NVIDIA has published nurec-skills, a repository of six agent skills that teach a coding agent to drive the NuRec neural reconstruction stack. Each skill is a single SKILL.md file with YAML frontmatter following the agentskills.io convention. These agents are thin coordination layers that redistribute no NVIDIA source and instead teach an agent to drive the public NGC containers, GitHub repositories and Hugging Face artifacts that make up the stack.
nurec-index picks the right sibling skill for a task, and physical-ai-datasets catalogs NVIDIA's PhysicalAI datasets on Hugging Face. ncore converts camera, LiDAR, radar, IMU, depth and stereo recordings into NCore V4 and is pinned to upstream 2026.04. nre trains 3DGUT and 3DGRT gaussian splat reconstructions, renders novel views locally or over gRPC, exports PLY, mesh and depth, and edits actors, pinned to NRE release_26.04 against the nvcr.io/nvidia/nre/nre and nvcr.io/nvidia/nre/nre-tools containers.
asset-harvester extracts per object 3DGS assets from sparse AV clip views using SparseViewDiT and TokenGS, and Asset Harvester was described in NVIDIA's SIGGRAPH NuRec libraries announcement. nurec-fixer post processes, evaluates or fine tunes novel view renders with DiffusionHarmonizer, after NVIDIA released Omniverse NuRec Fixer at NeurIPS to correct blurs, holes and missing geometry in NeRF and gaussian splatting workflows. The repository cites arXiv 2604.18468 for Asset Harvester and arXiv 2602.24096 for DiffusionHarmonizer.
Prerequisites are Linux x86_64 with CUDA 12.x drivers and an Ampere or newer GPU, with 16 GB of VRAM named as the practical floor for harmonizer inference and 24 to 48 GB recommended for NRE training. aarch64 is not supported by the NRE containers.
NVIDIA is not accepting external pull requests, and the repository carries no releases and no tags. It is available on GitHub.
