NVIDIA NuRec Neural Reconstruction Router
Routes NuRec requests to the right upstream workflow skill.
The evidence shows that the skill declares only Read/Shell, discloses Git, Docker, NGC, Hugging Face, GPU, and token dependencies, and provides safe verification, license-acceptance, and secret-handling guidance. It also requires reading upstream skills before mutating operations. Points are deducted because cloning, pulling, container execution, and potentially very large disk usage have no explicit per-action user confirmation, isolation boundary, or general rollback plan; some downstream permissions are delegated to external skills.
The evidence shows reasonably complete routing tables, workflow ordering, troubleshooting guidance, and limitations, plus reported static validation and task-evaluation results in BENCHMARK.md. Points are deducted because no key path was executed in this review; upstream content can drift and the skill depends on external repositories, containers, GPUs, and gated data. Evidence for diagnosable behavior on malformed input or unreachable upstream services is limited, so the score remains within the static ceiling of 10.
The evidence clearly identifies audiences, trigger terms, sibling-skill mappings, workflow scenarios, and non-fit boundaries, including explicit exclusions for SimReady and infrastructure deployment. Points are deducted for the NVIDIA-only hardware/toolchain requirements, dependence on GitHub, Hugging Face, and NGC, lack of Chinese-language or mainland-China reachability guidance, and the fact that the router itself cannot perform training, conversion, or rendering.
The evidence provides front matter with name, description, Apache-2.0 license, version, author, compatibility, and upstream metadata. The body includes progressive routing, workflows, limitations, troubleshooting, secret handling, maintenance, and teardown documentation. Points are deducted because BENCHMARK.md reports missing recommended Instructions and Examples sections and a nonconforming author format; maintenance responsibility is attributed to the NRS team, but maintainer contact, changelog, and a pinned upstream-version guarantee are absent.
The evidence shows that the skill can directly route NuRec requests and specify the sequence involving physical-ai-datasets, ncore, and nre; evals.json and BENCHMARK.md report positive routing, correctness, and effectiveness results. Points are deducted because this is a thin router whose commands and deliverables require separately fetched upstream skills, and the static source set contains no independently verifiable representative output, so the score stays near but below the static ceiling.
The evidence includes revision-scoped source material, upstream paths, multiple cited source declarations, evals.json, and BENCHMARK.md. Points are deducted because the benchmark is primarily an author-provided summary, with no committed CI workflow and test suite that independently reproduces key paths, and no cross-source corroboration verified during this static review.
- Core functionality depends on external upstream repositories, NGC containers, and gated Hugging Face data; it may be unusable on mainland-China networks or without an NVIDIA GPU.
- Before running upstream workflows, confirm tokens, licenses, network access, disk capacity, and container permissions individually, and check for upstream version drift.
- The benchmark report is a claim contained in the supplied source files and does not replace independent execution verification in this static review.
What does this skill do, and when should you use it?
This is a thin router for NVIDIA Neural Reconstruction (NuRec); it does not directly execute training, rendering, conversion, or cleanup commands. It maps user goals to upstream skills for datasets, NCore conversion, NRE, asset harvesting, or frame harmonization, and orders multi-step workflows. NuRec turns camera, LiDAR, radar, or stereo recordings into 3D scenes that can be rendered from new viewpoints. Downstream workflows generally require an NVIDIA GPU, container tooling, an NGC API key, and a Hugging Face token with the relevant gated licenses accepted.
Reads a NuRec-related request and selects the appropriate upstream sibling skill; clones or refreshes https://github.com/NVIDIA/nurec-skills with Git; reads the upstream index and selected sibling SKILL.md; and sequences NCore V4 conversion, NRE training, USDZ rendering, object processing, frame cleanup, or simulator integration. It also documents safe credential checks, troubleshooting paths, scope boundaries, and cross-skill teardown considerations.
- An autonomous-driving or robotics team needs to determine whether a request belongs to dataset download, NCore conversion, or NRE training.
- A user has camera, LiDAR, radar, or stereo recordings and needs to plan a NuRec reconstruction workflow.
- A team wants to render new views or LiDAR sweeps from an existing USDZ, or connect rendering to CARLA, Isaac Sim, or AlpaSim over gRPC.
- A user needs to extract, replace, or clean up cars, pedestrians, or rendered frames in a NuRec scene.
- A user is confusing NuRec with SimReady packaging, generic USD tuning, or infrastructure deployment.
What are this skill's strengths and limitations?
- Provides clear routing across NuRec datasets, NCore, NRE, asset extraction, and frame cleanup.
- Explicitly separates NuRec from SimReady, generic USD performance tuning, and infrastructure deployment.
- Documents upstream refresh, safe credential checks, common failures, and cross-skill teardown.
- Uses standard SKILL.md metadata and can be installed through NVIDIA's skills CLI.
- It is only a router and does not itself train, convert, render, or repair artifacts.
- Core recipes live in a fast-changing upstream repository, so stale local clones can drift.
- A complete workflow may consume more than 150 GB and does not support AMD, Intel, or Apple Silicon.
- Several Hugging Face datasets and models are gated and require prior license acceptance plus a token.
- The supplied material does not show an independent test suite or platform coverage report for this skill.
How do you install this skill?
Install the specific skill with the skills CLI documented in the repository README:
npx skills add nvidia/skills --skill physical-ai-neural-reconstruction --yes
You can target a client explicitly:
npx skills add nvidia/skills --skill physical-ai-neural-reconstruction --agent codex --yes
The README states that the skill becomes available the next time the agent loads skills for a relevant task.
How do you use this skill?
After installation, give the agent a concrete NuRec request, such as: "I have camera and LiDAR recordings; which NuRec skill should convert and train them?" or "How should I connect USDZ rendering to CARLA over gRPC?" The router selects an upstream sibling skill; before any mutating operation, fetch and read that upstream skill's SKILL.md. This router does not supply directly executable training, rendering, or conversion recipes.
How does this skill compare with similar options?
Unlike a SimReady pipeline, NuRec turns sensor recordings into renderable USDZ scenes rather than packaging CAD or source meshes. Unlike a generic USD performance-tuning skill, this router is limited to NuRec-related dispatch. Unlike AKS, OSMO, or NIM Operator infrastructure skills, it does not deploy infrastructure.