News & Events

Share         

​MAE AutoMan group won autonomous vehicle global challenge in top AI competition

Published on: 18-Dec-2020

Congratulations to our MAE AutoMan Team: Phd candidate Mr. Mo Xiaoyu, research fellow Dr. Xing Yang, supervised by Nanyang Asst. Prof. Lyu Chen, for winning the autonomous driving INTERPRET Challenge in top AI conference “Neural Information Processing Systems (NeurIPS)”. The result was recently announced during the NeurIPS week in USA in December 2020. 


About the INTERPRET challenge
Picture1.png 
Fig.1 The INTERPRET Challenge in NeurIPS 2020. 

The INTERPRET (INTERACTION-Dataset-Based PREdicTion) Challenge is a NeurIPS 2020 global competition, organized by the Mechanical Systems Control Lab of UC Berkeley. This challenge provided real-world traffic data collected from different locations worldwide and aimed to find machine learning solutions to trajectory prediction methods for autonomous driving. 

The proposed deep-learning-based trajectory prediction for AVs 

Autonomous driving is receiving more and more interest recently in both academia and industry. Trajectory prediction is of great importance for decision making of autonomous vehicles. However, trajectory prediction is challenging because driving is an interactive activity, and the driving behavior of a vehicle is affected by multiple factors. 

The AutoMan team proposes a deep learning model with heterogeneous graph, named ReCoG, to handle the challenge. As shown in Fig.2, the ReCoG model adopts the Encoder-Decoder structure and jointly uses Recurrent Neural Networks, Convolutional Neural Networks, and Graph Neural Networks as encoders. It models the relationships among vehicles and infrastructures as a directed heterogeneous graph, where different nodes represent different objects, and the edge from one node to another one indicates its impact on the other node. Finally, the dynamics feature of the target vehicle is concatenated with the interaction features and sent to an RNN decoder for prediction. ReCoG jointly considers the impacts of surrounding vehicles and the infrastructure on a target vehicle’s driving behavior. It outperformed all competitors in the challenge. 
Picture2.png
Fig. 2. The deep learning framework with a heterogeneous graph of the ReCoG model. 

The AutoMan group 

The Automated Driving and Human-Machine System (AutoMan) Group at MAE, directed by Nanyang Asst. Prof. Lyu Chen, focuses on automated driving, human-machine systems, human-in-the-loop AI, intelligent electric vehicles, and cyber-physical systems. 

More info: 
The research paper: https://arxiv.org/pdf/2012.05032.pdf 
The INTERPRET challenge: http://challenge.interaction-dataset.com/prediction-challenge/intro 
AutoMan group site: https://lvchen.wixsite.com/automan 

Back to listing