walk in a straight line from one end of a huge room to the other, and the moment you reach the middle, the light goes out. You do not see anything, and although you have a certain idea of ​​the direction of travel, you are easily knocked off course. ”

Tree shadows

Tesla Model 3 resorted to braking "unnecessarily, or by exercising excessive caution" 12 times per 290 km. Seven times this was due to the shadows of trees falling on the road, and the rest were caused by oncoming cars driving in a different lane or crossing the road far ahead. “These braking did not lead to unsafe situations, because the speed reduction was small and fast, so the car did not have time to slow down too much,” writes IIHS. “However, excessive braking may increase the risk of collisions during heavy traffic, especially if it is more abrupt. In addition, drivers who feel the irregularity of slowing down the car, can abandon the adaptive cruise control and will miss the benefits of the system associated with safety. "