Google sister Waymo has launched a transparency offensive to increase confidence in accident prevention.
It’s about the situation where the car recognizes that there could be an accident and reacts. But how rated Waymo the situation? Waymo has developed a test method for this, which they call Collision Avoidance Testing (CAT). How does this tool work, how are the test scenarios identified and what other tools are there?
One of Waymo’s methods of assessing the security of the “Waymo Driver” is scenario-based testing – a combination of simulated driving, using the test tracks and driving in real road conditions. The concept is one of many methods of assessing security readiness. The CAT system is designed for the latest version of the Waymo Driver.
This evaluates how well the system avoids accidents and reduces the risk of injury. It is compared to the behavior of reference model NIEON, which imitates an attentive person. This means a person who is not distracted or tired. But unlike humans, NIEON will never take their eyes off the road or get tired. So it’s perfection on an almost human level.
To identify the test scenarios, the data from real trips that Waymo has collected over the years since 2016 is used. In addition, the AI is provided with police accident databases, accidents recorded by dashcams and expertise. This data is supplemented with maps, driving conditions and road types and is regularly updated. The ongoing expansion is done by mapping and traffic tracking new areas where Waymo expands.
The primary data comes from Waymo’s closed proving ground called “Castle“. This is then modified and reproduced through the simulation. In this way the system is exposed until it is ready for the road. If dangerous situations arise, these are recorded and tested in the simulation.
In this way, you could compare the driving behavior of Waymo Driver with the reference model NIEON in terms of collisions or avoidance of injuries and draw a conclusion that, according to Waymo, is similar. Furthermore, a comparison is made with previous software versions to determine the performance of other evaluations.
The concept can be applied to cars or trucks as it is designed to be flexible and improves through a learning process.