RPX: Robot Perception X

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

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

Watch the RPX teaser video

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

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

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)

# 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:

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:

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

Citation

@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.