NVIDIA’s 3DGRUT toolkit has received an exciting update via PR #208, introducing native support for NCore v4 datasets alongside new LiDAR initialization capabilities and dataloader profiling tools. The changes span documentation updates, code formatting, bug fixes, and new features that expand 3DGRUT’s data ingestion pipeline.
The PR adds training configuration and dataset support for NVIDIA’s open source NCore v4 data format. Users can now point 3DGRUT training directly at NCore v4 sequence JSON files, with training defaults provided in a new ncore.yaml configuration. The NCore data format and tooling are maintained in the separate NVIDIA/ncore repository.
NCore is NVIDIA’s standardized format for multi sensor capture data used across their Physical AI and autonomous vehicle research. Bringing NCore v4 support into 3DGRUT bridges the gap between NVIDIA’s data infrastructure and their open source radiance field workflows.
A new configurable observation points option has been added to the LiDAR initialization pipeline. The update also fixes a point duplication bug that occurred when point counts fell below the upper bound threshold, and introduces batched observer point determination for NCore datasets. LiDAR initialization is increasingly important as radiance field methods move beyond RGB only inputs toward multi modal sensor fusion.
Additionally, they have added a new generic dataloader profiler script has been added with CUDA timing support, enabling developers to benchmark data loading performance.
