Third edition of self-driving car competition, the AI Driving Olympics, to occur at NeurIPS 2019
The Duckietown Foundation is excited to announce that the third edition of The AI Driving Olympics will be held Dec 13-14, 2019 at the Neural Information Processing Systems (NeurIPS) conference, the premiere machine learning conference in the world, which will be in Vancouver. In addition to being a fun and engaging event that is open and accessible, the major goals of the AI Driving Olympics are to offer a standardized baseline for comparison of learning-based and classical robotics approaches to self-driving cars. Learn more at https://driving-olympics.ai/.
The AI Driving Olympics debuted at NeurIPS 2018 on December 7, 2018. AI-DO 1 leveraged the standardized hardware and simulation environment of Duckietown, a miniature self-driving car platform used for autonomy education and research that originated at MIT in 2016. AI-DO 2 took place in conjunction with the international conference on robotics and automation (ICRA), the flagship conference in robotics, in May 2019 in Montreal. AI-DO 2 comprised 4 challenges of increasing complexity related to urban driving: 1) Drive down the road; 2) Drive down the road while avoiding other vehicles and obstacles; 3) Drive down the road and through intersections while avoiding other vehicles and obstacles; and 4) Assign tasks in a large-scale ride sharing simulation (“autonomous mobility on demand”).
This edition of the AI-DO will have three completely different events, representing much more diversity than at previous AI-DOs.
Event 1 – Urban Driving: As in previous editions, the urban driving event will use the Duckietown platform. Competitors will have access to a suite of professional development tools (simulators, logs, baseline implementations) as well as tutorial webinars and community communication platforms. Hardware kits are available through the Duckietown Foundation website for competitors who wish to do local testing in their own robotic environments. In addition or as an alternative, controlled testing environments (Duckietown robotariums) in several locations worldwide will be available to run and evaluate submissions.
Event 2 – Racing: This challenge focuses on simulation to real transfer (sim2real) across perceptual and dynamic changes for racing with AWS DeepRacer. Sim2Real is a major roadblock to large scale deployment of Reinforcement Learning (RL), hence we aim to promote research in this area. For NeurIPS 2019, we focus on generalization of RL models, first in a virtual simulation environment, second on a physical racing track with texture and lighting variations. We provide participants with a cloud hosted simulation environment, an image dataset of racing tracks, and an RL training environment on Amazon SageMaker.
Event 3 – Advanced Sensing: This event focuses on the perception tracking problems based on autonomous vehicle driving data from the nuScenes by Aptiv dataset. At CVPR 2019 we organized the nuScenes detection challenge. For NeurIPS 2019 we focus on tracking all traffic participants in the scene. To this end, users can use any of the 6 cameras, 5 radars and 1 lidar in nuScenes, as well as the semantic map. We provide the data, devkit and evaluation server.
The AI-DO 3 is presented in collaboration with three academic institutions: ETH Zürich (Switzerland), Université de Montréal (Canada), Toyota Technological Institute at Chicago (USA), as well as two industry co-organizers: Aptiv and Amazon Web Services (AWS).
The competition is will officially open on Oct. 28, 2019. Further information about the competition is available at https://driving-olympics.ai/.
AI-DO press kit materials available here.