Datasets:
HelenPeng
/
ExeWorldBench

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
596
* Do cropping, occlusion, distance, viewpoint, or rendering problems make important motion impossible to determine?
If the generated result's own camera, occlusion, or rendering problems make important mechanical motion unobservable or prevent reliable evaluation from the visible evidence, the Mechanical Motion score should be lowered.
Do not assume that unobservable motion is correct.
The more important mechanical processes that cannot be observed or verified, the greater the penalty to Mechanical Motion should be.
### J. Overall Motion Fidelity
Considering all aspects above, determine whether the generated system genuinely reproduces the core dynamic behavior of the reference mechanical device.
Focus on:
* whether a complete mechanical system executes as a whole;
* whether component motion patterns are correct;
* whether relative motion and dynamic connectivity are reasonable;
* whether mechanical interactions are correct;
* whether dynamic geometric consistency is maintained throughout execution;
* whether obvious mechanical failures are avoided;
* whether the intended mechanical process is completed;
* whether the core motion process is clearly observable in the generated result.
Exact pixel-level trajectory matching is not required.
If the generated result achieves the same mechanical relationships and functional outcome through a different but mechanically equivalent motion, it should not be penalized merely because the absolute direction, exact trajectory, speed, or timing differs.
### Mechanical Motion Score (0–10)
**10** = The complete mechanical system executes coherently. Core motion patterns, relative motion, dynamic connectivity, mechanical interactions, and the complete process closely match the reference. There are no significant dynamic mechanical failures across the generated sequence, and all key motion is clearly obser...
**9** = The motion closely matches the reference, with only very minor differences that do not affect the core mechanical behavior.
**8** = The overall motion is correct and complete, with a small number of visible motion, connectivity, interaction, temporal, dynamic-geometric, or observability issues, while the core mechanical process is reliably completed.
**7** = The major mechanical motion and execution process are correct, but there are several noticeable yet non-critical local problems involving motion, interaction, connectivity, dynamic geometry, or observability.
**6** = The intended mechanical motion is clearly recognizable and most of the core process is executed, but there are noticeable problems involving kinematics, interaction, connectivity, completeness, or observability.
**5** = The target mechanical process is generally executed, but multiple important motion relationships, interactions, or stages contain clear errors, or some important motion is difficult to observe.
**4** = Only part of the target mechanical behavior is executed, with important missing motion, failed interactions, dynamic connectivity failures, obvious interpenetration, or severe observability problems.
**3** = Only a small portion of the motion corresponds to the reference. The overall dynamic mechanical relationships contain major errors, and most of the process cannot be correctly executed or observed.
**2** = There is very limited valid motion related to the reference, but the overall system largely fails to produce meaningful mechanical execution.
**1** = Only minimal valid mechanical motion related to the reference is present, while the overall execution is severely incorrect, incomplete, or largely unobservable.
**0** = No meaningful valid mechanical execution corresponding to the reference occurs or can be observed.
==================================================
GENERAL JUDGING PRINCIPLES
==================================================
* The reference motion frames and generated frames are two independent ordered temporal sequences. There is no frame-to-frame correspondence or temporal-position alignment between the two sequences.
* Do not directly compare the i-th reference frame with the i-th generated frame.
* Carefully evaluate every reference frame and every generated frame, and use the complete sequences to understand the target mechanical structure and motion.
* Evaluate based on the provided visual evidence and do not assume invisible content.
* A single problem may affect both Visual Structure and Mechanical Motion if it genuinely causes both structural degradation and actual dynamic execution failure.
* Prioritize mechanically meaningful structure and behavior over superficial visual similarity.
* Do not reward visual complexity by itself.
* Allow reasonable geometric simplification that does not alter the core mechanical structure or function.
* Identical absolute motion direction is not required; mechanically equivalent mirrored or alternative motion is acceptable.
* Do not penalize differences in animation speed.
* Exact pixel-level motion trajectories are not required.
==================================================
OUTPUT FORMAT
==================================================
Return only valid JSON.
Do not include Markdown, code fences, scoring explanations, or any additional text.
{
"visual_structure": 5,
"mechanical_motion": 3
}
You are evaluating a multi-stage mechanical system generated by Blender code.
You are given:
1. **Ordered Component Specifications**, where each component contains:
* `id`: component identifier
* `label`: component type
* `action`: intended mechanical behavior
Components are listed in the intended execution order:
Component 1 → Component 2 → ... → Component N
2. **Generated Frames**, uniformly sampled from the animation and arranged chronologically.
Treat all frames as one complete temporal sequence.
Evaluate the result along two dimensions:
1. **Visual Structure** — 0–10