MIT's Stunning Driverless Car AI System

 AI System That Real-Time Forecasts Motion Around the Car

                                                                

                                                                        Source: MIT

M21 Appears to be Most Accurate Autonomous Driving System 

The next big thing in the automotive world is likely to be the spreading use of autonomous cars and trucks.  On the streets of Beijing, Chinese tech/AI company Baidu and California based autonomous car company Pony.AI have become the first robotaxi service in China to operate without backup human drivers. In the US, Alphabet's Waymo and GM's Cruise are also operating without human drivers in San Francisco and Arizona. The key to safety is the car's ability to predict the next movements of vehicles and people simultaneously around the vehicle.  Researchers at MIT may have a breakthrough system.  They have developed a new machine-learning system called M21 that helps driverless cars predict - in real time - the next moves of the vehicles, pedestrians and cyclists around it and thereby operate more safely when navigating city streets.  

Deceptively Simple

The MIT team says that their M21 system is "deceptively simple" in predicting some of the most complex behavior on the planet - the highly unpredictable and complicated next moves of vehicles, pedestrians and cyclists in traffic.  They took complex, multi-agent behavior road conditions and broke them into smaller components so a computer can solve and predict next moves - individually and in real-time.  MIT says that their M21 autonomous driving system is highly accurate.  In fact, they believe it is more accurate and more efficient in real-time than any other autonomous system now in operation.

How Does It Work?

To put it simply, MIT researchers trained their model using the Waymo Open Motion Dataset which has millions of real traffic scenes recorded by lidar, cameras and sensors. M21 first determines the relationship between two road users.  Specifically, which car, pedestrian or cyclist has the right of way and who will yield.  It then uses the information to predict future movement by multiple vehicles, pedestrians and vehicles.  Essentially, there are two big sets of data:  past trajectories of cars, pedestrians and cyclists at an intersection and a map of the intersection and lane configuration.  It predicts how the players will move through traffic for the next 8 minutes.  

Predicting the Nearly Unpredictable

What autonomous cars fundamentally need are highly reliable and accurate movement prediction systems to keep the vehicle safe from collisions. In essence, that is an AI system that can predict what is nearly unpredictable.  It appears that the MIT M21 is closing in on that goal.  The researchers say that by comparison M21 is more accurate and efficient than Waymo, which many consider the world's leader in autonomous vehicles.  For more news stories like this, go to Daily Vehicle Briefing