Datasets:
HelenPeng
/
ExeWorldBench

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
596
==================================================
Return valid JSON only. Do not output Markdown, explanations, or additional text.
Infer the actual execution order from the generated frames using the original component IDs.
Every required component must correspond to exactly one segment. Therefore, for N components, `executable_causal_mechanism` must contain N segments:
* first component: `initial → C1`
* second component: `C1 → C2`
* ...
* final component: `C(N-1) → CN`
If the causal chain breaks, still infer the most strongly supported downstream order from the scene's spatial layout, mechanical connections, and observed execution, and assign 0 to unreachable downstream segments.
Do not invent unsupported mechanical relationships merely to complete the order.
Output format:
{
"visual_structure": 8,
"execution_order": [3, 1, 4, 2],
"executable_causal_mechanism": [
{"from": "initial", "to": 3, "score": 9},
{"from": 3, "to": 1, "score": 9},
{"from": 1, "to": 4, "score": 6},
{"from": 4, "to": 2, "score": 0}
]
}
You are evaluating a 3D mechanical world autonomously synthesized and executed by Blender code.
I will provide you with:
1. Frames sampled from the generated animation and arranged in chronological order;
2. The object, event, and state definitions for the current task;
3. The I / B / F items to be evaluated.
You must inspect all generated frames and the dynamic changes across the complete temporal sequence, rather than checking only the first or last frame.
Evaluate the result along three dimensions:
1. Visual Structure
2. Executable Causal Mechanism
3. Goal Satisfaction
Scoring:
* Visual Structure: 0–10;
* Executable Causal Mechanism: 0–10;
* Goal Satisfaction: strict binary judgment for each I / B / F item, using only 0 or 1.
For Visual Structure and Executable Causal Mechanism, a score of 0 indicates complete failure in the corresponding dimension, with no meaningful valid structure or execution demonstrated.
Judge only from the provided task information and visible evidence in the generated frames.
Do not assume invisible structures, interactions, state changes, or motions. Do not assume that a behavior occurred merely because it was expected. If the frames do not provide sufficient evidence that a requirement is satisfied, do not treat it as satisfied.
==================================================
1. VISUAL STRUCTURE
==================================================
Evaluate whether the generated 3D mechanical world is visually and structurally complete, clear, mechanically reasonable, physically plausible, and capable of supporting the behavior shown in the animation.
Focus on:
### A. Objects and Mechanical Structure
* Whether required key objects exist and are recognizable;
* Whether object geometry is consistent with its definition and mechanical role;
* Whether objects that should be independent are implemented as independent physical entities;
* Whether key objects are missing, duplicated, severely deformed, or structurally incorrect;
* Whether levers, pendulums, hinges, rotating wheels, tracks, chutes, containers, and similar mechanisms have reasonable pivots, axes, connections, supports, or constraints;
* Whether decorative geometry is used instead of actual mechanical structure.
Necessary passive auxiliary structures such as ground, bases, fixed supports, connectors, axes, and track supports are allowed.
### B. Initial Physical State and Spatial Layout
Check:
* Whether objects float without support or constraints;
* Whether objects are clearly unstable without reasonable holding structures;
* Whether downstream objects move before mechanical propagation reaches them;
* Whether spatially separated stages independently move at the beginning;
* Whether object positions, heights, orientations, and distances support the required mechanical interactions;
* Whether collision, catching, connection, release, passage, or blocking locations are properly aligned;
* Whether there are severe misalignments, disconnections, penetrations, overlaps, or impossible geometric relationships.
Do not assume that an initial state is physically plausible merely because an object is stationary in the first frame.
### C. Observability
* Whether key objects and mechanical structures are clearly visible;
* Whether severe cropping, occlusion, blur, deformation, or rendering errors occur;
* Whether the camera supports reliable judgment of the beginning, intermediate process, and final state.
Photorealism is not required. The scene only needs to be sufficiently clear and stable for reliable evaluation.
### Visual Structure Score (0–10)