Dynamic Map Platform Launches Dataset Business for Physical AI, Releases Sample of AI-Native Data Integrating High-Precision 3D Data

Dynamic Map Platform Co., Ltd. (Headquarters: Shibuya-ku, Tokyo; President and CEO: Shuichi Yoshimura; hereinafter "the Company") has published a dataset targeting intersections as a sample of data designed for AI use (hereinafter "AI-native data") on Hugging Face, a machine learning community platform where AI developers worldwide share models and datasets.

This dataset is a multimodal integration of point cloud data, camera images, high-precision positioning information, high-precision 3D map data, and 3D Gaussian Splatting (3DGS) data, including information on locations with high accident risk.

This initiative aims to concretize the Company’s dataset business for physical AI and marks the full-scale launch of its "Data for AI" offering.

Public URL: https://huggingface.co/datasets/dynamic-maps/hard-intersection-multimodal-sample

Sample image of AI-native data (intersection data integrating point clouds, images, 3DGS, etc.)

To date, the Company has worked to build data infrastructure reflecting the real world across various fields, including autonomous driving, through the development and provision of high-precision 3D data. Simultaneously, it has advanced its "Data for AI" initiative, accumulating expertise on data formats beneficial for AI development.

In recent years, as AI technology has advanced, the application of AI targeting the physical world (physical AI) has expanded, increasing the importance of high-precision data designed for specific use cases.

Against this backdrop, the Company will further expand its "Data for AI" initiative, leveraging its accumulated expertise to promote the development and provision of AI-native data.

The AI-native data the Company aims for refers to high-precision 3D data designed and optimized for immediate use on platforms, suitable as training data for AI model development or as digital twin data in simulations.

Typically, AI learning for autonomous driving systems relies on real-world driving data from dashcams and similar sources, but collecting large volumes of data and rare scenarios involves significant time and cost. In response to these challenges, the importance of utilizing virtual environment data has grown.

The sample data released this time is a multimodal AI-native dataset leveraging the Company’s long-accumulated high-precision 3D data assets. It integrates point cloud data, multi-view camera images, high-precision positioning information, trajectory data, high-precision 3D map data, semantic annotations (data with meaning), and 3D Gaussian Splatting (3DGS) data in a temporally and spatially aligned format.

This enables the creation of highly accurate real-world-replicated environments for training and evaluation. Additionally, geographic feature information in map data can be used as annotations, contributing to enhanced spatial recognition by AI and reducing the Sim2Real gap (differences between real and virtual environments). Furthermore, thanks to 3DGS data, digital twins with realism close to the real world can be constructed, making this format consistently usable not only for AI model training but also for evaluation and validation of AI systems through simulation.

This sample targets real urban intersections with high accident risk, making it useful for advanced scene understanding and safety verification. It is particularly valuable for autonomous driving system development and is expected to be widely applied in AI use cases across industries such as infrastructure management, urban development, traffic flow analysis, and disaster prevention and response.

This sample release is positioned as the first phase. Moving forward, the Company will expand its data lineup and continue regular releases, while also working on data design and development with an eye toward productization.

The Company will continue to promote the development and provision of AI-native data centered on "Data for AI," advancing sophisticated data utilization for the era of physical AI.

[Reference]

The Company’s official website features a column by engineers on the progress of AI-native dataset development. This column technically explains the initial verification of 3DGS generation in the released sample data (prototype). Please take a look.

COLUMN "Next Stage of 3DGS — Redefining Spatial Generation with Multimodal Data" https://www.dynamic-maps.co.jp/column/column-1650/

＜About Dynamic Map Platform Co., Ltd.＞

Dynamic Map Platform Co., Ltd. was established with backing from the Japanese government and investment from 10 domestic automobile manufacturers. Headquartered in Japan, the Company has bases in North America, Europe, the Middle East, and South Korea, and currently operates in 26 countries. It provides high-precision 3D data for a wide range of applications, including autonomous driving and advanced driver assistance systems (ADAS), simulator environment construction, infrastructure management, and snow removal support.

With the vision of "Modeling the Earth" — digitizing the planet — the Company acts as a high-precision 3D data platform, co-creating innovation across various industrial fields.

Established: June 2016

Headquarters: Shibuya-ku, Tokyo

Representative: Shuichi Yoshimura

Business: Provision of high-precision 3D data for diverse industries, including autonomous driving and ADAS

Official Website: https://www.dynamic-maps.co.jp/

Official X: https://x.com/dynamic_maps