القائمة
Procedure for grounding scene plans in attached reference images — enumerate, describe, orient, then derive predicates.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Entry point for FSI-coupled hybrid plans (plan_type=fsi_in_scene) — any scene combining SPH fluid (water tanks, dam-break, wave channels) with multi-body dynamics, optionally with a wheeled vehicle. Pick this over mbs_in_scene whenever the plan involves a fluid domain (scene_objects with domain_type starting "sph_") or an FSI body registration (scene_objects with fsi_registration set). Routes to fsi/sph (always), veh/wheeled_vehicle (when vehicle present), and enforces FSI-specific invariants distinct from generic mbs_in_scene rigid scenes.
Entry point for rigid-body hybrid plans (plan_type=mbs_in_scene) combining a robot or vehicle with scene assets — NO fluid coupling. Routes to the correct domain skills and defines only high-level invariants. Use core/fsi_in_scene instead when the plan involves SPH fluid or FSI body registration.
Entry point for pure multi-body simulation plans (plan_type=mbs). Routes the agent to the correct mechanics, system, and camera skills and defines only high-level invariants.
Entry point for static scene plans (plan_type=scene). Routes the agent to the correct scene, system, and camera skills and defines only high-level invariants.
Set up SPH-based Fluid-Structure Interaction (FSI): create ChFsiFluidSystemSPH and ChFsiSystemSPH, configure fluid and SPH parameters, seed fluid particles with hydrostatic initialization, add container BCE boundary markers, register floating rigid bodies, optionally couple with a wheeled vehicle, and advance with sysFSI.DoStepDynamics(dT).
Translate a `geometry_relations` entry from the plan into correct PyChrono coordinate code. Read this whenever the plan declares a relation_name you have not previously encoded, and especially BEFORE writing SetPos / camera placement for any body that participates in a multi-body geometric constraint. Each subsection below is one canonical pattern named exactly as it appears in `plan.geometry_relations[i].relation_name`.
| name | image_grounding |
| description | Procedure for grounding scene plans in attached reference images — enumerate, describe, orient, then derive predicates. |
| compatibility | pychrono >= 8.0 |
| metadata | {"domain":"planning"} |
When a user attaches one or more images as the scene spec, the image is the authoritative source for layout and orientation — not a decorative prop. This skill defines the required procedure for turning an image into scene_predicates so the downstream simulation matches what the user drew or photographed.
Use this skill whenever a planning prompt has image_observation as a required section and one or more images are attached.
Perform these steps in order. Do not shortcut.
Enumerate visible objects. For every object you recognize, write one line: <object_name>: <what it is>. Use names that match the assets[] entries you plan to emit.
Relative positions. For each ordered pair of main objects, describe the relative position in image-native terms (left_of, right_of, in_front_of, behind, on_top_of, below). Use the schema <A>: <relation> <B> with generic placeholder names (<A>, <B>), not a specific object pair.
Orientations. For every directional object (monitor, TV, chair, laptop, camera, lamp head, robot, vehicle, any actor), state what it is FACING by naming the target object. Schema: <A>: facing <B>. If a chair has a clearly visible back/seat, the seat direction tells you where the occupant would be looking.
Viewpoint. Describe the camera viewpoint of the image relative to the primary subject(s) using the schema viewpoint: <relation_name> [— <one-line justification>]. Pick the closest match from the closed list:
top_down — the image is shot from directly above, looking down.side_minus_y — the image is shot from the −Y side of the subject (subject's left, world frame).side_plus_y — from the +Y side (subject's right).side_minus_x — from the −X side (subject's back).side_plus_x — from the +X side (subject's front).perspective — a corner / 3/4 view, neither pure side nor pure top-down. Use this when in doubt.Use ONLY these six names. The same enum drives topology.relation for camera objects in the plan, so the choice you write here will become the actual simulation viewpoint.
Cross-check user prompt vs image. If the prompt and the image disagree on placement or count, the image wins unless the user explicitly overrides the image in text.
Only after the image_observation block is written, derive scene_predicates. Apply these ordering rules:
FACING-TO as the primary orientation driver. Whenever you observed a "facing" relation in step 3, emit a FACING-TO predicate for it first.FRONT-OF / BACK-OF / LEFT-OF / RIGHT-OF FROM the facing direction — not from priors about "where things usually go". If <A> FACING-TO <B> and <A> sits on a support surface, <B> must lie on the side <A>'s forward axis actually points to in the scene's world frame.scene_coordinate_system worked example verbatim. Those are illustrative only. Recompute (x, y, z, deg_z) from YOUR predicates and the asset target heights.FACING-TO for it.viewpoint observation drives the first camera of step 1: at least one camera in the first implementation step MUST adopt the same relation as the observed viewpoint. Other cameras (when recording_mode: sensor_cams requires 2-3) remain free to pick complementary angles.The image_observation block must appear verbatim in the plan output as a top-level string field named image_observation, so it can be audited.
During refinement or repair, re-describe the image in a fresh image_observation block; do not reuse the previous one unless it is still correct. Compare new observations against existing scene_predicates and fix any contradictions, even if the user's clarification did not explicitly mention them.