Agent Skills › isaac-sim/IsaacSim › isaac-sim-validator

isaac-sim-validator

GitHub

Isaac Sim仿真交付前的最终质量校验工具，检查脚本规范、USD场景结构及渲染输出。拦截黑帧、废弃导入、硬编码路径等问题，确保物理稳定与渲染质量达标。

skills/isaac-sim-validator/SKILL.md isaac-sim/IsaacSim

Trigger Scenarios

验证Isaac Sim生成的Python脚本或USD场景检查渲染帧或训练产物的合规性

Install

npx skills add isaac-sim/IsaacSim --skill isaac-sim-validator -g -y

More Options

Use without installing

npx skills use isaac-sim/IsaacSim@isaac-sim-validator

指定 Agent (Claude Code)

npx skills add isaac-sim/IsaacSim --skill isaac-sim-validator -a claude-code -g -y

安装 repo 全部 skill

npx skills add isaac-sim/IsaacSim --all -g -y

预览 repo 内 skill

npx skills add isaac-sim/IsaacSim --list

SKILL.md

Frontmatter

{
    "name": "isaac-sim-validator",
    "description": "Validate Isaac Sim simulation scripts and outputs before delivery. Rejects black frames, deprecated `omni.isaac.*` imports, hardcoded user paths\/IPs, missing lights, wrong render mode, undersized SimulationApp configs, and bad RL output formats. Use when a sub-agent produces a script, a render, a training artifact, or any USD scene that's about to be handed to the user."
}

Isaac Sim Validator

Final gate before delivery. Reject outputs that fail the checks below; do not soften results.

Universal checklist (apply to every output)

Check	Pass	On fail
Script path	exists, readable	reject: `Script path invalid: <path>`
File size	> 1 KB	reject: `Script empty or corrupted`
`isaacsim` import	uses `isaacsim.*`, not `omni.isaac.core`	reject: `Use isaacsim.* namespace; omni.isaac.core is deprecated`
`SimulationApp` config	explicit `width` and `height`	reject: `SimulationApp without window args causes swapchain-capture size mismatch`
Lights present	`DomeLight` (intensity >= 100) and `DistantLight` (>= 500)	reject: `No lighting; render will be black`
Render mode	`RayTracedLighting` for iteration; `PathTracing` only for hero shots	reject: `PathTracing too slow for iteration; use RayTracedLighting`
ACES tonemap	`/rtx/post/tonemap/op=4`, `filmIso` 200 default / 600 deep-aisle / 400 aerial	suggest: `Enable ACES (op=4) for contrast`
Final render	>= 150 KB and mean_RGB > 30	reject: `Render is black or low-energy; check lighting`
User paths	no `/home/<user>/`, `C:\Users\<name>\`, or per-agent home paths	reject: `Hardcoded user path; use $ISAAC_SIM_DIR, $ISAAC_LAB_DIR, $WORKSPACE_DIR, or argparse`
Network endpoints	no bare IPs (`192.168.x.x`, `10.x.x.x`) or non-localhost hosts	reject: `Hardcoded endpoint; parameterize via env var or config`
Output dir	configurable (env var, CLI, or function param)	reject: `Output dir hardcoded; accept as parameter or read $OUTPUT_DIR`
VRAM headroom	`nvidia-smi` headroom > 20% after settle	warn at >90% utilization; likely leak or scale issue
`pkill -f "kit/kit"` used	flag for kill+relaunch bug	warn: `Verify zero kit processes before next launch`
Sim duration	>= 3 s of physics	reject: `Sim ran < 3 s; physics has not settled`

File structure validation (USD scenes)

Item	Valid	Invalid	Action
`$ISAAC_LAB_DIR/scripts/`	read-only reference	writing outputs there	redirect to `$OUTPUT_DIR` or `$WORKSPACE_DIR/outputs/`
`get_assets_root_path() -> s3://...`	runtime OK	baked into delivered output	cache locally; reference via configurable path
`Melting` / `Splat` MDL-only shaders	OK in raw form	MDL-only at export	filter out; MDL-only causes black render
`MakeInstanceable` for RL	required (`make_instanceable: true`)	absent	reject
`usdPreviewSurface` on surfaces	required	absent	reject; suggest adding `UsdPreviewSurface`
`convex_decompose_mesh`	`false` (with `clean_sdf_dc`)	`true`	recommend `false` for stability
`mpu=0.01`	OK for warehouse shell (13 m tiles)	on a robot	reject

Render validation (frame analysis)

Metric	Good	Bad	Action
File size	1–4 MB	~82 KB	82 KB = black, reject
Mean RGB	80–160	< 20	reject: `Too dark; add lights`
Max RGB	200–250	0	reject: `No light in scene`
Pixel variance	> 15	< 5	flat = missing shadows; add accents
Color histogram	multi-modal	single peak	overexposed or flat lighting; adjust intensity
Object visibility	>= 3 robots or 50+ assets	< 10 visible	suggest: `Scene too sparse; add set dressing`
Shadow consistency	consistent shadows	random black blobs	reject: `Shadow artifacts; check light cone bounds`

RL training validation

Issue	Action
`AppLauncher` not first	reject: `AppLauncher must precede any isaaclab imports`
Task `__init__.py` non-empty	reject: `Task __init__.py must be empty; do not import configs`
`omni.isaac.lab` import	reject: `Use isaaclab.*; omni.isaac.lab is deprecated`
PhysX backend (when Newton applies)	reject: `Use Newton; it is the default on Kit 110`
Missing `PYTHONUNBUFFERED=1`	warn: `Set PYTHONUNBUFFERED=1 to see logs in real time`
No GPU memory logging	suggest: `Log torch.cuda.memory_allocated()/1e6 to W&B`
Batch size exceeds VRAM	suggest: `Reduce num_envs (try 2048)`
Wrong checkpoint name	reject: `Must be model_XXXX.zip, not policy.pt`
Single seed	suggest: `Run seeds 42, 123, 456 for valid comparison`

Asset & shell validation

Input	Check	Action
`docker run` command	`-v` mount for `/workspace/output`	reject: `Output folder not mounted`
Sub-agent spawn	`workdir=$WORKSPACE_DIR`	reject: `Sub-agent workdir hardcoded; use $WORKSPACE_DIR`
Blender USD export	correct mpu, dual-shader (UsdPreviewSurface + MDL), correct export type	reject if any missing
Scene complexity	renders within 10 settle steps	reject: `Scene too complex; simplify assets`
Ghost artifacts	cleanup of `/dev/shm/carb-*`	reject: `Use workspace_cleanup.py to clear /dev/shm/carb-*`

Decision tree

graph TD
    A[Received Output] --> B{Has script?}
    B -->|No| C[Reject: "No script provided"]
    B -->|Yes| D{File size >1KB?}
    D -->|No| E[Reject: "Script empty or corrupted"]
    D -->|Yes| F{Python?}
    F -->|Yes| G{isaacsim.* imports?}
    G -->|No| H[Reject: "Use isaacsim.* not omni.isaac.core"]
    G -->|Yes| I{SimulationApp or --exec?}
    I -->|Neither| J[Reject: "Must use SimulationApp with window args OR isaac-sim.sh --exec"]
    I -->|SimulationApp| K{width and height set?}
    K -->|No| L[Reject: "SimulationApp requires explicit width and height"]
    K -->|Yes| M{Dome + Distant lights?}
    M -->|No| N[Reject: "Add DomeLight and DistantLight (min intensity 100 and 500)"]
    M -->|Yes| O{RayTracedLighting?}
    O -->|No| P[Reject: "Use RayTracedLighting for iteration"]
    O -->|Yes| Q{Render output exists?}
    Q -->|No| R[Reject: "No render output"]
    Q -->|Yes| S{>= 150 KB and mean_RGB > 30?}
    S -->|No| T[Reject: "Render is black, too dark, or corrupted"]
    S -->|Yes| U{Hardcoded user paths or bare IPs?}
    U -->|Yes| V[Reject: "Replace hardcoded user paths/IPs with env vars or args"]
    U -->|No| Y[VALID - deliver]

Training output

Item	Action
`policy_only_XXXX.pt`	accept
`model_XXXX.zip`	accept
`log.csv` or `wandb` URL	accept
`results.json` with episode data	accept
No reward curve / metrics	reject: `Log episode/reward_mean and train/value_loss`

Reply format

Pass: Valid: <one-line summary>
Fail: Invalid: <list of failed checks>
Suggest (only when otherwise valid): Could be improved: <tip>

Do not hedge ("I think", "maybe"). Do not deliver a black frame.

Input sources

isaac-sim-orchestrator, isaac-sim-rendering, other producer skills, and sub-agent outputs.
Invoked on: "validate this", "check this render", "review this code".