physics-agent-cli

name	physics-agent-cli
description	Run the Physics Agent CLI for VLM-based physics property classification, tuning, and refine-loop workflows for 3D assets. Use when the user wants to run the physics-agent CLI directly, classify asset components, identify asset type, build datasets from USD files, predict component material/type/physics properties, author simulation-ready USD physics schemas, tune authored physics parameters, or run iterative refine.
version	0.1.2
author	NVIDIA Content Agents
tags	["content-agents","physics-agent","cli","usd","vlm"]
tools	["Shell","Filesystem","Python","wu"]
compatibility	Requires Linux, a Linux container, or WSL2; the physics-agent CLI; a repo Python environment; provider credentials for the selected VLM backend; a render endpoint for remote rendering configs; and a scene optimizer endpoint or local scene optimizer when optimize_usd is enabled.

Physics Agent CLI

The Physics Agent renders USD components, identifies the asset, predicts material/component/physics properties for each part, and can author UsdPhysics schemas into a simulation-ready output USD.

When to Use

• Use when the user asks to run physics-agent directly from the command line.
• Use when the user wants component classification for a USD asset.
• Use when the user wants to author rigid-body, collision, mass, and physics material schemas from VLM predictions.
• Use when the user needs the instanced-USD/deinstance path for writable physics authoring.

Route requests outside the local CLI to the exact sibling skill:

User intent	Required sibling skill
Call an already-running Physics Agent REST API, monitor its sessions, or download service artifacts	physics-agent-client
Build, start, stop, or configure the local Physics Agent Docker Compose service	deploy-physics-agent-docker
Provision Brev-hosted render/VLM dependencies or run the Brev hybrid Physics Agent workflow	deploy-physics-agent-brev
Assign or refine visual/PBR materials rather than physical properties	material-agent-cli

Limitations

• Keep secrets out of chat and commits. Tell the user to set provider keys in their local environment or repo-root .env; never ask them to paste keys.
• Official runtime targets are Linux, Linux containers, and WSL2. Native Windows shell execution is not supported; direct the user to WSL2 or Linux, not to a sibling skill that does not support the requested runtime.
• Config paths such as input.usd_path resolve relative to the config file, not the current shell directory.
• Instanced USD descendants are instance proxies and cannot be authored on directly. Use optimize_usd with deinstance enabled before apply_physics when predictions target instance proxies.
• restore_usd maps predictions back to original paths for reporting; when optimization runs, apply_physics authors onto the optimized/deinstanced USD using raw prediction paths.
• The classification pipeline does not apply visual/PBR materials or render a final beauty image. It stops after physics authoring unless the user invokes other tooling.