Autonomous Cars Should Have Better "Reasoning" to Navigate Tricky Intersections
Wed, April 21, 2021

Autonomous Cars Should Have Better "Reasoning" to Navigate Tricky Intersections

The model gauges the risk of potential collisions and other traffic disruptions / Photo Credit: Zapp2Photo (via Shutterstock)


MIT and Toyota researchers have developed a new model to help self-driving vehicles determine when it’s safe to “merge into traffic at intersections with obstructed views,” according to Rob Matheson of MIT News, a science and technology news platform. Automated systems that help driverless cars and human drivers navigate through intersections “require direct visibility of the objects” they need to avoid. These systems can fail when a vehicle’s line of sight is blocked by nearby structures or other obstructions. 

Hence, the model uses its own uncertainty to gauge the risk of potential collisions and other traffic disruptions. The system weighs several factors such as nearby visual obstructions, sensor noise and errors, the speed of other vehicles, and the attentiveness of drivers. Then, the system may advise the vehicle to pull into traffic, stop, or nudge forward to collect data. 

The system was tested in over 100 trials of remote-controlled cars turning left at a busy, obstructed intersection in a makeshift city, with other vehicles constantly driving “through the cross street.” Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science.  Daniela Rus, along with her colleagues, divided the road into small segments, allowing the model to detect if each segment is occupied to determine the likelihood of a collision. 

For example, when a sensor detects a passing car driving through a visible segment, the model uses the car’s speed to predict its prediction throughout the segments. A probabilistic “Bayesian network” considers other uncertainties like noisy sensors or abrupt changes in speed to determine the probability of each segment being occupied by a passing car.  

Notably, the risk estimate is updated continuously regardless of the car’s location in the intersection. Running the model on remote-controlled cars in real-time proved that it is fast and efficient enough to be deployed into full-scale autonomous test cars in the future, but the model still needs to undergo more rigorous testing first. “In this work, we use a predictive-control model that’s more robust to uncertainty, to help vehicles safely navigate these challenging road situations,” explained Rus.  

Rus and her team plan to include other challenging risk factors such as pedestrians in and around the road junction.