AVs Need More Snow Training, Moscow Trial Reveals

Bad roads are a well-known fact about Russia and the dependence of driverless vehicles on high quality of infrastructure is another known. Yet, nobody in the country believes in a fast upgrade of the roads even for the sake of all the benefits offered by autonomous transport.

Local developers prefer to follow the think of an old local anecdote. It tells of a foreign engineer who, upon seeing an off-road ambulance back in 1965, said “Russians are incredibly inventive when they don’t feel like building roads”. Half a century later, the country is teaching autopilots how to survive in existing infrastructure, seemingly a more feasible task than fixing motorways. As a part of this strategy, the federal state had given 162M rubles ($2.45M) to arrange a ‘Winter City’ contest among AV developers and another 175M ($2.65M) to the winner.

The track prepared for the contest’s qualifying sessions imitates a picture familiar to local drivers. Street illumination is uneven. Snow blowers provide blizzard conditions hiding road markings and signs beneath a thick covering of snow. Construction zone barriers and parked cars turn a two lanes motorway into a slalom on ice. Stuntmen act as idiot drivers running a red light at an intersection. In these conditions, competitors were offered eight tasks, each adding up one point. Four points was the minimum score to qualify for the full competition.

No deep snow driving, no icy slopes, no bears, no -50C degree frost. “Some said we made the quest overtly easy,” said Viacheslav Gershov, director and head of development of technological challenges and initiatives department at Skolkovo Foundation, a co-organizer of Winter City. Experts from Israel’s Mobileye said otherwise, he added and possibly confirming this, no team reached the top scores.

Head of the AV project at Starline, Boris Ivanov, said: “DARPA and similar contests are typically conducted in good weather, with infrastructure in working order. In Moscow we had to deal with snow on sensors.” Starline is a car security systems producer from Saint Petersburg. Its team was among the leaders, albeit unable to gain a point for crossing a junction with the red-light runner.

Most participants said they were attracted by an opportunity to run their tech in country’s best AV-dedicated town simulator rather than the cash prize. “I’m not expecting a competition,” a participant Andrey Karpenkov, head of department of robotics at Kovrov State Technological Academy, said a few days before qualification had started. “It’s a place where we can run algorithms in urban environments and work the weaknesses. It’s a step toward street tests.” The academy’s team Alpha failed to qualify.

While 33 teams had applied for the contest, only nine had been technically validated for qualifying sessions starting on February 1 and ending on March 5. Six of them successfully qualified. Currently, the three best performers are Starline and two academic teams MADI (Moscow Automobile and Road Construction State Technical University) and NNSTU (Nizhny Novgorod State Technical University), each with seven points. Also on the list of qualified teams are InnoTeam (the University of Innopolis from Kazan), Avto-RTK (joint team of Scientific-design bureau of computing systems from Taganrog, Southern Federal University, and Southwestern State University) and BaseTrack (company Energo from Moscow).

Obviously, cars exposed at the Winter City still need more development to be called “Russian-winter-proof”. Now the participants have a nine month break to fix weak areas in their technology. All found tasks of interpretation and prediction were the biggest challenges facing their machines. One example was a sequence of parked cars and road repair barriers which should have been interpreted as slalom. A second example was the prediction of a child’s appearance in front of the car as it chased a ball into the road.

Autopilots systems already outperform humans in certain tasks like recognizing traffic signs, said Victor Shirshin, leader of SmartVision team from Tomsk, allowed by the organizers to try the event as a non-competitor. In fact, the poorer illumination, the bigger advantage machinery vision has over drivers’ eyes. “Autonomous cars are not a presence but a near future,” he repeated the industry’s usual mantra.

Practice sessions due in December 1-25, 2019 will reveal the winner… or not. A vehicle must complete a 50kms (30 mile) route in three hours on the first attempt without help from humans and in the presence of traffic and other AVs. Up to three teams with the best times will share the prize money assuming any vehicle makes it!

Leave a comment

Your email address will not be published. Required fields are marked *