Add frozen evaluation episode specifications
Browse filesAdds the unchanged pre-submission 500-episode navigation specification and synchronizes the anonymous dataset documentation and Croissant metadata.
- README.md +18 -2
- croissant.json +10 -10
- episodes.jsonl +0 -0
README.md
CHANGED
|
@@ -25,8 +25,9 @@ IntentionNav is a 500-item benchmark for intent-driven object navigation in indo
|
|
| 25 |
**Target categories:** 64
|
| 26 |
**Instruction variants:** 4 per item, for 2000 English instructions
|
| 27 |
**Photo format:** 1024 x 1024 PNG, rendered in Isaac Sim
|
| 28 |
-
**Target policy:** each item points to one USD object instance whose target category has exactly one physical instance in that scene.
|
| 29 |
-
**
|
|
|
|
| 30 |
|
| 31 |
Primary intent-mode counts are: event-script 202, inner-state 167, physical-state 72, and affordance 59.
|
| 32 |
|
|
@@ -36,6 +37,7 @@ Primary intent-mode counts are: event-script 202, inner-state 167, physical-stat
|
|
| 36 |
README.md
|
| 37 |
croissant.json
|
| 38 |
selected_500_intents.jsonl
|
|
|
|
| 39 |
<scene_id>/
|
| 40 |
intents.json
|
| 41 |
manifest.json
|
|
@@ -56,6 +58,12 @@ selected_500_intents.jsonl
|
|
| 56 |
- `room` / `room_type`: where the target is located
|
| 57 |
- see `selected_500_intents.jsonl` for the full schema
|
| 58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
## Loading Example
|
| 60 |
|
| 61 |
```python
|
|
@@ -67,6 +75,12 @@ with open("selected_500_intents.jsonl", "r", encoding="utf-8") as f:
|
|
| 67 |
item = items[0]
|
| 68 |
print(item["selection_id"], item["scene_id"], item["target_category"])
|
| 69 |
print(item["natural_en"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
```
|
| 71 |
|
| 72 |
Image paths are scene-relative. For an item with `scene_id == "kujiale_0262"` and `photo == "surface_photos/21_x.png"`, the packaged image is under `kujiale_0262/photos/21_x.png`.
|
|
@@ -75,4 +89,6 @@ Image paths are scene-relative. For an item with `scene_id == "kujiale_0262"` an
|
|
| 75 |
|
| 76 |
The dataset is intended for benchmark evaluation of embodied AI agents and VLM-based navigation systems. It does not contain human subjects or personally identifiable information. The current release includes model-generated English intents and automated VLM judge metadata; users should treat the annotations as benchmark labels rather than naturally collected human utterances.
|
| 77 |
|
|
|
|
|
|
|
| 78 |
This peer-review snapshot intentionally omits author identities, affiliations, and identifying project links. The benchmark payload is otherwise unchanged from the fixed pre-submission release.
|
|
|
|
| 25 |
**Target categories:** 64
|
| 26 |
**Instruction variants:** 4 per item, for 2000 English instructions
|
| 27 |
**Photo format:** 1024 x 1024 PNG, rendered in Isaac Sim
|
| 28 |
+
**Target policy:** each item points to one USD object instance whose target category has exactly one physical instance in that scene.
|
| 29 |
+
**Intent-record SHA-256:** `fcf28886c8c853f9e3f9158dc420d63090a6fec58c44f094e54f9657e2b22b27`
|
| 30 |
+
**Episode-spec SHA-256:** `b9cd9699b90f6275e631c47d447abf52ee48c525d95668851948dafadb29a9bc`
|
| 31 |
|
| 32 |
Primary intent-mode counts are: event-script 202, inner-state 167, physical-state 72, and affordance 59.
|
| 33 |
|
|
|
|
| 37 |
README.md
|
| 38 |
croissant.json
|
| 39 |
selected_500_intents.jsonl
|
| 40 |
+
episodes.jsonl
|
| 41 |
<scene_id>/
|
| 42 |
intents.json
|
| 43 |
manifest.json
|
|
|
|
| 58 |
- `room` / `room_type`: where the target is located
|
| 59 |
- see `selected_500_intents.jsonl` for the full schema
|
| 60 |
|
| 61 |
+
## Fixed Navigation Episodes
|
| 62 |
+
|
| 63 |
+
`episodes.jsonl` contains the frozen 500-episode navigation specification used for evaluation.
|
| 64 |
+
Each row is joined to the intent records by `selection_id` and includes the exact `scene_id`, target object and position, start position and quaternion, start/target rooms, and geodesic/euclidean distances.
|
| 65 |
+
The file was frozen before submission and is released unchanged.
|
| 66 |
+
|
| 67 |
## Loading Example
|
| 68 |
|
| 69 |
```python
|
|
|
|
| 75 |
item = items[0]
|
| 76 |
print(item["selection_id"], item["scene_id"], item["target_category"])
|
| 77 |
print(item["natural_en"])
|
| 78 |
+
|
| 79 |
+
with open("episodes.jsonl", "r", encoding="utf-8") as f:
|
| 80 |
+
episodes = {row["selection_id"]: row for row in map(json.loads, f)}
|
| 81 |
+
|
| 82 |
+
episode = episodes[item["selection_id"]]
|
| 83 |
+
print(episode["start_position"], episode["target_position"])
|
| 84 |
```
|
| 85 |
|
| 86 |
Image paths are scene-relative. For an item with `scene_id == "kujiale_0262"` and `photo == "surface_photos/21_x.png"`, the packaged image is under `kujiale_0262/photos/21_x.png`.
|
|
|
|
| 89 |
|
| 90 |
The dataset is intended for benchmark evaluation of embodied AI agents and VLM-based navigation systems. It does not contain human subjects or personally identifiable information. The current release includes model-generated English intents and automated VLM judge metadata; users should treat the annotations as benchmark labels rather than naturally collected human utterances.
|
| 91 |
|
| 92 |
+
The USD scene assets, raw evaluation logs, and evaluation agents are not included in this snapshot. The episode file fixes the benchmark geometry and identifiers but does not by itself reproduce the reported reference-agent scores.
|
| 93 |
+
|
| 94 |
This peer-review snapshot intentionally omits author identities, affiliations, and identifying project links. The benchmark payload is otherwise unchanged from the fixed pre-submission release.
|
croissant.json
CHANGED
|
@@ -84,7 +84,7 @@
|
|
| 84 |
{
|
| 85 |
"version": "1.0.0",
|
| 86 |
"datePublished": "2026-05-05",
|
| 87 |
-
"description": "Initial release: 500 unique intents, 176 scenes, 64 categories, 4 styles, 4 intent modes."
|
| 88 |
}
|
| 89 |
],
|
| 90 |
"rai:dataUseCases": "Evaluating zero-shot VLM/VLA navigation systems on intent-driven object localization; diagnosing model robustness across instruction style and intent type; ablating perception vs. language vs. planning bottlenecks in embodied agents.",
|
|
@@ -119,13 +119,13 @@
|
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"@type": "cr:FileObject",
|
| 122 |
-
"@id": "
|
| 123 |
-
"name": "
|
| 124 |
-
"description": "
|
| 125 |
-
"contentUrl": "
|
| 126 |
-
"encodingFormat": "application/
|
| 127 |
-
"sha256": "
|
| 128 |
-
"contentSize": "
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"@type": "cr:FileObject",
|
|
@@ -134,8 +134,8 @@
|
|
| 134 |
"description": "Dataset README with field schema and directory layout.",
|
| 135 |
"contentUrl": "README.md",
|
| 136 |
"encodingFormat": "text/markdown",
|
| 137 |
-
"sha256": "
|
| 138 |
-
"contentSize": "
|
| 139 |
}
|
| 140 |
],
|
| 141 |
"recordSet": [
|
|
|
|
| 84 |
{
|
| 85 |
"version": "1.0.0",
|
| 86 |
"datePublished": "2026-05-05",
|
| 87 |
+
"description": "Initial release: 500 unique intents, 500 fixed navigation episodes, 176 scenes, 64 categories, 4 styles, and 4 intent modes."
|
| 88 |
}
|
| 89 |
],
|
| 90 |
"rai:dataUseCases": "Evaluating zero-shot VLM/VLA navigation systems on intent-driven object localization; diagnosing model robustness across instruction style and intent type; ablating perception vs. language vs. planning bottlenecks in embodied agents.",
|
|
|
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"@type": "cr:FileObject",
|
| 122 |
+
"@id": "episodes-jsonl",
|
| 123 |
+
"name": "episodes.jsonl",
|
| 124 |
+
"description": "Frozen navigation specification for 500 evaluation episodes, including exact target instances and positions, start positions and orientations, room metadata, and geodesic/euclidean distances.",
|
| 125 |
+
"contentUrl": "episodes.jsonl",
|
| 126 |
+
"encodingFormat": "application/jsonlines",
|
| 127 |
+
"sha256": "b9cd9699b90f6275e631c47d447abf52ee48c525d95668851948dafadb29a9bc",
|
| 128 |
+
"contentSize": "443469 B"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"@type": "cr:FileObject",
|
|
|
|
| 134 |
"description": "Dataset README with field schema and directory layout.",
|
| 135 |
"contentUrl": "README.md",
|
| 136 |
"encodingFormat": "text/markdown",
|
| 137 |
+
"sha256": "e5bfbe9d4e40239ed5a8550fc36709560447ef0ddf602047bff460392feb5100",
|
| 138 |
+
"contentSize": "3936 B"
|
| 139 |
}
|
| 140 |
],
|
| 141 |
"recordSet": [
|
episodes.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|