physical-ai-neural-reconstruction

This is a thin "router" skill for NVIDIA's Neural Reconstruction (NuRec) ecosystem. Its job is not to run reconstruction workflows itself but to help an agent figure out which of NVIDIA's five upstream sibling skills — physical-ai-datasets, ncore, nre, asset-harvester, and nurec-fixer — answers a given request, then locate, clone, or refresh the canonical `nurec-skills` checkout from GitHub. It solves the discoverability and orchestration problem in a large, multi-repository toolchain by ordering multi-step NuRec pipelines (data conversion to training to rendering to cleanup) before the agent opens the upstream recipe. By design it never copies or reconstructs the actual commands, which live upstream.

The skill provides picker tables that map user intents to sibling skills, an upstream-fetch recipe that uses only `git`, workflow ordering guidance, a "mix-ups" reference clarifying easily-confused concepts (NuRec vs NRE, built-in Difix vs standalone DiffusionHarmonizer, ncore vs nre ordering), maintenance notes for keeping the router in sync, and safe secret-verification steps (checking that tokens are present without echoing their values). It ships a benchmark report (BENCHMARK.md) documenting NVSkills-Eval results across security, correctness, discoverability, effectiveness, and efficiency.

The intended users are teams working with NVIDIA Physical AI, autonomous-vehicle simulation, and novel-view-synthesis pipelines who need to navigate NuRec tooling. Downstream execution requires Docker, the NVIDIA Container Toolkit, a GPU, an NGC API key, a Hugging Face token with gated PhysicalAI licenses accepted, and Python 3.10+ with huggingface_hub; optional CARLA/Isaac Sim/AlpaSim integration is supported over gRPC.