Follow
Subscribe

Hyundai's Self-Driving Arm Motional Releases the World's Largest Open Data Set For Autonomous Vehicle Path Planning

Home > News > Content

【Summary】​Autonomous driving joint venture Motional, which was formed by Hyundai Motor Group and Aptiv to develop autonomous driving technology for the automaker as well as ride-hailing company Lyft, announced the launch of the initial version of a new open dataset called “nuPlan”, which the company says is the world’s largest public dataset for autonomous vehicle prediction and planning.

Eric Walz    Dec 10, 2021 10:00 AM PT
Hyundai's Self-Driving Arm Motional Releases the World's Largest Open Data Set For Autonomous Vehicle Path Planning

Autonomous driving joint venture Motional, which was formed by Hyundai Motor Group and Aptiv to develop autonomous driving technology for the automaker and as well as ride-hailing company Lyft, announced the launch of the initial version of a new open dataset called "nuPlan", which the company says is the world's largest public dataset for autonomous vehicle prediction and planning.

The new open-source planning dataset will allow researchers to better understand how a driverless vehicle can find its way through a dynamic environment, including driving in the city like a human driver. Motional says the nuPlan data set is the world's first benchmark for autonomous vehicle planning. 

For developers of autonomous vehicles, this type of data is used to train machine learning algorithms so self-driving vehicles can navigate safely. The data in nuPlan will be available to the public and can be used to teach an autonomous vehicle how to handle unique driving situations. 

While ML-based planning has been studied extensively, the lack of published datasets that provide a common framework for closed-loop evaluation has limited progress in this area, according to Motional. The company believes that nuPlan will help fill this gap by providing an ML-based planning dataset, closed-loop evaluation, and planning related metrics.

The initial release of nuPlan and follows the release of "nuScenes" in March 2019. nuScenes is composed of millions of photos and data-points collected from the vehicles' full sensor suites, were then hand-annotated, and used to inform and advance machine learning models to build the safest possible self-driving vehicles.

nuPlan, builds upon nuScenes. It's essentially a "virtual driving test" that contains a large-scale machine learning dataset and a toolkit for measuring the performance of motion planning techniques for self-driving vehicles. 

The dataset contains 1,500 hours, or 4.7 years of average driving data that was collected across four different cities where Motional is testing its robotaxis: Boston, Pittsburgh, Las Vegas, and Singapore. It contains around 500 million images and 100 million lidar scans.

The dataset includes all of the complexities that human drivers encounter every day in urban areas, such as navigation in busy intersections and identifying road signs and following traffic laws It also includes extreme real world edge-case scenarios that are only experienced roughly one time every 1,000 hours of driving. 

These rare scenarios can include a shopping cart in the middle of the roadway. Although these situations are not experienced everyday, a self-driving vehicle still needs to learn how to deal with it. No humans are annotating this real world data, rather its being entirely done by machine learning, with nearly the same quality, according to Motional.

Another significant advantage of nuPlan is its "closed-loop" testing capability. With an open-loop system, the input is independent of the system's response, regardless of the system's behavior. 

Open-loop is sometimes called imitation learning, since the system simply checks that the planned route is similar to the one the driver took. In closed-loop evaluation however, the planned route is used to control the vehicle. The vehicle may deviate from the original route and other drivers will then react accordingly.

Motional says that autonomous vehicles are currently at a crossroads between perception and planning and nuPlan will help fill this gap.

In 2018, Motional said its machine learning team examined the competitive landscape in the autonomous driving space and found that most of the attention was focused on perception, which included teaching AVs how to identify cars, pedestrians, bicycles and other objects in camera images and lidar scans in order to gain a better understanding of their surroundings.

While perception allows an autonomous vehicle with the ability to see the world, the nuPlan dataset helps a vehicle more safely navigate it. Now that perception has improved to allow autonomous vehicles to more accurately identify what is around them, Motional is now focused on autonomous vehicle path planning with nuPlan, which the company believes is one of the final frontiers in autonomous driving.

Motional says it was the first to make its vast trove of data publicly available. Sharing data between companies can help accelerate the development of autonomous vehicles through collaboration. 

Since the release of nuScenes in 2019, Motional said its been downloaded by more than 16,000 students, researchers and developers so far. It has also been referenced in more than 1,000 academic publications. 


Prev                  Next
Writer's other posts
Comments:
    Related Content