Over the years, Google/Waymo and Uber have been recognized as leaders in the autonomous car industry. However, their approach based on the "laser radar" has very serious drawbacks, due to which it is practically unsuitable for fully autonomous traffic on the roads.
Lidars cannot do anything in fog and heavy rain - IR radiation is well absorbed by water vapor. So they will not be able to safely drive in difficult conditions or will have to create a normal vision system using cameras. But then it is not clear why they need an expensive lidar, while a set of cameras costs several times less.
What does Tesla offer? A computer can be taught to see with the help of eight standard video cameras looking in all directions using its standard cars. Of course, they do not see in the sense that man see. When the driver recognizes the image (for example, a pedestrian whose clothes are red on one side and white on the other), then even if the pedestrian turns to him on the other (differently colored) side, the driver will still understand that this is a person. This is because the brain models possible changes in the images it sees, based on their possible behavior.
A computer cannot simulate a complete image, therefore, to distinguish complex objects with confidence, it needs to have data on how a complex object can look from any point of view. Tesla cannot teach a computer to think, but it is capable of giving its computers enough images to make them even distinguish complex objects in any imaginable environment.
The basic autopilot software looks like a typical neural network. That is, it consists of many elements - software analogues of neurons. At first, the neural network is "not trained." To train the neural network to the necessary (safe) level of driving support, you need a huge amount of information (training set).
Tesla uses its own cars as the source of its training set. Each time a driver makes a maneuver at an intersection or when changing lanes, he acts as a “teacher” for training the neural network of all Tesla computers in general.
Typically, the performance of a deep learning system is limited, at least in part, by the quality of the training set used to train the model. In many cases, significant resources are invested in the collection, processing and annotation of training data. The effort required to create a training kit can be significant and often tedious. In addition, it is often difficult to collect data for specific use cases for which the machine learning model needs to be improved.
That is why Tesla is constantly striving to improve the system and method of obtaining training data. The patent 'System and method for obtaining training data' was filed in 09/13/2019 and published on 03/19/2020. Inventor Andrej Karpathy, Tesla’s head of AI and Autopilot software.
It describes several innovations that can be implemented in many ways, including as a process; apparatus; system; the composition of the substance; computer software product implemented on a computer-readable storage medium; and / or a processor, such as a processor, configured to execute instructions stored and / or provided by memory associated with the processor.
For this, various methods are used: