xiaojiahao/embodied-visual-intervention

Embodied Reference and Signal Collection

A research prototype containing 1,000 selected episodes, 3,628 reference images, and 90,700 single-factor signal interventions. Images and metadata total approximately 9.92 GB.

Source	Episodes	Reference images	Intervention images
LIBERO-R	82	267	6,675
RoboInter-Data / DROID	918	3,361	84,025
Total	1,000	3,628	90,700

Files and loading

reference_collection/
  references.json
  images/{dataset}/{episode}/progress_XX.png
signal_collection/
  signals.json
  images/{dataset}/{episode}/progress_XX/{factor}_{severity}.png

Images are distributed in ZIP shards under archives/ to avoid uploading 94,328 individual image files. archives.json lists image counts, sizes and SHA-256 checksums. ZIP storage preserves the PNG bytes exactly. The tree above is the layout AFTER extraction. Download as a file-based dataset. Each JSON's image paths are relative to that JSON's directory. Keep both top-level collection folders together; signals.json references the original reference images using ../reference_collection/ paths.

import json
import zipfile
from pathlib import Path
from huggingface_hub import snapshot_download
from PIL import Image

root = Path(snapshot_download(repo_id="xiaojiahao/embodied-visual-intervention", repo_type="dataset"))
for archive in sorted((root / "archives").glob("*.zip")):
    with zipfile.ZipFile(archive) as z:
        z.extractall(root)

references = json.loads((root / "reference_collection/references.json").read_text(encoding="utf-8"))
signals = json.loads((root / "signal_collection/signals.json").read_text(encoding="utf-8"))
sample = signals["references"][0]
original = Image.open(root / "signal_collection" / sample["reference_image"])
variant = Image.open(root / "signal_collection" / sample["variants"][0]["image"])

Reference selection and annotations

One episode per normalized instruction (case, whitespace and final punctuation). This is lexical deduplication, not a guarantee of semantically distinct tasks. All 82 LIBERO-R instruction strings are covered through libero-100-r; its 83 task IDs include a repeated instruction. RoboInter uses DROID episodes in index order, requires Primary instruction/subtask annotation quality, rejects empty instructions, and also deduplicates annotated instructions and source trajectories across cameras.

LIBERO references use the last frame of each distinct reasoning-update state (end-exclusive interpretation). RoboInter uses the first frame of each contiguous subtask plus the final observation. Episode metadata includes instruction, source revision, frame/timestamp information, subtasks, automatically extracted noun phrases, and original state/action chunks (up to 8 steps, truncated at episode end).

LIBERO images are 256x256; RoboInter images are 320x180 from the processed source videos. Original high-resolution video trials are NOT included. Original camera orientations and pixel dimensions are retained. Source frame indices are distinguished from episode-local indices where necessary.

Seven signal factors

Each factor is applied independently to the original reference. There are 25 nonzero variants per reference. Severity 0 reuses the original image without a duplicate file.

Factor	Nonzero severities	Parameters in order
luminance	-2,-1,+1,+2	linear exposure gains 0.35,0.65,1.3,1.6
contrast	-2,-1,+1,+2	mean-centered encoded-RGB gains 0.35,0.65,1.2,1.45
chromatic_information	-2,-1,+1,+2	linear-RGB chroma scales 0,0.5,1.3,1.6
spatial_frequency	-2,-1,+1,+2	Gaussian blur sigma 3/1.3 px; unsharp sigma 1, amount 0.25/0.5
spatial_sampling	-3,-2,-1	downsample to 1/8,1/4,1/2 then restore original dimensions
snr	-3,-2,-1	additive RGB Gaussian noise sigma 25,12,5 in 8-bit units
amplitude_capacity	-3,-2,-1	uniform 2,4,6-bit/channel quantization stored in RGB PNG

signals.json records reference IDs, paths, image hashes, actual transform parameters, noise seeds, implementation hashes and library versions. Noise strengths share a standard-normal field within a reference; seeds differ across references.

Limitations

• This is a reference/intervention collection, not a fully adjudicated benchmark or train/test split.
• RoboInter task_complete is null: episode termination is not treated as verified success. Intermediate progress assumes preceding annotated subtasks were completed; final progress is unknown.
• Automatically extracted perception noun phrases require review. No bounding-box GT is supplied.
• Original task/action annotations remain linked to the scene, but interventions can obscure evidence; image-only task solvability and label visibility are not guaranteed.
• Equal pixel-based blur parameters need not imply equal difficulty across source resolutions.
• No model evaluation results, downloaded video caches, credentials or model checkpoints are included.

Sources, attribution and terms

This collection redistributes selected source images/annotations and adds image sampling, metadata organization, lexical selection and deterministic signal transformations. Original data and annotation authors retain their rights. This card does not relicense third-party source material or grant rights beyond its upstream terms.

• NVIDIA LIBERO-R, revision 61ac0be5c084b61e60d50c21c97052b58a65a31b, based on LIBERO. Reasoning annotation work: Yilin Wu, Anqi Li, Tucker Hermans, Fabio Ramos, Andrea Bajcsy, Claudia Perez-D'Arpino, Do What You Say: Steering Vision-Language-Action Models via Runtime Reasoning-Action Alignment Verification (2025), https://arxiv.org/abs/2510.16281. The inspected pinned dataset README does not state a dataset license; consult the source maintainers and underlying dataset terms for intended reuse.
• InternRobotics RoboInter-Data, revision 0208b79c34eca4d214cdb5d07f5ae0f7cc634340, based on DROID. RoboInter's source notice specifies CC BY-NC-SA 4.0. Preserve attribution, noncommercial and share-alike obligations for applicable derived material, together with underlying DROID terms. See the RoboInter project for the authors' citations and documentation.

Where and What intervention release

This release adds the frozen Where collection and the available partial What (9/11) collection. The original reference and signal data, licensing and source attribution above remain unchanged.

Collection	Reference frames	Intervention images	Status
Where	3,628	174,848	Generated; 87,424 regions, each with removal and retain
What	3,628	32,652	Partial: nine methods; state and identity not generated

Where includes 87,424 binary masks, 87,424 feathered alpha masks and 3,628 copied reference PNGs, in addition to removal/retain images (353,324 PNGs total). Not every concept has a valid region on every reference; the manifest records actual available regions.

What methods available on all references: color, texture, shading, size, shape, material, parts, text_symbol and flat_2d. Quality scores exist for 6,456 images only; images are not filtered by quality score. Missing scores are not passes. Local judge reports and caches are excluded. State and identity must not be reported as evaluated in this snapshot.

Metadata: where_collection_final/where.json, what_collection/what.json, and each collection's methods.json. Paths are relative to the corresponding manifest. intervention_archives.json records ZIP sizes, SHA-256 checksums, file counts and collection coverage. Extract selected ZIPs at the dataset root, preserving their top-level folder. Existing archives.json continues to describe only the original reference/signal release.

Download and evaluate with the Where/What evaluation instructions. Cognition uses paired A/B orders; planning uses all cyclic option rotations. Source labels can become visually unanswerable after an intervention and are not new human annotations of edited images.

Evaluation correction: What means Where × What

The what_collection PNGs in this dataset are full-frame source assets, not the final What evaluation inputs. Official What evaluation combines each source What image with eligible Where regions once, replacing only the selected region, using the unchanged saved feathered alpha. Pixels outside the region are preserved, apart from the existing feathered boundary. For uint8 alpha w, composition is (reference*(255-w) + what*w + 127)//255. There is no complementary outside-region What variant. Only the standalone Where suite keeps both removal and retain modes.

The current assets support 786,816 Where × What composites (nine available What methods × 87,424 Where regions, with one inside-region replacement each). These composites are generated and cached locally on demand, not included as a second set of precomputed ZIP archives. State and identity remain unavailable. The current v5 selection intersects measured non-negligible pixel change with the frozen region/attribute eligibility policy, retaining 457,549 composites. Category eligibility alone retains 465,464; 7,915 of those fail the pixel-change rule. The pixel rule excludes exact equality or both mean absolute RGB difference below 0.25 uint8 levels and fewer than 0.1% of pixels with maximum-channel difference at least 8. The policy is an experimental scope, not a quality certification; no GT or model outputs enter selection. Source assets are unchanged. Qwen and PaliGemma now use fixed batch32, compact per-sample probabilities/raw outputs and per-batch durable writes. Both cognition and planning are queued, reference excluded; old single-item results are kept separately because batch inference can change numerical probabilities. See the linked instructions for preparation, exact eligibility matrix, resume and output paths.

Use the unified cognition/planning runner: one command loads a selected model once, evaluates reference, Where and Where × What, and saves three separate result directories. Reference is evaluated once. Old full-frame-only What results are diagnostic and must not be presented as Where × What evaluation.