File size: 9,977 Bytes
65d1483
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76a5063
 
 
65d1483
 
 
 
76a5063
 
65d1483
 
9630685
65d1483
9630685
65d1483
9630685
91bad34
9630685
91bad34
 
9630685
65d1483
76a5063
 
 
9630685
65d1483
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76a5063
9630685
 
76a5063
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9630685
76a5063
9630685
 
76a5063
65d1483
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9630685
65d1483
 
 
 
 
 
 
76a5063
9630685
 
76a5063
9630685
 
 
65d1483
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
---
license: cc-by-4.0
pretty_name: "RPX: Robot Perception X"
task_categories:
- image-segmentation
- depth-estimation
- object-detection
- visual-question-answering
language:
- en
tags:
- robotics
- embodied-ai
- rgb-d
- benchmark
- perception
- manipulation
- stereo
- tabular
- image
- video
size_categories:
- 100B<n<1T
configs:
- config_name: multi_object
  description: "MOS phase records, tertile-cut by Effort-Stratified Difficulty."
  default: true
  data_files:
  - split: "easy"
    path: "splits/easy.parquet"
  - split: "medium"
    path: "splits/medium.parquet"
  - split: "hard"
    path: "splits/hard.parquet"
- config_name: media_preview
  description: "Data Studio media table with one video-layout RGB/depth/mask image preview per MOS phase."
  data_files:
  - split: "preview"
    path: "preview/media_preview.parquet"
- config_name: single_object
  description: "SOS selected-object catalog; one 360 collection per object, no difficulty split."
  data_files:
  - split: "objects"
    path: "manifest/selected_sos_objects_v1.parquet"
---

# RPX: Robot Perception X

A real-world RGB-D benchmark for evaluating robot perception under
embodied deployment conditions.

* Code: https://github.com/IRVLUTD/RPX

![RPX teaser](assets/rpx_teaser.png)

[Watch the RPX teaser video](assets/rpx-jumbotron.webm)

## Dataset at a glance

| | |
|---|---|
| Multi-object scenes (MOS) | **100** (`scene001` to `scene100`, 3 phases each: clutter / interaction / clean) |
| Single-object scenes (SOS) | **70 selected objects** (one 360 collection per object) |
| Frame manifest rows | **110,000** (75,000 MOS + 35,000 selected SOS) |
| MOS mask-object rows | **2,100** local mask IDs mapped to global object IDs |

Scene renames are documented in `manifest/scene_name_mapping_v1.csv`,
mapping original scene names to `scene001` through `scene100`.

The Hugging Face Dataset Viewer exposes `multi_object` as structured split
tables and includes a `media_preview` config with an actual image media column:
one video-layout RGB, depth, and mask preview still for each MOS phase.

## Effort-Stratified Difficulty (ESD)

ESD is the difficulty protocol used to split multi-object scenes. Each
`sceneXXX.phaseY` receives an RPX Difficulty Score (`rpx_ds`); higher scores
indicate harder perception conditions. The score is produced by
`effort_stratified_v1` using `primary_method: mean_pn`, a weighted aggregate of
the normalized feature set recorded in `splits/scene_splits.json`.

The released split files expose two levels:

| level | where | construction |
|---|---|---|
| phase-level splits | `splits/easy.txt`, `medium.txt`, `hard.txt` and CSV copies | all 300 scene phases are sorted by `rpx_ds` and tertile-cut into 100 easy, 100 medium, 100 hard entries |
| scene-level tiers | `splits/scene_splits.json` | each scene score is the mean of its three phase scores, then 100 scenes are tertile-cut into 33 easy, 33 medium, 34 hard scenes |

Because scene tiers are aggregated, a hard scene can still contain an easy
phase. Use the phase-level split when downloading or evaluating MOS tasks.

The ESD feature weights are fully accounted for in
`splits/scene_splits.json`: 27 feature names, 27 weights, no missing weights,
and total weight `1.0`. `iter_mean` and `iter_max` each carry weight `0.125`;
the other 25 features each carry weight `0.03`.

| feature group | features |
|---|---|
| refinement effort | `iter_mean`, `iter_max` |
| object complexity | `obj_mean`, `obj_std`, `obj_consist` |
| occlusion | `occ_mean`, `occ_p90`, `occ_heavy` |
| depth quality | `depth_invalid`, `depth_invalid_mask`, `depth_std`, `depth_std_mask` |
| image appearance | `specular`, `dark` |
| mask/visibility stability | `area_cv`, `area_drop`, `vis_instability` |
| motion and trajectory | `trans_mean`, `trans_p90`, `rot_mean`, `rot_p90`, `jerk` |
| fisheye quality | `fisheye_dark`, `fisheye_bright`, `fisheye_sharpness`, `fisheye_corr`, `fisheye_texture` |

The split tables include the final `rpx_ds`, `difficulty`, `scene_tier`, and
`scene_score` fields for Dataset Viewer filtering. Raw per-feature values are
not duplicated in those tables; `splits/scene_splits.json` is the source of
truth for the score/tier assignments and the feature/weight provenance.

## Modality inventory

The table below describes this cleaned release. `cam_pose_icp` is not included in this cleaned release.
Use `manifest/frames_v1.parquet` and the identity manifests under `manifest/`
as the source of truth.

| modality | files | bytes |
|---|---:|---:|
| `cam_pose` | 110,000 | 163.3 MB |
| `depth` | 110,000 | 16.9 GB |
| `fisheye` | 220,000 | 63.4 GB |
| `rgb` | 110,000 | 41.0 GB |
| `masks` | 110,000 | 341.7 MB |

## Quick start

```bash
pip install "rpx-benchmark[hub]"
hf auth login
```

```python
from rpx_benchmark.dataset_hub import download_for_task

# Pull just RGB + masks for the Easy difficulty tier — never the whole repo.
res = download_for_task(task="segmentation", split="easy",
                          repo_id="IRVLUTD/RPX")
print(res.local_dir, res.matched_scenes)
```

```bash
# Or from the CLI:
python -m rpx_benchmark.dataset_hub.cli download \
    --task segmentation --split easy \
    --repo-id IRVLUTD/RPX
```

A subsequent call for a different task on the same split (e.g.
`relative_pose`) reuses the cached RGB tars and only fetches the new
modality (`cam_pose`) as the delta.

## Repo layout

```
IRVLUTD/RPX/
├── manifest/
│   ├── frames_v1.parquet                 # per-frame metadata
│   ├── scene_name_mapping_v1.csv         # original scene names to scene001..scene100
│   ├── selected_sos_objects_v1.csv       # selected 70-object SOS catalog
│   ├── selected_sos_objects_v1.parquet   # Dataset Viewer SOS catalog table
│   ├── object_catalog_v1.json            # SOS object/global-ID catalog
│   ├── mos_raw_mask_object_map_v1.csv    # MOS local mask IDs from sam2 metadata
│   ├── mos_mask_object_map_v1.csv        # MOS local mask IDs joined to SOS global IDs
│   ├── mos_mask_object_map_v1.parquet    # parquet copy of the MOS map
│   └── current.json                      # default version per label modality
├── splits/
│   ├── scene_splits.json
│   ├── easy.txt  medium.txt  hard.txt      # phase-level split IDs
│   ├── easy.csv  medium.csv  hard.csv      # human-readable split tables
│   └── easy.parquet  medium.parquet  hard.parquet  # Dataset Viewer split tables
├── preview/
│   ├── media_preview.parquet               # Dataset Viewer image media table
│   ├── data_studio_preview.csv             # preview index with source shard links
│   └── image_examples/preview/             # source JPEGs for the media preview
├── scenes/<scene_id>/<phase>/                     # MOS
│   ├── rgb.tar  depth.tar  fisheye.tar
│   └── labels/{cam_pose,masks,masks_aux,sam2_meta}/v1.tar
├── objects/<object_id>/0/                         # SOS
│   └── (same modality structure)
├── objects_meta/                                  # questionnaire dedup
│   ├── _index.json
│   └── <object_id>/questionnaire.json
└── README.md   ←  this file
```

## Object identity manifests

SOS objects use dataset-wide IDs from `manifest/selected_sos_objects_v1.csv`
and `manifest/object_catalog_v1.json`.

| field | meaning |
|---|---|
| `global_object_id` | New integer ID, 1 to 70 |
| `source_catalog_id` | Original catalog/PDF ID, kept as a string such as `88.2` |
| `object_id` | Actual folder name, such as `mug.2` |
| `questionnaire_path` | Linked questionnaire under `objects_meta/<object_id>/questionnaire.json` |

MOS masks use local IDs. A `local_mask_id` is only meaningful within one
`scene_id + phase`; it is not a global object ID. The raw source is
`scenes/<scene_id>/<phase_index>/labels/sam2_meta/v1.tar:sam2/mask_to_object.json`.

Use `manifest/mos_mask_object_map_v1.csv` or
`manifest/mos_mask_object_map_v1.parquet` to join MOS masks to SOS objects:

```csv
scene_id,phase,local_mask_id,object_id,global_object_id,source_catalog_id,object_name
scene001,phase0,2,boot.2,11,18.2,boot
```

That row means `scene001/0` mask ID `2` is `boot.2`, whose global object ID
is `11`, with questionnaire
`objects_meta/boot.2/questionnaire.json` and SOS template `objects/boot.2/`.

## Tasks

### Multi-object (use a difficulty split)

| recipe | inputs → labels |
|---|---|
| `monocular_depth` | ['rgb'] → ['depth'] |
| `rgbd_segmentation` | ['depth', 'rgb'] → ['masks'] |
| `segmentation` | ['rgb'] → ['masks'] |
| `relative_pose` | ['rgb'] → ['cam_pose'] |
| `rgbd_relative_pose` | ['depth', 'rgb'] → ['cam_pose'] |
| `stereo_depth` | ['fisheye'] → ['depth'] |
| `object_tracking` | ['rgb'] → ['masks'] |
| `vqa` | ['rgb'] → ['questionnaire', 'vqa'] |

### Single-object (no split — these are object templates)

| recipe | inputs → labels |
|---|---|
| `object_templates` | ['rgb'] → ['masks'] |
| `object_templates_rgbd` | ['depth', 'rgb'] → ['masks'] |
| `object_pose_library` | ['depth', 'rgb'] → ['cam_pose', 'masks'] |

## Label versioning

Labels live at `labels/<name>/v<N>.tar`. Newer versions land at new
paths; old versions stay reachable for reproducibility.

| modality | current version |
|---|---|
| `masks` | `v1` |
| `masks_aux` | `v1` |
| `sam2_meta` | `v1` |
| `cam_pose` | `v1` |

To pin to a specific version:

```python
download_for_task(
    task="relative_pose", split="easy", repo_id="itaykadosh/RPX",
    label_versions={"cam_pose": "v1"},   # don't auto-upgrade to v2
)
```

## Citation

```bibtex
@misc{rpx2026,
    title  = {RPX: Robot Perception X — A real-world RGB-D benchmark for
              embodied perception},
    author = {IRVL UT Dallas},
    year   = 2026,
    url    = {https://huggingface.co/datasets/itaykadosh/RPX},
}
```

## License

Released under the **cc-by-4.0** license.