Yandex rolls out proprietary lidars across AV fleet
Lidar is an important sensor for self-driving cars. Lidars use laser light to scan their surroundings and create a 3D picture of the outside world. Using reflected signals, lidars detect the shapes of various nearby objects and distance to them down to a centimeter even in the dark. They also enable autonomous vehicles to pinpoint their position on the road by comparing the 3D images they create with the 3D maps that are built into them.
We started making our own lidars in the spring of 2019, and nine months later we began testing first driverless vehicles equipped with our lidars on city streets. Today, after another two years of development, we have completely switched to using our own proprietary software-defined lidars as the main sensor. All our fourth-generation cars are already equipped with them, and they will also be included in all of our new cars going forward. This blog post details key features of our lidar and explains why having a proprietary sensor is so important for the development of self-driving technology.
To start with, what is a software-defined lidar? It is a sensor whose parameters including the number of beams, angle of view, range and others can be changed in advance or when the vehicle is on the move, depending on the tasks and traffic conditions it is facing. For example, this sensor can increase point cloud density in the area near the vehicle when it is moving through a courtyard or increase range when driving at a high speed on a highway. This feature can also be used for more complex problems, such as increasing the resolution precisely in the area where the algorithms report a certain level of uncertainty in object recognition.
Due to the software-defined nature of our lidar, it doesn’t have any fixed essential parameters, such as detection range or resolution. Varying the parameters of laser and scanning system we can reach up to 500 meters range and 0.1 resolution in all directions. This means that the scanning pattern of our lidars can change in real time, choosing the scanning characteristics that are best suited to each situation.
So, when driving on a narrow street with active pedestrian traffic, we can focus most of the rays on the road in front of the vehicle, greatly increasing the cloud density for recognizing nearby objects. On wide high-speed highways, on the other hand, it is more important to see objects in the distance. Changing the corresponding settings will increase the density in the desired zone and make it possible to recognize a passenger car at a distance of 200 meters or a tractor trailer at a distance of 500 meters. This range is also important for the localization of the vehicle with the help of lidar localization algorithms in areas without very dense development near the road, such as wide avenues, highways and overpasses.
This sort of flexibility in setting parameters is possible thanks to the lidar’s scanning system design. In addition, there are no rotating electronic elements inside it, which means that these kinds of lidars are more reliable and less susceptible to thermal distortions, which we encountered earlier when using rotating lidars. Our solid-state lidars function equally well on real roads at temperatures from -30 Celsius to +30 Celsius, the usual annual temperature range for Russia and other countries and regions with a pronounced continental climate.
The use of a software-defined approach becomes possible when the sensor and the self-driving technology stack are tied. Control of the main sensor also provides other benefits, such as access to raw data, collection of specialized data sets and, of course, price optimization. We invested a lot of effort into the development of our own lidar. As a result, we were able to go, in less than three years, from concept to a production model that, in many respects, surpasses the lidars that we were using before, while costing us the same amount. We were able to maintain this pace by immediately testing our hypotheses and improvements using the lidars on real self-driving cars. To date, our vehicles have already traveled more than 500,000 kilometers using our proprietary lidars, proving their functionality and reliability.
We have reached an important milestone today, and we will continue to perfect our lidars as well as develop new ones. We are already testing prototypes of lateral lidars for near-field detection, and we are also designing a lidar for our autonomous delivery robots. The use of our own lidars in our fleet has made us even more confident in the fact that a close relationship between the technology and the sensor can lead to impressive capabilities, and we hope that more breakthroughs and meaningful improvements are still to come.
Written and published by Yulia Shveyko
