Changing lanes is something all drivers have to do at some point almost every time they drive. It may not seem like a big deal, but changing lanes is one of the hardest things autonomous cars have to attempt. Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a new algorithm that helps autonomous cars switch lanes quickly and easily.
Current algorithms that help autonomous cars change lanes rely on statistical models of the driving environment around the car. These models are difficult to assemble and can be too complex for the car to quickly analyze in a real-life driving situation. The algorithms can also be too simple and force the car to make impractical and often apprehensive decisions, which can lead to the car never changing lanes.
The new algorithm from MIT’s CSAIL splits the difference between these two problems. The algorithm allows autonomous cars to make aggressive lane changes that a human driver would typically make. The algorithm only relies on information that is immediately available about the surrounding cars.
Autonomous cars manage to avoid collisions by calculating and creating buffer zones around the cars surrounding them. The buffer zones account for the current positions of the cars and what the future positions of the cars may look like. A lane change is basically trying to stay out of the buffer zones of the surrounding cars.
When creating an algorithm, designers need to prove that it will avoid collisions. In order to do this, the researchers used a mathematical model that describes traffic patterns. In current autonomous cars, the mathematical model for the buffer zones around cars is computed in advance. The cars then bring up the predetermined buffer zones that fit the situation that the car is in. The predetermined buffer zones are too restrictive to be used in fast and dense traffic. This leads the autonomous cars to avoid changing lanes in situations that a human typically would.
The new algorithm creates buffer zones in real time. The autonomous cars using this new system maintain collision avoidance even though the buffer zones are created in real time. To create the new system, the researchers started with the Gaussian distribution. Gaussian distribution is the bell curve probability that, in this case, represents the position of the car while estimating the length and location. The new system creates a new logistic function in real time based on the estimated direction and velocity. The logistic function is multiplied by Gaussian, skewing the distribution by the car’s movement. Higher car speeds lead to an increase in the skew. The skewed distribution is the key to creating a new buffer zone developed while the car is driving in real time, and the mathematical equation allows the autonomous car to create a buffer zone in real time.
The researchers provided 16 autonomous cars with the new algorithm in order to test it. The cars were then placed in an environment with hundreds of other cars that were driven by humans.
"The autonomous vehicles were not in direct communication but ran the proposed algorithm in parallel without conflict or collisions," said Alyssa Pierson, a postdoc at CSAIL and first author on the new paper. "Each car used a different risk threshold that produced a different driving style, allowing us to create conservative and aggressive drivers. Using the static, precomputed buffer zones would only allow for conservative driving, whereas our dynamic algorithm allows for a broader range of driving styles."