RPX: Same Scene, Different Story

Evaluating Robot Perception Across Scene Phases in the Wild

Jishnu Jaykumar P*, Itay Kadosh*, Narendhiran Vijayakumar*, Srinanditha Kamath, Sai Haneesh Allu,
Govind Tyagi Rangappa, Animesh Maheshwari, Jikai Wang, Yu Xiang
_{UT Dallas · UF Gainesville · IIIT Hyderabad · UT Austin *equal contribution}

Toolkit docs · Code · pip install rpx-benchmark · Docker images · Cite

A robot's scene does not hold still. A tabletop starts cluttered, a hand rearranges it, and it ends clean. RPX captures the same real scenes before, during and after manipulation on robot-grade sensors, and scores six perception tasks on them through one pipeline. Two numbers summarize every model on a common scale: phase robustness Φ (does quality hold as the scene changes?) and worst-phase quality 𝒥_min (how good is it at its weakest?).

100	3 + 1	70	133,121	6	65
real scenes 50 indoor · 50 outdoor · 20 sites	phases + ego view Clutter · Interaction · Clean · Ego	everyday objects each scanned alone in 360°	annotated frames RGB-D · stereo · pose · masks	tasks depth to grounding	zero-shot evaluations

What it shows. Depth and tracking change significantly across phases (p < 0.001); model rankings shift across tasks, prompting modes and input modalities on the very same scenes; and no evaluation reaches both 𝒥_min ≥ 0.75 and Φ ≥ 0.86.

Six tasks, one pipeline

	Task	Scored on	Native metrics	Reference runtime
T1	Image depth	3 phases	AbsRel · RMSE · SILog · δ1	irvlutd/rpx:depth-zipdepth-latest
T2	Video depth	3 phases	T1 metrics + TGM · TGSE	irvlutd/rpx:depth-zipdepth-latest
T3	Multi-object tracking	3 phases + Ego	HOTA · MOTA · IDSW	irvlutd/rpx:tracking-dam4sam-rpx-latest
T4	Relative camera pose	Clutter · Clean	AUC@5° · @10° · @20°	irvlutd/rpx:rcpe-monst3r-rpx-latest
T5	VQA localization	3 phases + Ego	Acc@0.5 · GIoU · Center-in-GT	irvlutd/rpx:vqa-vllm
T6	In-context VQA	3 phases + Ego	Acc@0.5 · GIoU · Center-in-GT	irvlutd/rpx:vqa-vllm

Every task keeps its field's native metrics; each result is mapped onto the same scene–phase scale, from which Φ and 𝒥_min are computed. The Docker guide lists which image holds which model.

Get started in two minutes

python -m pip install 'rpx-benchmark[hub,schemas]'

# Check the install on all six tasks with synthetic data (CPU, no downloads):
python -m rpx_benchmark.examples.benchmark_tasks --task all --smoke --output results/six-task-smoke

Score your own model. Wrap any callable, local model or API, and the toolkit fetches only the shards that task needs, runs it, and writes the metrics:

import rpx_benchmark as rpx
from my_model import predict_depth          # RGB array -> depth in metres

model = rpx.make_numpy_depth_model(predict_depth, name="my-depth")
result, report, paths = rpx.run_monocular_depth(
    rpx.MonocularDepthRunConfig(
        model=model, split="easy", repo_id="IRVLUTD/RPX",
        output_dir="rpx_results/my-depth/easy",
    )
)
print(result.aggregated)

The same pattern covers video depth, tracking, relative pose and VQA; see the toolkit docs. The toolkit pins a dataset commit by default (rpx.DEFAULT_REVISION), so two runs always read identical data. Do not download the whole 237 GB repository for one task.

Just browse the data:

from datasets import load_dataset

easy = load_dataset("IRVLUTD/RPX", "multi_object", split="easy")   # the paper's Easy tier: 33 scenes × 3 phases
objects = load_dataset("IRVLUTD/RPX", "single_object", split="objects")

Difficulty tiers

Scenes are stratified into Easy (33), Medium (33) and Hard (34) by annotation-refinement effort, the number of correction passes a scene's masks needed, as a proxy for perceptual ambiguity. A tier always contains all three phases of its scenes, exactly as in the paper and in every benchmark manifest.

File	Contents
splits/{easy,medium,hard}.parquet (.csv, .txt)	one row per scene–phase of that tier: 99 / 99 / 102 rows
splits/scene_splits.json	scene lists per tier, scores, per-phase scores, and scoring provenance
manifests/<task>/<tier>.json	the exact samples each benchmark task evaluates

Each split row also keeps the phase's own score (rpx_ds) and the tier that score alone would give (phase_difficulty), for analyses at phase level. Ego captures are not scored for difficulty; they follow their scene's tier.

What is included

Collection	Coverage	Modalities
Multi-object scenes (MOS)	100 scenes × 3 phases × 250 frames = 75,000	RGB, metric depth, fisheye stereo, 6-DoF pose, instance masks, tracklets
Egocentric view (Ego)	100 Interaction clips, 23,121 frames	RGB (1920 × 1080), instance masks, tracklets
Single-object scans (SOS)	70 objects × 500 frames = 35,000	RGB, metric depth, fisheye stereo, pose, masks, questionnaire
VQA question bank	4,529,322 single-image + 749,083 two-image records	attribute, spatial-box and spatial-binary questions
Benchmark VQA subset	31,500 runnable questions	fixed, seeded evaluation manifests
Tracking text prompts	70 objects, 2,739 scene–condition entries	stable IDs, canonical names, fixed prompts

Capture: Intel RealSense D435 (RGB-D, 640 × 480) with a T265 (fisheye stereo and visual–inertial pose) on a hand-held rig, plus a head-mounted GoPro during Interaction. Every object keeps one global identity across phases, views and the single-object catalogue.

Repository layout and file formats

assets/                         dataset-card media
manifest/                       canonical inventories, object catalog and checksums
manifests/<task>/<tier>.json    executable benchmark sample manifests
objects/                        70 single-object scan shard trees
objects_meta/                   per-object metadata and questionnaires
preview/                        lightweight MOS and Ego previews
scenes/                         100 MOS phase trees and 100 Ego view trees
splits/                         difficulty tiers and scoring provenance
tracking/metadata/              text-initialization vocabularies and checksums
vqa/                            full VQA Parquets and benchmark JSONL manifests

Large modalities are packed in tar shards. RGB and fisheye images are lossless WebP; depth and masks are PNG; camera poses are (7,) float64 NPY vectors [x, y, z, qx, qy, qz, qw]. Labels are versioned under labels/<name>/v<N>.tar. manifest/frames_v1.parquet is the canonical frame inventory (133,121 records, with absent Ego depth, fisheye and pose marked explicitly); frames_v2.parquet is historical only.

Object and mask identifiers

Field	Scope and rule
global_object_id	canonical integer identity, unique and contiguous from 1 to 70
source_catalog_id	original catalog identity, stored as a string; dotted values such as "18.2" must stay strings
object_id	canonical SOS folder key, for example boot.2
local_mask_id / mask_index	local to one scene condition; join with scene, view and phase before mapping to a global ID
sample_id	stable identifier of a VQA record; never derive object identity from it

The source of truth is manifest/object_catalog_v1.json. Join MOS masks through manifest/mos_mask_object_map_v1.parquet and tracking prompts through tracking/metadata/text_initialization_v1/scene_condition_vocab.parquet.

VQA files

File	Rows	Purpose
vqa/attribute.parquet	3,241,564	colour, material, function, composition, count and related attribute questions
vqa/spatial_bbox.parquet	543,017	spatial questions with box answers
vqa/spatial_binary.parquet	744,741	left/right and above/below questions
vqa/incontext_mos_attribute_bbox.parquet	317,873	MOS two-image attribute transfer
vqa/incontext_ego_attribute_bbox.parquet	113,193	Ego two-image attribute transfer
vqa/incontext_mos_spatial_bbox.parquet	318,017	MOS two-image spatial reasoning

The runnable manifest is vqa/metadata/benchmark_subset_v1/benchmark_available_31500.jsonl. benchmark_pending_4500.jsonl holds reserved descriptors, not runnable questions: Ego has no depth, so those spatial questions are not fabricated. A model answers with a target-image box; the evaluator never passes ground truth to the model.

Reproducibility

• Pin the dataset commit you used and report it with your results.
• Verify downloads with the SHA-256 manifests under manifest/, vqa/metadata/ and tracking/metadata/.
• manifest/current.json is the active metadata index.
• Read source_catalog_id as text from CSV; prefer Parquet or JSON when types matter.

Citation

@misc{rpx2026,
  title  = {Same Scene, Different Story: Evaluating Robot Perception Across Scene Phases in the Wild},
  author = {{Jishnu Jaykumar P} and Kadosh, Itay and Vijayakumar, Narendhiran and Kamath, Srinanditha and
            Allu, Sai Haneesh and Rangappa, Govind Tyagi and Maheshwari, Animesh and Wang, Jikai and Xiang, Yu},
  year   = {2026},
  note   = {Dataset: \url{https://huggingface.co/datasets/IRVLUTD/RPX}}
}

License

The RPX dataset is released under Creative Commons Attribution 4.0 (CC BY 4.0): share and adapt it with attribution. The benchmark code is MIT-licensed in the GitHub repository.