Crowdsourced Generation of Parking Maps Using Mass-Production Vehicle Data for Parking-to-Parking Automated Driving

Yudai Kato･Kazutaka Hayakawa･Hiroshi Okabe (Aisin)･Keisuke Maekawa･Iori Nagase (Panasonic Advanced Technology Development)

Parking to Parking (P2P), which enables automated driving from an origin parking space to a destination parking space including both public roads and parking environments, is an important capability for next-generation mobility systems. To realize P2P, this study proposes a crowdsourced mapping technique for parking environments using sensor data collected from mass-production vehicles. Even under GNSS-denied conditions such as indoor and underground environments, the proposed method integrates observations from multiple vehicles to generate maps enriched with both geometric and semantic information. Furthermore, the approach enables continuous map updates without relying on dedicated mapping vehicles. The proposed framework aims to achieve a low-cost and scalable mapping solution, and its system architecture and effectiveness are investigated.

Development of Cloud-Based Parking Map Generation Technology for Automated Valet Parking

Morihiko Sakano･Takahiro Sakai (Hitachi)･Yoshinobu Ogasawara･Hidehiro Toyoda (Astemo)

This paper presents a cloud-based parking map generation and high-speed localization technology for Automated Valet Parking (AVP). A dual-descriptor map is introduced, which combines deep learning-based features for high-precision map construction with handcrafted features for real-time localization on edge devices. In addition, a divide-and-conquer map generation approach is proposed to significantly reduce processing time for large-scale parking environments. Experimental results using real-world data demonstrate that the proposed system achieves sub-meter localization accuracy with processing times suitable for real-time AVP applications, confirming its practical applicability.

Development of a Shape Optimization Framework Using AI-Based Shape Generation for Electric Water Pump Design and Its Practical Implementation Challenges

Yuka Ibaraki (Aisin)･Toshiki Terabe (Toyota Motor)･Kyohei Kitamura･Shohei Nakai (Aisin)

A shape optimization method combining DeepSDF-based shape generation and surrogate modeling was investigated for an electric water pump. By leveraging the relationship between latent variables and performance, optimization and performance prediction of newly generated shapes were achieved. This report discusses the challenges identified during practical application, along with approaches to addressing these issues and future perspectives for improving the applicability of the method.

Study on Performance-Driven 3D Shape Generation Using Latent Space

Keisuke Ootake･Hiroaki Suzuki･Yusuke Hasaba (Yamaha Motor Engineering)･Kiyofumi Shijo･Hibbanurrohim Ibram Muhammad (Astraea Software)

This study proposes a method for 3D shape generation and design support using latent space. A hybrid generative model combining a variational autoencoder (VAE) and a diffusion model is developed. By mapping CAE analysis results of the generated 3D shapes onto the latent space, the method enables inverse exploration of shapes based on target performance and demonstrates the potential for novel shape generation. The effectiveness of the proposed approach as a design support method is also confirmed.