Spatial AI stands as a pivotal element in autonomous vehicles, crucial for ensuring the safety of passengers through advanced 3D situational comprehension and vehicle control requiring high computational power. The challenge is real: as autonomous vehicles travel, they cannot replenish energy, making the consumption of computing power a direct factor in reducing driving range. To address this, our research is intensely focused on developing energy-efficient computational methods. Our innovative approach involves a unique HW-SW co-design and a sensor-friendly architectural design, leading to the creation of specialized AI SoCs (System on Chips) for autonomous driving. We’re dedicated to pushing the boundaries of technology, ensuring safer and more efficient autonomous journeys.”

.

Recently, the rapid growth of interest in autonomous vehicles (e.g., drones, robots, cars) has actively led the world’s market. For the safety of both drivers and pedestrians, the 3D perception that recognizes the surrounding environment has aggressively progressed over the past decade as an essential function of autonomous driving technologies. The advancement of 3D sensors (e.g., depth camera, radar, LiDAR) also propelled the performance of 3D perception by directly providing accurate distance information together with the shape of objects. Over the past few years, research utilizing LiDAR has drawn attention due to its ability to provide 3D information (a.k.a. point cloud) without data loss, enabling enhanced accuracy. The challenge arises when simply adopting the traditional 2D CNN(convolutional neural network) processors because of the properties of point cloud. Point cloud is unstructured and unordered data that is distributed in 3D space. Therefore, each point data does not contain neighbor information like 2D array image pixels. As a result, the 3D PNN(point cloud neural network) processors should be designed to support different functions with CNN.