File size: 3,559 Bytes
98e8615 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | # Embodied Navigation VLM-Based Evaluation
Evaluates whether a generated image/video correctly performs an embodied-navigation task by asking **Gemini-2.5-Pro** to judge it against the initial scene and the destination reference. This folder provides the exact evaluation prompts; run them through any Gemini access path (Google AI Studio, Gemini API, Vertex AI, or your own client).
There are **four tasks**, each with a separate **video** and **image** evaluation prompt (8 prompts total), organized one subfolder per task:
| Subfolder (task) | Description |
|------------------|-------------|
| `panoramic_view_last-mile_navigation` (task01) | Last-mile navigation in a 360° panoramic FPV scene |
| `top-down_view_real-world_navigation` (task02) | 2D top-down multi-floor navigation |
| `3D_real-world_navigation` (task03) | 3D dollhouse/cutaway real-world navigation |
| `simultaneous_localization_and_generation` (task04) | Split-screen: 3D navigation + synchronized 2D map |
Each subfolder contains:
- `video_evaluation_prompt.txt` — used to evaluate generated **videos**
- `image_evaluation_prompt.txt` — used to evaluate generated **images** (the final navigation state)
## Evaluation Protocol
For each generated output, send one user message to `gemini-2.5-pro` containing:
1. **The matching evaluation prompt** (text) — the `video_` or `image_` prompt from the task's subfolder
2. **The initial scene reference** — the starting panorama / top-down map / scene image
3. For object-goal levels, **the location description** text specifying the destination
4. **The generated video or image** to evaluate
The prompts are self-contained (criteria and output schema are inline) and contain no runtime placeholders — the references above are supplied as separate inputs.
## Evaluation Criteria
The prompts score a set of **binary** metrics (`"0"`/`"1"`), returned as JSON. The exact set varies slightly by task, but generally covers:
| Metric | Meaning |
|--------|---------|
| `SUCCESS_SCORE(_3D)` | Robot reaches and stops on/next to the specified destination by the end |
| `ORACLE_SUCCESS_SCORE(_3D)` | Trajectory ever passes through / pauses inside the correct destination region |
| `OBJECT_SEMANTIC_SCORE` | No collisions or clipping through walls/furniture |
| `AGENT_CONSISTENCY_SCORE` | Motion is temporally continuous (no teleporting, pose pops, size jumps) |
| `SPATIAL_ALIGNMENT_SCORE` | Heading / walking direction / elevation changes stay coherent |
| `DESTINATION_INTEGRITY_SCORE` | Destination in the output matches the reference (position, appearance, semantics) and is unchanged |
| `SCENE_CHANGE_SCORE` | The environment stays static aside from expected viewpoint changes |
| `TRAJECTORY_ALIGNMENT_SCORE` *(task04)* | The 2D-map trajectory matches the 3D navigation path |
Refer to each prompt for the precise metric list and definitions for that task. The prompts instruct Gemini to be fail-safe (assign `0` when ambiguous) and to penalize physical violations even when the goal is reached.
## Response Format
Each prompt instructs Gemini to return **only a JSON object** with a `SCORES` field (string `"0"`/`"1"` per metric) plus a brief reasoning field. Parse the JSON from the response.
## Metrics
`SUCCESS_SCORE(_3D)` is the primary task-success metric. For a set of generations, report:
```
success_rate = (# of outputs with SUCCESS_SCORE = 1) / (total # of outputs)
```
`ORACLE_SUCCESS_SCORE` gives a more lenient upper bound; the remaining metrics are physical-plausibility / fidelity diagnostics.
|