In the field of self-driving cars that change quickly, it is very important for perception systems to be accurate and strong. Recently, there have been big improvements in how birds see things and in a technology called Light Detection and Ranging (LiDAR), which helps cars understand their surroundings better. But we don't know enough about how well these new ways of understanding the world around us work when things get tough or different.

The RoboDrive Challenge at the IEEE International Conference on Robotics and Automation (ICRA) 2024 aims to improve the limits of reliable self-driving car perception. This competition is one of the first to test the strength of cutting-edge self-driving car perception models when dealing with unusual situations and problems with sensors. There are five tracks to this challenge, each focusing on a different way of helping cars understand what's happening around them, including BEV Detection, Map Segmentation, Occupancy Prediction, Depth Estimation, and Multi-Modal BEV Detection.

Track 2 - Robust Map Segmentation requires contestants to use advanced machine learning algorithms to perform accurate map segmentation on high-resolution bird’s-eye views. This task involves a detailed analysis of various urban geographic features, such as segmented lanes, sidewalks, and green spaces. In addition, this track also tests contestants’ image segmentation capabilities under different lighting, weather conditions, and noise conditions.

The ICRA 2024 RoboDrive Challenge has attracted widespread attention from academia and industry, gaining 140 team registrations from 93 institutions (universities, companies, etc.) in 11 different countries. Among all participants, the Advanced Research Lab of Samsung R&D Institute China – Beijing (SRC-B) has achieved promising results on Track 2 - Robust Map Segmentation.

In this challenge, a team of researchers from SRC-B explored several methods to improve the robustness of the map segmentation task. They drew significant findings by conducting large-scale experiments, as summarized below:

By analyzing the impact of different configuration options on corruption robustness, the recipes for a robust map segmentation model were found to possibly include utilizing a temporal fusion module and strong backbone (e.g., Swin-Base Transformer). Moreover, some data augmentation methods are effective in improving the robustness of map segmentation models. Nonetheless, further investigation into more advanced augmentation methods is warranted for future research. These novel findings earned SRC-B the “Innovative Honorable Award.”

In the future, SRC-B will use this technology in various scenarios where intelligent agents such as autonomous driving and robots can autonomously explore and interact with the surrounding environment, especially when exceptions occur in the sensor, to provide a solution to ensure the robustness of the algorithm.

SRC-B Team for the 2024 RoboDrive Challenge