Multimodal episode pipeline
One Xperience-10M public sample episode is converted into aligned windows and a documented feature contract.
Project overview and contributions.
The page is organized like a compact research project: motivation and scope, dataset sample, task suite, method, baselines, research directions, interactive walkthroughs, and resources for continuing the work.
Task suite and baseline heads
Every core task has a minimal baseline and a compact PyTorch MLP head over the same windows, splits, and labels.
Dataset source alignment
The public description is aligned to the official gated Xperience-10M dataset card, including modalities, scale, access, and current project coverage.
Public research artifacts
Metrics, figures, walkthrough data, baseline weights, and project metadata are packaged across GitHub, GitHub Pages, and Hugging Face.
Omni-model scale-up path
The 32-episode LoRA path is prepared; full training results require gated data access, held-out splits, training, and evaluation.
Data governance
Raw MP4/HDF5/RRD files, private gated Xperience-10M data, and full Qwen weights are excluded from the public repo and HF mirrors.
Evaluation protocol is explicit.
The protocol is generated from committed metric artifacts so readers can see the exact data unit, split, task targets, leakage controls, and current limitations before comparing scores.
Data unit
One 20-frame aligned window from the public sample episode, stride 5 frames, 1,161 windows total, represented by the current 8,378-d feature vector.
Split policy
Single-episode chronological 70/30 train/test split. This avoids random future-window mixing; cross-episode generalization is measured in the later multi-episode pilot.
protocol docMetric contract
All 12 tasks list input, target, primary metric, minimal baseline score, and neural MLP score from committed result files.
summary metricsLeakage controls
Scalers fit on train windows only; future labels, target feature blocks, caption/object labels, and contact labels stay on the target side unless explicitly queried.
builder scriptCurrent limits
This public-sample run does not evaluate cross-episode generalization, audio-visual learning, pixel-depth reconstruction, neural rendering, or full 32-episode Qwen3-Omni training.
status checkScale-up gate
The Omni pilot requires at least 32 valid episodes, held-out episode splits, no train/test episode leakage, training metadata, predictions, metrics, and a run report.
data gateResearch progress and next milestones.
The project shows the completed public-sample task suite, then lays out the data requirements for the Qwen3-Omni scale-up path.
Aligned Xperience-10M sample windows
5,821 frames become 1,161 synchronized 20-frame windows with an explicit 8,378-d feature contract.
12 minimal heads + 12 neural MLP heads
Every task has a minimal interpretable head and a matching neural MLP run over the same windows, splits, and task contract.
Four research directions are mapped by evidence type
The Ropedia directions are labeled as direct, proxy, or diagnostic coverage, plus one coded extension probe per direction.
Qwen3-Omni pilot setup
The current Qwen3-Omni artifacts use one episode and 128 train windows. The 32-episode evaluation is still pending.
Scale-up status is machine-checked
The status check confirms historical 32ep run/path strings stay in setup-file provenance and are not presented as completed 32-episode results.
Prepared mirrors are byte-checked
The parity report compares critical JSON, figure, and validator files across the repo, HF Space bundle, artifact dataset bundle, and model bundle before upload.
Figures are indexed as evidence
The figure index records public visual assets, dimensions, SHA-256 hashes, source scripts, and the role each figure plays in the project narrative.
Brand assets are packaged consistently
The generated logo system is packaged into the website header, favicon, README/HF cards, Open Graph preview, and brand-asset manifest.
Publication bundles are checked before release
The validator checks required assets, raw-data exclusion, Python cache exclusion, heavy archive exclusion, accidental HF token strings, and public-card figure freshness across GitHub and the HF bundles.
Website integrity is checked
The site validator checks local links, anchors, JSON bundles, and referenced image dimensions before publishing.
Release gates are explicit
The quality-gate manifest collects the automated validators and the live post-publish checks required before the release is presented as current.
Official dataset card is aligned
The source-alignment note mirrors the public Hugging Face dataset-card facts, sample-card facts, and API metadata: gated access, sample license/tooling, modality coverage, episode layout, intended uses, and current project coverage.
Research reading path.
A newcomer should be able to move from the dataset sample to the task design, model baselines, current limitations, and scale-up plan without reading every file first.
Understand the current scope
Start with the project status, evidence contract, artifact index, scale-up status check, publication package check, and website integrity report. They separate implemented single-episode work from the prepared Qwen3-Omni scale-up.
Inspect one model input
Use the window table and feature manifest to see the exact aligned sample unit, feature blocks, dimensions, and omitted audio feature status.
Compare minimal vs neural heads
Every task has a small interpretable baseline and a matching neural MLP head over the same feature contract and chronological split.
Check the scale-up gate
The multi-episode Qwen3-Omni path is prepared. The 32-episode result will be added after the data gate and held-out evaluation pass.
Verified nowOne public episode, 5,821 frames, 1,161 windows, 8,378 current features, 12 minimal heads, 12 neural heads, and 4 direction-extension probes.
Pending scaleA 32-episode held-out Qwen3-Omni LoRA pilot is gated on Xperience-10M access and must pass manifest, training, and evaluation checks.
Not redistributedRaw videos, raw annotations, full Qwen weights, and private gated Xperience-10M data are not included in the public repo or HF bundles.
Aligned with the official dataset card.
The official Xperience-10M card describes a gated, large-scale 4D egocentric multimodal dataset. This project records that full upstream scope while focusing the implemented artifacts on one public sample episode.
Official scale
About 10M experience units and 10,000 hours, with RGB video, audio, depth, camera pose/SLAM, hand/body mocap, IMU, captions, metadata, and calibration.
HF file-size display
The live Hugging Face page/API currently shows 31.9 TB hosted. This is recorded separately from the card's about-1PB full-scale storage statement.
source JSONHF access path
The source dataset is manually gated for approved non-commercial use, with an external agreement step noted by the public HF metadata.
official HF datasetAPI listing snapshot
HF API metadata observed 803 session folders and 12,103 episode folders with annotation.hdf5. This snapshot supports planning; it is not a local data inventory.
Public sample card
The sample repo lists cc-by-nc-4.0, HOMIE Toolkit for videos/annotations, and Rerun 0.29.0 for .rrd visualization.
Source alignment
The source-alignment validator checks full-dataset facts, public sample-card facts, API-listing caveats, and current-project markers across the repo, website, and HF cards.
alignment reportEpisode layout
Expected folders contain six MP4 streams and annotation.hdf5; visualization.rrd is treated as a viewer artifact and excluded from training downloads.
Current project subset
One public sample episode, 5,821 frames, 1,161 windows, 8,378 current features, audio documented but not yet featurized, and no raw-data redistribution.
modality atlasCovered now
Action/subtask labels, next-action prediction, temporal diagnostics, hand trajectory, contact, object relevance, caption grounding, retrieval, reconstruction, and misalignment.
summary metricsResponsible use
The official card notes limited diversity and showcase/production quality. This project excludes identity, surveillance, biometric, sensitive-attribute, and safety-critical uses.
use notesLater milestones
Full audio-visual learning, caption generation, depth-pixel prediction, SLAM estimation, neural rendering, policy learning, cross-episode generalization, and 32-episode Qwen3-Omni evaluation.
data gateRopedia Xperience-10M 12-task suite, first.
Start with the full 12-task map, then inspect the large native modality atlas below it. Audio is present in the sample MP4 stream, but the current 8,378-d baseline manifest does not featurize it.
Readable modality atlas.
Each Xperience-10M stream gets a large thumbnail, a plain sample-content line, and the exact current-baseline use. These are small derived images only; no raw MP4, HDF5, or RRD data is redistributed.
01
visual streamVideo
sample contains
6 synchronized camera MP4 streams
current baseline use
RGB/fisheye/stereo frame statistics
02
acoustic streamAudio
sample contains
AAC stream embedded in MP4
current baseline use
Documented, not featurized in the 8,378-d vector
03
geometry mapDepth
sample contains
Depth map + confidence channel
current baseline use
Spatial geometry feature block
04
camera posePose / SLAM
sample contains
Trajectory + sparse SLAM map
current baseline use
Position + orientation features
05
human motionMotion Capture
sample contains
Body + hand joint tracks
current baseline use
3D mocap feature statistics
06
wearable sensorInertial
sample contains
Accelerometer + gyroscope
current baseline use
Wearable motion statistics
07
semantic annotationLanguage
sample contains
Object tags + action captions
current baseline use
Task labels + semantic targets
The atlas redistributes only small derived thumbnails and metadata. Raw MP4, HDF5, and RRD files remain excluded from this repo and the Hugging Face mirrors.
From raw episode to checked artifacts.
Every script works from one data contract: aligned multimodal windows, explicit labels, cached feature extraction, and a manifest that makes omitted modalities visible.
What this project proves
It demonstrates the full development loop: reading Xperience-10M sample data, aligning modalities, converting them into model-ready windows, defining meaningful tasks, producing metrics, and packaging artifacts for continued research.
What it does not prove
It does not claim general embodied intelligence. A single episode cannot support cross-environment generalization; that requires many episodes and held-out episode splits.
What the current results actually say.
A generated takeaways layer reads the committed metrics and separates useful research signals from claims that still require held-out episodes.
One episode becomes a benchmark contract
The public sample is converted into 5,821 frames, 1,161 aligned 20-frame windows, and an 8,378-dimensional feature contract.
Chronological split exposes class shift
All-feature action reaches 0.9791 macro-F1 on its local split, while the 12-task chronological action head is 0.0500 macro-F1 with four unseen later action labels.
takeawaysNeural heads help dynamics
Hand MPJPE improves from 0.8223 to 0.1116; temporal-order F1 rises from 0.5487 to 0.8718; misalignment F1 rises from 0.4866 to 0.7335.
metricsRetrieval and reconstruction remain open
Ridge/cosine retrieval remains stronger than the neural projection here, and cross-modal feature reconstruction still has negative R2.
retrieval metricsScale means held-out episodes
The next credible model-quality unit is a 32-episode held-out pilot across 32 sessions, not more adjacent windows from one sample.
scale-up statusSmall baselines, no hidden machinery.
Motion-only and current all-feature classifiers use lightweight heads so the comparison stays readable on a laptop and easy to inspect. The neural run keeps the same features and splits, then swaps in PyTorch MLP heads.
Motion-only action
0.9688Current all-feature action
0.9791Motion-only subtask
0.9528Current all-feature subtask
0.9308
Neural MLP heads, same task contracts.
The neural baseline uses small PyTorch MLP classifiers/regressors on the same 8,378-d window features, chronological splits, and leakage filters. This isolates the value of a nonlinear head before moving to heavier Qwen/Omni experiments.
Neural hand forecast
0.1116Neural temporal order
0.8718Neural misalignment
0.7335Neural cross-modal retrieval
0.1530
The 12 tasks organized into four research directions.
Each task is mapped as direct, proxy, or diagnostic evidence for the Ropedia research tracks. The mapping uses two current baselines: minimal interpretable heads and neural MLP heads over the same feature contract.
A. Human Modeling & Motion Understanding
Direct evidence comes from hand trajectory forecasting and contact prediction; action and object relevance are supporting proxies.
2direct2proxy0diagnostic
B. 3D/4D Reconstruction & Neural Rendering
Cross-modal retrieval, modality reconstruction, and misalignment detection check reconstruction prerequisites, not full geometry.
0direct2proxy1diagnostic
C. Egocentric Vision & Interaction
Action, subtask, transition, next-action, object, caption, order, and alignment tasks directly stress egocentric understanding.
6direct2proxy3diagnostic
D. Scene Reconstruction & World Modeling
Current probes cover task state, object relevance, retrieval, reconstruction, temporal order, and alignment but no persistent map yet.
0direct6proxy3diagnostic
Baseline 1: minimal heads
Softmax, logistic, ridge, and retrieval heads keep every input/output contract readable. They are the first sanity check for whether a task is well-posed.
Baseline 2: neural MLP heads
Small PyTorch MLP classifiers/regressors reuse the same features and splits. They test nonlinear gains before heavier Omni fine-tuning.
Four extra probes make the directions actionable.
These are new data-backed extension tasks computed from the same single-episode feature tensor. They add one concrete input, process, output, and metric for each research direction, while keeping the single-episode limitation explicit.
Body and Hand Motion Intensity
Case: classify fast reach/pour windows as high motion and steady holding windows as low motion.
Input: non-mocap video, depth, pose, IMU, SLAM, calibration, and language features.
Output: high_motion or low_motion.
0.7827minimal macro-F10.7986neural macro-F1
Multi-View Consistency Retrieval
Case: retrieve the synchronized stereo-left window from a fisheye-camera query.
Input: fisheye_cam0 video features against stereo_left candidate features.
Output: ranked synchronized view candidates.
0.5534minimal MRR0.3469neural MRR
Action Phase Progress Estimation
Case: estimate whether a Pour coffee window is near the start, middle, or end of its action segment.
Input: non-caption multimodal features.
Output: 0-to-1 progress inside the current action.
0.3416minimal MAE0.3038neural MAE
Short-Horizon Ego-Motion Forecasting
Case: predict how the camera translation changes over the next 20 frames.
Input: current sensors excluding camera translation and captions.
Output: future camera-translation delta vector.
0.1989minimal MAE0.0989neural MAE
What changed
The four research directions now have coded extension probes, prediction/rank CSVs, JSON metrics, a Markdown summary, and a website chart generated from real sample-window features.
What still needs scale
A full research result still needs many Xperience-10M episodes, held-out episode splits, stronger encoders, and direction-specific models such as body priors, renderers, or persistent scene graphs.
The 12 tasks share four head families.
The diagram separates the shared episode-window feature pipeline from the task-specific heads, and notes that audio remains dataset context rather than a current baseline feature block.
Interactive task walkthrough.
Each task uses a common research name and a concrete case study, then opens into the input, middle modules, output, modality evidence, metric, and current limitation.
Step 1 / 4 · Input
Action Recognition
Egocentric Action Recognition
Input: inspect the 20-frame multimodal window before choosing the target.
01 / 12
supervised
multiclass classifier
Action Recognition
In the coffee-making sample, a pouring window maps to the current action label.
Metric: macro-F1. Minimal 0.0500; neural MLP 0.0263.
Current limitation: single-episode chronological split.
Task cards and metrics.
The 12 task cards use readable research names, representative modality thumbnails, explicit input-process-output contracts, and verified minimal versus neural scores from the committed result files.
Every feature block has a source.
The point is not hidden complexity. Every block has a source modality, a dimensional footprint, and a manifest entry.
Diagnostics separate memorization from signal.
The charts make the main lesson visible: within-episode supervised labels are easy under some splits, while retrieval, grounding, forecasting, and alignment remain the useful probes.
Research artifacts for the next experiments.
Metrics, predictions, manifests, lightweight model weights, and derived window artifacts are organized so the project can be inspected, extended, and scaled before rerunning the full pipeline. Raw Xperience-10M data and Qwen weights are not redistributed.
Research artifacts
From one episode to task heads
Start with the files that define the sample windows, feature blocks, task contracts, metrics, walkthroughs, and research-direction mapping.
Task-suite report
One JSON file with every task definition, split detail, feature dimension, and minimal/neural metric.
Windows table
Window start/end frames and aligned action/subtask labels for the public sample episode.
windows.csvFeature manifest
Start/end index and dimension for every current feature block in the 8,378-d window vector.
feature_manifest.jsonNeural MLP task results
Per-task PyTorch MLP metrics, predictions, histories, and checkpoints for the same 12 task contracts.
neural_mlp/Four-direction taxonomy
Generated JSON, CSV, Markdown, and website data mapping all 12 tasks to the four research tracks.
research_directions/Direction extension probes
Four coded probes, one per research direction, with minimal and neural metrics plus prediction/rank CSVs.
research_direction_extensions/Task walkthroughs
Case studies for all 12 tasks, including input, middle process modules, output, metric, limitation, and task-player data.
task_walkthroughs/Cross-modal retrieval
The strongest self-supervised signal from the single episode.
metrics.jsonPublic surfaces
Project map, mirrors, and runnable code
Use these files to navigate the whole project, open the published mirrors, or reproduce the public-sample pipeline.
Artifact guide
Human-readable map from project scope to data contract, task evidence, platform mirrors, and scale-up status.
Reproduction scripts
Training, visualization, taxonomy, walkthrough, validator, and omni-readiness scripts.
scripts/Hugging Face Space
The dashboard packaged as a public static Space.
HF SpaceDerived HF artifacts
Metrics, predictions, docs, and lightweight derived files without raw data redistribution.
dataset repoHF baseline models
Minimal NumPy softmax, ridge baselines, and neural task-head model files.
model repoHF collection
Space, artifacts, and model baselines grouped into one public project collection.
collectionCurrent all-feature action model
Classifier metrics, predictions, confusion matrix, and model weights.
metrics.jsonProject packet
Machine-readable 90-second project path and scope summary.
project_packet.jsonScale-up path
Prepared for multi-episode training
The multi-episode Qwen3-Omni path is documented and scripted. Full-pilot metrics come after the data gate and held-out evaluation pass.
Project scope
Connects implemented single-episode artifacts, setup-stage Omni work, pending data access, and later multi-episode milestones.
Multi-episode access status
Public data-access path and selected 32-episode pilot plan, without private infrastructure details.
MULTI_EPISODE_ACCESS_STATUS.mdQwen3-Omni setup artifacts
Manifests, metadata, metrics, and progress logs from the current setup run.
episode_manifest.json32-episode data gate
The data gate defines what must be available before full pilot training and held-out metrics.
DATA_BLOCKER_REPORT.mdConsistency checks
Release checks behind the research site
These validator outputs support the public research artifacts by keeping links, mirrors, figures, source wording, and package boundaries consistent.
Artifact index
Selective source-of-truth catalog with existence checks, sizes, and stable-file hashes.
artifact_index.jsonTask-surface integrity
Checks that public task cards use readable research names, modality thumbnails, and the interactive walkthrough/player contract.
task_surface_integrity.jsonWebsite integrity
Checks local links, anchors, JSON files, and referenced website image dimensions.
website_integrity.jsonQuality gates
One release checklist for automated validators and live post-publish checks.
quality_gates.jsonMirror parity
Prepared repo, HF Space, artifact dataset, and model bundle parity for critical data, figures, website HTML, and validator files.
mirror_parity.jsonLive publication
Last public GitHub/HF URL verification after upload.
live_publication_status.jsonScale-up status check
Machine check that historical 32ep identifiers stay in setup-file provenance and are not presented as completed 32-episode results.
Public presentation check
Checks repo, website, and Hugging Face presentation quality, accessibility semantics, public links, and copy consistency.
public_surface_qa.jsonPublication package check
Checks raw-data exclusion, cache exclusion, heavy-archive exclusion, token-string scanning, and public-card figure freshness.
publication_audit.jsonQwen3-Omni pilot is approval-ready.
The full Xperience-10M Hugging Face dataset is gated. While access is pending, the public plan has selected a 32-episode pilot across 32 different session UUIDs.
Selection
Stratified round-robin over 64 top-level sessions; 680 complete candidates scanned; 32 sessions selected.
Transfer
Download raw episodes only from official gated sources, exclude visualization.rrd, validate files, then stage them for training.
Boundary
The current LoRA artifact is a readiness checkpoint. A real 32-episode result requires local gated data and held-out evaluation.
Reproduce the suite.
Raw Xperience-10M data is not redistributed here. The public reproduction contract states the commands, expected outputs, exact-match reproduction evidence, and current non-reproducible scale-up boundary.
Reproducibility contract
Human-readable commands, expected artifacts, and boundaries for the public single-episode pipeline.
REPRODUCIBILITY.mdReproducibility matrix
Machine-readable command matrix covering sample download, baselines, 12 tasks, figures, and validation.
reproducibility_matrix.jsonExact-match reproduction check
The last metric check rebuilt the public-sample outputs from fresh cache and matched the committed metrics.
reproducibility_audit.mdWebsite integrity
Local HTML references, anchors, JSON bundles, and image dimensions checked before publishing.
website_integrity.jsonScale-up boundary
The 32-episode Qwen3-Omni pilot is prepared but not publicly reproducible until gated data access and held-out evaluation pass.
DATA_BLOCKER_REPORT.mdMinimal path: install the toolkit dependencies, download the official sample, run the 12-task suite with neural heads, regenerate visualizations, then run the artifact index and publication validator.
git clone https://github.com/Ropedia/HOMIE-toolkit.git
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r HOMIE-toolkit/requirements.txt huggingface_hub hf_xet
git clone https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite.git
pip install -r ropedia-xperience-10m-task-suite/requirements.txt
pip install torch
hf download ropedia-ai/xperience-10m-sample \
--repo-type dataset \
--local-dir data/sample/xperience-10m-sample
cd ropedia-xperience-10m-task-suite
export WORKSPACE=/path/to/workspace
python scripts/episode_task_suite.py --workspace "$WORKSPACE" --include-neural
python scripts/research_direction_extension_tasks.py
python scripts/task_walkthroughs.py
python scripts/generate_visualizations.py
python scripts/render_overview_figures.py
python scripts/render_task_suite_infographic.py
python scripts/export_modality_atlas_assets.py
python scripts/validate_website_integrity.py
python scripts/validate_scope_claims.py
python scripts/build_artifact_index.py
python scripts/validate_mirror_parity.py
python scripts/validate_publication_package.py