simready-foundation-conform-fet-005-standard

SimReady Conform FET_005_STANDARD Simulate Grasp Physics

Purpose

Use this workflow skill after Core, Minimal, and rigid-body requirements are already in a reasonable state and the next profile validation gate is FET_005_STANDARD@0.1.0. It adds or repairs a USD grasp vector line that tells a robotic gripper where to grasp the asset.

This is a visual-semantic authoring skill, not a bounding-box shortcut. The grasp line must be chosen from visual evidence of the asset shape and intended grasp behavior, then authored deterministically as a UsdGeom.BasisCurves prim.

Capability Gate

This skill requires a vision-capable agent or model. Before authoring, replacing, or invoking any script that creates a grasp line, confirm that the current agent can inspect visual evidence directly, such as renders, screenshots, viewport captures, or generated mesh previews.

If the current agent cannot use vision, stop immediately. Do not add or repair a grasp line from bounds, text metadata, file names, validator failures, or user descriptions alone. Tell the user that FET_005_STANDARD grasp-line repair requires a vision-capable model/agent and ask them to rerun this skill with vision enabled.

If the current agent can use vision and visual evidence is available, the agent is expected to choose explicit local-space grasp points from that evidence. Do not skip authoring merely because the user did not pre-supply points. User approval is only required when the visual evidence, coordinate mapping, gripper constraints, or grasp intent remain genuinely ambiguous after inspection.

Prerequisites

Read the source-of-truth files named below before editing. Work on staged outputs where the skill requires them, and keep validation evidence with the result. This grasp workflow also requires a vision-capable agent or rendered visual evidence.

Source of Truth

Before changing an asset, load the selected FET_005_STANDARD manifest and requirement docs:

• nv_core/tiers/simready_foundation_tier_core/simready/foundation/tier_core/features/FET_005_STANDARD-0.1.0.json
• nv_core/tiers/simready_foundation_tier_core/simready/foundation/tier_core/features/FET_005_STANDARD.md
• nv_core/tiers/simready_foundation_tier_core/simready/foundation/tier_core/capabilities/physics_bodies/physics_graspable/requirements/graspable-vector-line.md
• nv_core/tiers/simready_foundation_tier_core/simready/foundation/tier_core/capabilities/physics_bodies/physics_graspable/validation.py

The current manifest lists PMT.001 and GSP.001. The current FET_005_STANDARD prose and graspable validator focus on GSP.001: at least one BasisCurves prim under the default prim, named with the grasp_identifier prefix, with at least two points. If validation reports PMT.001, read references/fet005-requirements.md and handle physics-material binding separately from grasp-line placement.

Inputs

Collect these before editing:

Instructions

Use this checklist when changing the repository:

1. Confirm the input asset exists and that earlier gates needed by the selected profile are already staged. For a prop workflow, FET_000_STANDARD and FET_001_STANDARD should pass; the selected FET_003_* rigid-body physics feature should usually pass before FET_005_STANDARD because grasping assumes a physical prop.
2. Parse the validation report and filter to FET_005_STANDARD@0.1.0 unless the user selected a different future version. If GSP.001 already passes, do not add duplicate grasp lines.
3. Inspect existing grasp vector prims under the default prim:
   • type is BasisCurves
   • name starts with grasp_identifier
   • points has at least two points
   • curve is under the asset default prim
   • it may be below a clear annotation scope rather than below the rigid body; do not reparent a valid line solely to satisfy the runtime test
4. Generate or collect visual evidence before choosing points. Prefer real renders from render-usd, Kit, Omniverse viewport screenshots, or existing preview images when available. If no renderer is available, use assets/scripts/render_grasp_preview.py to generate a four-panel point-cloud PNG with top, front, side, and isometric views. If the staged USD opens but has no renderable mesh triangles, use the nearest source visual asset from conversion context or the asset fixture/source tree, then record that fallback explicitly. If the current agent cannot inspect the generated or collected images with vision, write a first-class BLOCKED report and stop.
5. Use visual reasoning to choose a grasp region:
   • Prefer stable, rigid, central body geometry with enough contact area for opposing gripper fingers.
   • Avoid handles, holes, voids, thin rims, fragile appendages, decorative parts, sharp protrusions, and table-contact bottoms unless the user or asset semantics explicitly say those are intended.
   • For handled containers, prefer the main body or sidewall region, not the handle, unless a handle grasp is explicitly requested.
   • For box-like props, prefer a line across two broad opposing side faces with clearance.
   • For cylindrical props, prefer a diameter across the main body at a stable height, avoiding top and bottom edges.
   • For tools, use the designed grip handle only when it is the intended robotic grasp region.
6. Convert the selected visual grasp into USD local-space points. The line should intersect the graspable geometry and represent the gripper contact/closing axis or agreed runtime convention. If the visual source and staged USD use different units, root transforms, or meter-normalization scales, convert points into the authoring parent's local space and record the conversion note. If the coordinate mapping is uncertain, inspect bounds, xform ops, layer scale, and default-prim transforms before authoring. If explicit points still cannot be chosen, write a BLOCKED report with the visual evidence path, the reason points could not be selected, and the exact inputs needed to continue.
7. Stage a copy under output_root; do not mutate the source unless the user explicitly asks for in-place repair.
8. Author or replace one BasisCurves prim named grasp_identifier_## under the default prim, with:
   • type = "linear"
   • curveVertexCounts = [point_count]
   • points = [...]
   • a small display width
   • purpose = "guide"
   • visible display color for review A clear grasp annotation scope below the default prim is valid. When using one, verify that the line's world-space segment crosses the intended rigid body so runtime body resolution is unambiguous.
9. Render or inspect the repaired asset again. Confirm the line is visible, intersects the intended region, and is not accidentally placed through an avoided region. When possible, save overlay images or a small decision JSON beside the staged output.
10. Rerun the same profile validation gate, or the narrowest available FET_005_STANDARD validation gate. Stop when FET_005_STANDARD passes or when the remaining FET_005_STANDARD issue requires user or runtime-gripper judgement. If the full profile still fails on another feature, report the first remaining failing gate separately from the FET_005_STANDARD result.

Authoring Script

Use assets/scripts/render_grasp_preview.py to create visual evidence when a renderer or viewport is not available. This script uses USD mesh points and pure-Python PNG output, so it does not require PIL, matplotlib, Kit, or a display server:

uv run --python 3.12 python <skill-dir>/assets/scripts/render_grasp_preview.py <input-usd> \

After choosing explicit points, rerun the preview with the proposed line overlaid and inspect it before authoring:

The preview helper is evidence generation only. It does not choose the grasp region and does not replace vision review.

Use assets/scripts/author_grasp_line.py for deterministic USD authoring after the vision decision has produced explicit points. Prefer uv run --python 3.12 python or python3; do not assume python exists or points to Python 3.10+ on Linux:

When repair is blocked before explicit points are available, use the same script to produce a distinguishable BLOCKED report instead of a failed repair report:

If the Physical AI Skill Hub CLI is available, author-grasp-vectors can also author explicit points. Do not rely on its bounding-box heuristic unless visual review confirms the generated line is semantically correct for the asset.

The script only authors explicit points. It does not choose the grasp region, render evidence, or infer scale. Use the optional evidence and rationale flags so the report preserves the human/vision decision that led to those points. If no points are supplied, the script returns BLOCKED with requirements_blocked = ["GSP.001"]; workflow summaries must preserve that status instead of converting it to FAIL.

Vision Policy

Do not use this skill to author a grasp line unless the current agent can inspect the visual evidence directly. A non-vision agent must report blocked for GSP.001 repair and should not call assets/scripts/author_grasp_line.py.

Do not place the final line from a bounding box alone when the asset has meaningful shape, openings, handles, or multiple plausible grasp regions. A bounding-box line is acceptable only for simple primitives or after render review confirms it intersects the intended graspable body.

If there is no way to see the asset, block and request visual evidence or a specific user-approved grasp region. Passing GSP.001 with an arbitrary line is not a useful SimReady repair.

If there is a way to see the asset and the current model is vision-capable, make the repair actionable: inspect the image, choose two or more explicit points, author the curve, then rerun validation. A statement such as "visual evidence existed but no user-approved points were supplied" is not sufficient unless the asset has unresolved ambiguity that the report names precisely.

Examples

Example request:

Expected result summary:

Validation Handoff

Preserve reports under the staged output directory. If the Physical AI Skill Hub validation commands are available, use:

Count this skill as successful when FET_005_STANDARD@0.1.0 passes, even if the full profile still fails on unrelated features. If PMT.001 fails, report it as a physics-material binding issue and do not hide it by only adding a grasp line.

Limitations

• Do not silently mutate the source asset; work on the requested staged output.
• Do not hide later profile failures after the selected feature gate passes or fails.
• Do not invent geometry, metadata, or runtime behavior that conflicts with the asset intent.

Troubleshooting

• Error: validation tooling is unavailable. Solution: run the narrowest available USD or static check and report the gap.
• Error: a repair would change asset intent. Solution: stop and ask for direction or stage the smallest reversible edit.
• Error: later profile gates still fail. Solution: report the next failing feature and hand off to the matching conformance skill.
• Error: runtime reports Gripper rebuild failed. Solution: preserve the original setup exception. Import/backend failures belong to the environment; body-resolution or empty-bound errors require inspecting the grasp segment and rigid-body geometry. Do not respond with generic mass/friction tuning.

Resources

• assets/openai.yaml preserves optional UI metadata for clients that read skill display hints. It is not required for the workflow.
• assets/scripts/ contains the deterministic preview and authoring helpers for this skill.
• references/ contains detailed requirement notes; load only the files needed for the active validation failure.

Available Scripts

Summary Format

Report:

Keep the user-facing summary short: what visual evidence was used, where the grasp line was placed, why that region was chosen, and what validation reported.