Changing lanes in a car is a relatively straightforward process for human drivers. But autonomous cars must process a massive amount of information to do it safely, with the car relying on detailed statistical models of the driving environment or making extremely conservative decisions, such as not changing lanes at all.
Now, researchers from the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory (CSAIL) are trying to fix this problem using a new lane-change algorithm that allows for more aggressive lane changes and more human-like behaviour.
Autonomous cars calculate buffer zones around other vehicles, which describe where the vehicles are and where they are likely to be in a certain timeframe. The cars make decisions based on staying out of another vehicle’s buffer zone.
However, these buffer zones are pre-computed and can be restrictive, particularly in fast or busy traffic. This can result in situations where safety stops a vehicle from changing lanes at all.
“Many of the current lane-change algorithms assume a single style of driving for the autonomous vehicle or require computationally intensive algorithms that become intractable in congestion,” said Alyssa Pierson, a Postdoctoral Associate at CSAIL.
CSAIL’s algorithm is designed to quickly plan a sequence of lane changes for an autonomous vehicle and allow users to choose the amount of risk an autonomous vehicle takes on based on the behaviour of other drivers around the autonomous vehicle.
“First, we look at planning the sequence of lane changes, whereas existing lane-change algorithms deal with a single lane change. By planning for a sequence of lane changes, we look ahead and create a ‘weaving through traffic’ type of behaviour,” Pierson said.
“Second, we only require knowing other cars’ position and velocity to calculate the risk. This allows us to calculate the autonomous vehicle’s actions online without communication with other vehicles.”
Checking blind spots
One of the challenges of the CSAIL team’s research has been finding an algorithm that has rich emergent behaviours for the vehicle, while also being computationally efficient.
“Our key insight is how we choose to quantify risk. Based on our formula for calculating risk, we are able to prove that cars have a maximum allowable risk threshold, and any actions that remain below that threshold will avoid collisions,” Pierson said.
“A main priority was to maintain collision avoidance guarantees for the vehicle while incorporating riskier behaviours for the vehicle.”
CSAIL’s work is still in the prototype stage, with the team currently extending the scope of their research to experiments on autonomous vehicles, validating on human data and increasing the complexity of scenarios in simulation.
This article was originally published as “AI in the fast lane” in the November 2018 issue of create.