Skip to main content
conference lunch move company map contacts lindholmen lindholmen 2 travel info


Yandex's AV winter testing milestone

Yandex's autonomous cars have driven over six million miles in the challenging conditions of Moscow’s icy roads. 

Yandex, Russia's ’Google-like’ internet giant, began testing its autonomous cars on Moscow's icy winter roads over three years ago. The goal was to create a ’universal’ self-driving vehicle that could safely maneuver around different cities across the globe. Now, Yandex claims its trials have been a resounding success. The vehicles recently hit a major milestone by driving over six million miles (10 million kilometers) in autonomous mode, with the majority of the distance travelled in the Russian capital.

To overcome the hazardous conditions, Yandex says it cranked up its LIDAR performance by implementing neural networks to filter snow from the lidar point cloud, thereby enhancing the clarity of objects and obstacles around the vehicle. The system was also able to adjust to both sleet and harder icy conditions over time, allowing it to gradually make better decisions on everything from acceleration to braking to switching lanes. In addition, the winter conditions pushed the system's built-in localization tech to adapt to hazards such as hidden road signs and street boundaries and snow piles the "size of buildings." This was made possible by the live mapping, motion, position and movement data measured by the system's mix of sensors, accelerometers and gyroscopes.

The article includes a short time-lapse video recorded from one of their vehicles driving around during fairly snowy conditions that is well worth a look.  

Personal comment:
Assessing the quality of the Autonomous Driving System is not easy since it requires some idea of what driving in Moscow is like. That said there are some interesting sequences that caught my eye. At 48 seconds into the video the AV makes a nice manoeuvre in relation to a truck that suddenly decides to change lane. Judging by the safety driver’s stillness in the driver’s seat the application of brakes was reasonably smooth. After 1 minute 40 seconds the AV makes a swerving entry onto the highway after a left turn. It is difficult to know if this is due to difficulties identifying where the lane is since the surface is covered by snow, the snow making the surface slippery or if the AV is adapting to the vehicle approaching from behind at a high speed. Would a human driver cause the vehicle to behave in the same way? It’s difficult to know. The lack of clear lane markings due to the snow could also explain why the AV seems to often stray to the right or the left compared with the vehicle in front. But that could also be explained by the driver in front not caring too much about such details. Again, it is difficult to assess the quality of the system without contextual information.

So, just for fun I went back to one of my all-time favourite AV driving videos to see how the Yandex vehicle driving in Moscow compares to highway driving in California in 2015. The driving task in Moscow is obviously more complicated, adding left-turns and traffic lights while the weather in California delivers as expected, making it easier to identify lane markings. Both systems deliver the same overall information to the driver in terms of surrounding traffic and environment, but the Russian system has a much neater graphic representation. 

It would be interesting to see how Yandex’s system behaves in similar traffic conditions to that shown in the video, but under milder weather conditions since that would enable assessment of the weather’s impact on the system. Could the Yandex system be more generally resilient and adaptive to other sorts of environmental conditions as well? Or will it be overtrained for specific conditions and then try to assume everything is snow? After all, everything looks like a nail if you have a hammer.

The article and accompanying video raise the important question of how well we can replicate and, in a controlled way, alter the driving scenarios in order to assess if the behaviour of the complete system is improving or not.

Written by Håkan Burden,
RISE Mobility & Systems (Mobilitet i transformation)