DART Simulation Verification
Load this skill when verifying that a DART 3D scene or physics simulation is
correct — implementing, debugging, benchmarking, or reviewing dynamics,
collision, contact, or GUI output. Modern image-capable agents can inspect a
capture, but pixels do not expose solver state and machine image checks are not
semantic inspection. This tooling grounds visual reasoning without a GUI.
Lead with text, corroborate with images. Measured A/B evidence: per-step
metrics and trajectories detect nearly all seeded physics defects; a rendered
image alone misses static geometry defects (penetration, interpenetration).
Decide correctness from text; use images for scene comprehension and gross
dynamic failures.
Applicability contract
Use this skill for any task whose claim depends on 3D structure or behavior:
model/scene loading, dynamics, collision/contact/constraints, simulation
stepping, GUI/rendering, or visual examples. First run a text oracle (metrics,
scene diff, trajectory/contact comparison, or focused behavioral test), then
corroborate it end to end with an assessed headless view and only the debug
layers needed by the claim. If rendering is unavailable or genuinely
irrelevant, record why and name the replacement evidence; never treat an image
as the sole correctness oracle. When the active agent accepts image input,
actually open and semantically inspect the selected capture; a passing view
report or image-verdict is not visual review.
Full documentation
docs/onboarding/agent-sim-verification.md
— the durable guide. docs/ai/verification.md owns the gate policy;
docs/onboarding/profiling.md owns text-first profiling.
Image-capable review loop
Every lane in the model-routing owner (docs/ai/README.md § "Model
Routing") is image-capable and supports native image input and
original-detail inspection; a model-upgrade audit re-verifies that property
and updates the owner, not this skill. Keep this loop capability-based.
- State one claim, its expected visible observation, and the text oracle that
decides correctness.
- Capture one assessed, claim-tied view first. Add only the debug layers needed
for the claim; use paired plain/debug views when an overlay could obscure the
underlying scene. For a turntable or motion sequence, inspect at least the
capture sidecar's start/middle/end frame targets; add intervening frames or a
grid when a transient event is part of the claim.
- Run
image-verdict for artifact integrity, then open the selected local PNG
with the active agent's native image viewer. Use original detail for small
contacts, labels, bounds, or frame axes. Add a grid or another view only when
motion, occlusion, or ambiguity requires it.
- Record the visible observation separately from the text result. If they
disagree, do not average them into a pass: report fail/uncertain, inspect the
capture sidecar, reframe or recapture, and investigate the simulation state.
- Close with pass/fail/uncertain, artifact path, view/layers, reproduction
command, what the image shows, and what it does not prove. If native image
review is unavailable, use
verification-bundle for an image-capable
reviewer and record that limitation.
Quick commands
Text (primary):
world.compute_step_metrics() — energy/momentum/penetration/contacts/residual
(solved pairs only; static- or kinematic-only overlap needs world.collide())
dartpy.dump_scene_json(world) / dump_scene_text(world) — "what is in this
world?" (glTF/USD-flavored hierarchy + flat index)
pixi run scene-diff — structural JSON verdict for intended-vs-actual scene
dumps
pixi run trajectory-record / pixi run trajectory-compare — per-body TSV +
contact JSONL; bit-exact or tolerance diff with first-divergence; scratch
scenes via --factory path/to/scene.py:callable
Visual (corroboration):
dart.gui.render(world, camera=None, size=(w, h), debug=(...layers...)) →
headless image with optional world-derived debug layers (grid,
world_frame, body_frames, coms, inertia_boxes, collision_bounds,
velocities, contacts, labels; trajectories additionally requires a
sampled dart.gui.TrajectoryTracker via debug_scene_for_world);
dart.gui.render_annotated(...) composites label text; .png_bytes() for
notebooks; dart.gui.orbit_camera(...) / look_at(...)
dart.gui.assess_view(world, camera, size, focus=...) → ViewReport with
issues (cropped/too-far/too-close/occluded/ambiguous);
dart.gui.select_viewpoints(...) picks deterministic best views;
dart.gui.frame_body/frame_region reframe onto a subject. Assess first;
fix flagged views before capturing evidence. Descriptor bounds, transformed
corners, viewport FOV, and fit distance stay in the shared dart::gui core;
Python performs only focus resolution and capture/search orchestration.
- viewer camera flags:
--view {three-quarter|front|side|top},
--camera-azimuth/-elevation/-distance/-target, --turntable N, --fit
pixi run py-demo-capture — headless PNG/MP4 capture from Python
pixi run agent-capture — deterministic evidence harness: auto/explicit
cameras, debug layers, stills/turntable/motion video, reproducible sidecar
pixi run image-compose — side-by-side / blend / diff-heatmap composites
pixi run evidence-select — claim-driven artifact selection with recorded
rationale; pixi run evidence-publish — PR "Visual verification" section
with a required semantic verdict and claim boundary plus GitHub-hosted media
(manual placeholders by default; gh-release upload only with --yes +
maintainer approval)
pixi run image-verdict / image-golden / image-sheet — JSON verdict,
golden diff, contact sheet (contrast is report-only; --require-contrast to
gate); these are machine pixel checks, not semantic visual review
pixi run image-ab-study — blind-judge detection deltas for single-view,
multi-view, turntable, and annotated captures
pixi run image-ab-round2 — prepare a blinded round-2 packet and score
completed judge observations
Opt-in:
pixi run render-golden-gate — opt-in golden gate (backend-specific golden,
curated locally with -- --update; not default CI)
pixi run rerun-trajectory — rerun.io inspection (opt-in; graceful when
rerun-sdk is absent)
pixi run verification-bundle — package text evidence plus still/grid images
for a provider-neutral VLM or reviewer call
Default capture for agent review: one ~1280 px frame, UI hidden, 3/4 view; add
a 9-frame grid for motion. Keep images as corroboration, never the sole oracle
for static geometry.
DART 6 (release-6.20)
Use that branch's own dart-verify-sim skill; its OpenSceneGraph capture path
and task names differ from main.
1---2name: dart-verify-sim3description: DART Verify Sim: text-first and visual checks for 3D scenes and physics (metrics, scene dump, trajectories, headless render, image verdict/golden)4---56<!-- AUTO-GENERATED FILE - DO NOT EDIT MANUALLY -->7<!-- Source: .claude/skills/dart-verify-sim/SKILL.md -->8<!-- Sync script: scripts/sync_ai_commands.py -->9<!-- Run `pixi run sync-ai-commands` to update -->1011# DART Simulation Verification1213Load this skill when verifying that a DART 3D scene or physics simulation is14correct — implementing, debugging, benchmarking, or reviewing dynamics,15collision, contact, or GUI output. Modern image-capable agents can inspect a16capture, but pixels do not expose solver state and machine image checks are not17semantic inspection. This tooling grounds visual reasoning without a GUI.1819**Lead with text, corroborate with images.** Measured A/B evidence: per-step20metrics and trajectories detect nearly all seeded physics defects; a rendered21image alone misses static geometry defects (penetration, interpenetration).22Decide correctness from text; use images for scene comprehension and gross23dynamic failures.2425## Applicability contract2627Use this skill for any task whose claim depends on 3D structure or behavior:28model/scene loading, dynamics, collision/contact/constraints, simulation29stepping, GUI/rendering, or visual examples. First run a text oracle (metrics,30scene diff, trajectory/contact comparison, or focused behavioral test), then31corroborate it end to end with an assessed headless view and only the debug32layers needed by the claim. If rendering is unavailable or genuinely33irrelevant, record why and name the replacement evidence; never treat an image34as the sole correctness oracle. When the active agent accepts image input,35actually open and semantically inspect the selected capture; a passing view36report or `image-verdict` is not visual review.3738## Full documentation3940[`docs/onboarding/agent-sim-verification.md`](../../../docs/onboarding/agent-sim-verification.md)41— the durable guide. `docs/ai/verification.md` owns the gate policy;42`docs/onboarding/profiling.md` owns text-first profiling.4344## Image-capable review loop4546Every lane in the model-routing owner (`docs/ai/README.md` § "Model47Routing") is image-capable and supports native image input and48original-detail inspection; a model-upgrade audit re-verifies that property49and updates the owner, not this skill. Keep this loop capability-based.50511. State one claim, its expected visible observation, and the text oracle that52 decides correctness.532. Capture one assessed, claim-tied view first. Add only the debug layers needed54 for the claim; use paired plain/debug views when an overlay could obscure the55 underlying scene. For a turntable or motion sequence, inspect at least the56 capture sidecar's start/middle/end frame targets; add intervening frames or a57 grid when a transient event is part of the claim.583. Run `image-verdict` for artifact integrity, then open the selected local PNG59 with the active agent's native image viewer. Use original detail for small60 contacts, labels, bounds, or frame axes. Add a grid or another view only when61 motion, occlusion, or ambiguity requires it.624. Record the visible observation separately from the text result. If they63 disagree, do not average them into a pass: report fail/uncertain, inspect the64 capture sidecar, reframe or recapture, and investigate the simulation state.655. Close with pass/fail/uncertain, artifact path, view/layers, reproduction66 command, what the image shows, and what it does not prove. If native image67 review is unavailable, use `verification-bundle` for an image-capable68 reviewer and record that limitation.6970## Quick commands7172Text (primary):7374- `world.compute_step_metrics()` — energy/momentum/penetration/contacts/residual75 (solved pairs only; static- or kinematic-only overlap needs `world.collide()`)76- `dartpy.dump_scene_json(world)` / `dump_scene_text(world)` — "what is in this77 world?" (glTF/USD-flavored hierarchy + flat index)78- `pixi run scene-diff` — structural JSON verdict for intended-vs-actual scene79 dumps80- `pixi run trajectory-record` / `pixi run trajectory-compare` — per-body TSV +81 contact JSONL; bit-exact or tolerance diff with first-divergence; scratch82 scenes via `--factory path/to/scene.py:callable`8384Visual (corroboration):8586- `dart.gui.render(world, camera=None, size=(w, h), debug=(...layers...))` →87 headless image with optional world-derived debug layers (`grid`,88 `world_frame`, `body_frames`, `coms`, `inertia_boxes`, `collision_bounds`,89 `velocities`, `contacts`, `labels`; `trajectories` additionally requires a90 sampled `dart.gui.TrajectoryTracker` via `debug_scene_for_world`);91 `dart.gui.render_annotated(...)` composites label text; `.png_bytes()` for92 notebooks; `dart.gui.orbit_camera(...)` / `look_at(...)`93- `dart.gui.assess_view(world, camera, size, focus=...)` → ViewReport with94 issues (`cropped`/`too-far`/`too-close`/`occluded`/`ambiguous`);95 `dart.gui.select_viewpoints(...)` picks deterministic best views;96 `dart.gui.frame_body`/`frame_region` reframe onto a subject. Assess first;97 fix flagged views before capturing evidence. Descriptor bounds, transformed98 corners, viewport FOV, and fit distance stay in the shared `dart::gui` core;99 Python performs only focus resolution and capture/search orchestration.100- viewer camera flags: `--view {three-quarter|front|side|top}`,101 `--camera-azimuth/-elevation/-distance/-target`, `--turntable N`, `--fit`102- `pixi run py-demo-capture` — headless PNG/MP4 capture from Python103- `pixi run agent-capture` — deterministic evidence harness: auto/explicit104 cameras, debug layers, stills/turntable/motion video, reproducible sidecar105- `pixi run image-compose` — side-by-side / blend / diff-heatmap composites106- `pixi run evidence-select` — claim-driven artifact selection with recorded107 rationale; `pixi run evidence-publish` — PR "Visual verification" section108 with a required semantic verdict and claim boundary plus GitHub-hosted media109 (manual placeholders by default; `gh-release` upload only with `--yes` +110 maintainer approval)111- `pixi run image-verdict` / `image-golden` / `image-sheet` — JSON verdict,112 golden diff, contact sheet (contrast is report-only; `--require-contrast` to113 gate); these are machine pixel checks, not semantic visual review114- `pixi run image-ab-study` — blind-judge detection deltas for single-view,115 multi-view, turntable, and annotated captures116- `pixi run image-ab-round2` — prepare a blinded round-2 packet and score117 completed judge observations118119Opt-in:120121- `pixi run render-golden-gate` — opt-in golden gate (backend-specific golden,122 curated locally with `-- --update`; not default CI)123- `pixi run rerun-trajectory` — rerun.io inspection (opt-in; graceful when124 `rerun-sdk` is absent)125- `pixi run verification-bundle` — package text evidence plus still/grid images126 for a provider-neutral VLM or reviewer call127128Default capture for agent review: one ~1280 px frame, UI hidden, 3/4 view; add129a 9-frame grid for motion. Keep images as corroboration, never the sole oracle130for static geometry.131132## DART 6 (release-6.20)133134Use that branch's own `dart-verify-sim` skill; its OpenSceneGraph capture path135and task names differ from `main`.