Dynamic Map Platform Co., Ltd. (Head Office: Shibuya-ku, Tokyo; CEO & President: Shuichi Yoshimura; “Dynamic Map Platform”) presented at the Voxel51-hosted online event “ADAS, AV, and AI Meetup,” held on September 17, 2026 (Pacific Time; September 18 in Japan).
In a session titled “From Survey-Grade Maps to Physical AI: Scaling Real-World Data for Training and Simulation,” Dynamic Map Platform introduced initiatives that apply High-Precision 3D Map Data and point cloud data to AI training and evaluation. The presentation also showcased examples of extending real-world data into AI development, including the generation of photorealistic driving videos using High-Precision 3D Data and the creation of digital twin environments using 3D Gaussian Splatting (3DGS) and NVIDIA Isaac Sim.
■Real-World Data for Driving Video Generation and Digital Twin Environments
The presentation introduced Dynamic Map Platform’s work on simulation and synthetic data generation using High-Precision 3D Map Data and point cloud data.
In the first case, scenario videos generated in a simulation environment using High-Precision 3D Map Data were converted into photorealistic driving videos with the NVIDIA Cosmos Transfer 2.5 AV Model.
By generating videos under different weather and time-of-day conditions, Dynamic Map Platform demonstrated the potential to efficiently create diverse data for AI training and evaluation.
Dynamic Map Platform also presented its work on 3D Gaussian Splatting (3DGS), a technology for reconstructing surrounding environments in three dimensions from numerous images captured in the real world. By combining 3DGS with the High-Precision 3D Data developed for autonomous driving, Dynamic Map Platform is working to create digital twin environments that reflect real-world locations.
The presentation showed a digital twin environment in which driving scenarios generated from High-Precision 3D Map Data, real-world environments reconstructed with 3DGS, vehicle models, and other assets were integrated in OpenUSD and simulated in NVIDIA Isaac Sim. By combining vehicle models and driving scenarios, Dynamic Map Platform demonstrated how diverse conditions can be reproduced in environments grounded in the real world for use in autonomous driving AI and Physical AI training and evaluation.
Dynamic Map Platform also introduced a vehicle dynamics simulation use case using OpenCRG data, which reproduces road surface geometry from high-density, high-precision point cloud data.
■Real-World Data to Support AI Training and Evaluation
The presentation also introduced Dynamic Map Platform’s efforts to use High-Precision 3D Map Data and point cloud data, originally developed for autonomous driving, as data for AI training and evaluation. High-precision 3D map data contains information on real-world features such as lane markings, road edges, and traffic signs, together with lane centerlines, road connectivity, and road geometry. This information can be used as training labels for AI and as ground truth for comparing and validating recognition results.
Dynamic Map Platform also presented semantic data created by adding meaning to image and point cloud data. Combining raw data captured from the real world with semantic data can help streamline annotation work and support the evaluation and validation of recognition accuracy in AI development.
Building on the High-Precision 3D Data developed for autonomous driving, Dynamic Map Platform is advancing its Data for AI business initiative.
Part of the data used in the presentation is available as a sample dataset on the AI development platform Hugging Face.
■Voxel51 Demonstration Using Dynamic Map Platform’s Public Dataset
Dynamic Map Platform’s public sample dataset is also available through the Voxel51 Dataset Zoo. During the seminar, Voxel51 demonstrated how the dataset can be visualized and analyzed in its multimodal data platform.
The demonstration combined multiple camera views, semantic data, and High-Precision 3D Map Data, highlighting how AI developers can efficiently analyze and evaluate real-world data.
With Dynamic Map Platform’s data now available through the Voxel51 Dataset Zoo, AI researchers and developers around the world can more easily access the dataset.
■Presentation Archive
The recording of the presentation published by Voxel51 is available at the following link:
Dynamic Map Platform will continue to leverage the expertise it has developed in High-Precision 3D Data for autonomous driving and advance the provision of data infrastructure for AI training, evaluation, and simulation through its Data for AI business initiative.
＜Event Overview＞
| Event Name | ADAS, AV, and AI Meetup |
| Date & Time | September 17, 2026, 9:00 AM – 11:00 AM (PDT) September 18, 2026, 2:00 AM – 4:00 AM (JST) |
| Format | Online (Zoom) |
| Participation Fee | Free |
| Speaker | Ryoto Miyake (Software Technology Department, Technology Unit) |
| Organizer | Voxel51 |
