Driverless startup Cruise today detailed a homegrown tool — the Continuous Learning Machine — that tackles on-the-road prediction tasks. Cruise claims the Continuous Learning Machine, which automatically labels and mines training data, allows some of the AI models guiding Cruise’s self-driving cars to predict things like whether bicycles will swerve into traffic or kids will run into streets.
One of the challenges of autonomous vehicles is predicting intent. People don’t always follow the rules of the road, and even when they do, they’re liable to bend those rules. According to the U.S. National Highway Traffic Safety Administration, 94% of serious crashes are due to drivers’ errors or dangerous choices.
That’s why Cruise built Continuous Learning Machine. Leveraging a technique called active learning, it automatically identifies errors made by perception models running on Cruise’s cars, and only scenarios with a significant difference between prediction and reality are added to the training data sets. Cruise says this enables extremely targeted data mining, minimizing the number of “easy” scenarios that enter the corpora.
The Continuous Learning Machine also labels data autonomously using model predictions as “ground truth” for all scenarios. Essentially, the framework observes what a person or vehicle might do in the future and compares that against what they actually end up doing. The final step is training a new model, running it through testing, and deploying it to the road while ensuring performance exceeds that of the previous model.
Cruise says the Continuous Learning Machine has enabled it to make highly accurate predictions for a number of rare scenarios its models encounter in the real world. These include U-turns, which Cruise’s cars see fewer than 100 times a day, on average, and cut-ins, when people change their trajectory to avoid slowing or stationary objects. Another example is K-turns — three-point turns that require drivers to maneuver forward and in reverse. Cruise says these are about as half as common as U-turns.
“Our machine learning prediction system has to generalize to both completely novel events as well as events that it sees very infrequently,” Cruise senior engineering manager Sean Harris wrote in a blog post. “We need to understand both the intent of other agents on the road and reason about the sequence and interactions between different agents and how they will evolve over time. The complexity of this problem is its own field of research, which is another reason why autonomous vehicles are the greatest engineering challenge of our generation.”
The audio problem: Learn how new cloud-based API solutions are solving imperfect, frustrating audio in video conferences. Access here