PROJECT_README.md

Ropedia Xperience-10M Task Suite

An audit-first embodied-AI learning repo built around one public Xperience-10M sample episode released by Ropedia.

The public dashboard and generated figures deliberately follow the visual language of ropedia.com: near-black 4D-world canvas, lime-green identity accents, thin green-tinted cards, point-cloud texture, and the Inter Tight / Space Grotesk typography pairing. The layout is original to this project, but the style stays aligned with Ropedia's own product site.

The project does one narrow thing carefully: it turns a raw multimodal episode into:

• manifested sliding-window features over the currently extracted modalities,
• motion-only and current all-feature baseline models,
• 12 end-to-end episode-level tasks,
• lightweight neural MLP heads for the same 12 task contracts,
• a generated four-direction research taxonomy matching the Ropedia job tracks,
• four additional direction-extension probes with minimal and neural baselines,
• junior-friendly walkthroughs for every task, with case study, input, process, and output,
• a next TODO track for Qwen3-Omni fine-tuning and sensor-bridge evaluation,
• metrics, predictions, model weights, manifests, charts, and a static website,
• a clear explanation of what a single episode can and cannot prove.

Evidence Contract

This repo is organized around an explicit proof boundary:

Claim layer	Evidence	Boundary
Data windows	results/episode_task_suite/windows.csv, shared_windows.npz, summary_report.json	one public sample episode
Feature contract	results/episode_task_suite/feature_manifest.json, available_modalities.json	8,378 current features; audio documented but not featurized
12-task suite	scripts/episode_task_suite.py, per-task metrics.json, predictions	chronological single-episode split
Neural heads	scripts/neural_task_models.py, results/episode_task_suite/neural_mlp/	compact MLP heads, not a foundation model
Research directions	research_direction_taxonomy.json, extension probe results	direct/proxy/diagnostic evidence, not full solutions
Qwen3-Omni	results/omni_finetune/DATA_BLOCKER_REPORT.md, A100_HF_RELAY_STATUS.md	smoke-only until 32 valid episodes are available
Publication hygiene	scripts/validate_publication_package.py, docs/data/publication_audit.json	public repo and HF bundles only; ignored local scratch files are excluded
Citation and metadata	CITATION.cff, codemeta.json, docs/data/project_manifest.json, LICENSE	code is MIT-scoped; raw-data use follows Xperience-10M terms

Read the full contract in EVIDENCE_CONTRACT.md, or consume the machine-readable copy at docs/data/evidence_contract.json. The current publication audit is at docs/data/publication_audit.json. Project citation and machine-readable metadata live in CITATION.cff, codemeta.json, and docs/data/project_manifest.json.

Dataset Modality Coverage

The Xperience-10M sample is a 4D multimodal episode source spanning video, audio, depth, pose, motion capture, inertial sensing, and language annotation. This repo keeps that distinction explicit:

• the raw sample files include six MP4 video streams with AAC audio streams,
• annotation.hdf5 includes depth, SLAM/camera pose, hand/body mocap, IMU, and language annotation,
• the current minimal 8,378-d baseline feature manifest includes video, depth, pose/SLAM, mocap, IMU, calibration, and language blocks,
• audio is documented in the figures but is not yet extracted as a model input feature block in this minimal baseline.

Start with the visual dashboard:

chaoyue0307.github.io/ropedia-xperience-10m-task-suite

Hugging Face Space app:

cy0307-ropedia-xperience-10m-task-suite.static.hf.space

Read This Project In Three Layers

Layer	What to inspect	Why it matters
Data contract	windows.csv, feature_manifest.json, modality manifests	Confirms what each sample window contains before modeling
Minimal heads	softmax, ridge projection/regression, multi-label logistic heads	Keeps every input/output contract visible and debuggable
Neural heads	PyTorch MLP classifiers/regressors under neural_mlp/	Checks whether nonlinear heads improve each task without changing features
Evidence	metrics, predictions, confusion matrices, diagrams, dashboard	Makes the single-episode claims reviewable without rerunning first
Publication audit	docs/data/publication_audit.json	Confirms public bundles contain no raw Xperience-10M data, Python caches, heavy archives, or token strings
Citation metadata	CITATION.cff, codemeta.json, LICENSE	Makes the repo easier to cite, index, and reuse without confusing code license and dataset terms

Links

Resource	Link
This GitHub repo	github.com/ChaoYue0307/ropedia-xperience-10m-task-suite
This project website	chaoyue0307.github.io/ropedia-xperience-10m-task-suite
This Hugging Face Space	huggingface.co/spaces/cy0307/ropedia-xperience-10m-task-suite
Live Hugging Face static app	cy0307-ropedia-xperience-10m-task-suite.static.hf.space
Derived artifacts on Hugging Face	huggingface.co/datasets/cy0307/ropedia-xperience-10m-task-suite-artifacts
Minimal and neural task baselines on Hugging Face	huggingface.co/cy0307/ropedia-xperience-10m-task-baselines
Hugging Face collection	huggingface.co/collections/cy0307/ropedia-xperience-10m-task-suite
Xperience-10M dataset website	ropedia.com/dataset
Xperience-10M release page	ropedia.com/blog/20260316_xperience_10m
Ropedia GitHub organization	github.com/Ropedia
HOMIE Toolkit	github.com/Ropedia/HOMIE-toolkit
Xperience-10M Hugging Face dataset	huggingface.co/datasets/ropedia-ai/xperience-10m
Xperience-10M sample on Hugging Face	huggingface.co/datasets/ropedia-ai/xperience-10m-sample
Ropedia Hugging Face organization	huggingface.co/ropedia-ai

Citation, License, And Metadata

Use CITATION.cff when citing this project. The repository also includes codemeta.json for machine-readable software metadata and docs/data/project_manifest.json for website/Hugging Face surface metadata.

The code files are MIT-licensed. Raw Xperience-10M data is not redistributed here, and dataset use remains governed by the official Ropedia/Xperience-10M terms. See LICENSE and DATA_NOTICE.md.

The infographic uses a ChatGPT-image-generated text-free research background and larger modality-atlas thumbnails extracted from the public sample episode. The task names, input/output summaries, and metrics are overlaid from results/episode_task_suite/summary_report.json with scripts/render_task_suite_infographic.py, so the published PNG is a presentation graphic with verified labels and metrics, not a hallucinated metric sheet.

The pipeline and architecture figures use the same pattern: ChatGPT-image provides text-free visual backgrounds, while scripts/render_overview_figures.py overlays exact labels, dimensions, and metrics from the committed result files.

Scope

This is a learning, inspection, and pipeline-validation repo. It does not claim cross-episode generalization because the public sample used here is one episode. The correct next step for real model claims is to run the same suite over many episodes and split train/test by held-out episode.

What Is Inside

scripts/
  train_min_action_model.py         # motion/IMU baseline
  train_all_modalities_model.py     # current all-feature lightweight baseline
  episode_task_suite.py             # 12 end-to-end task definitions
  neural_task_models.py             # optional PyTorch MLP heads for all 12 tasks
  research_direction_taxonomy.py    # maps 12 tasks to the four research tracks
  research_direction_extension_tasks.py # one extra data-backed probe per track
  task_walkthroughs.py              # beginner explanations for each task contract
  generate_visualizations.py        # refreshes SVG charts + summary JSON
  render_task_suite_infographic.py  # renders the ChatGPT-image-backed PNG
  render_overview_figures.py        # renders polished pipeline/architecture PNGs
  validate_publication_package.py   # checks public repo + HF bundle hygiene
  omni/
    download_sample_modelscope.py   # mainland-China friendly sample download
    build_episode_manifest.py       # metadata-only multi-episode scanner
    plan_finetune_sample_budget.py  # H20 storage/sample-count planner
    qwen3_omni_adapter_smoke.py     # real-data Qwen3-Omni adapter smoke test

results/
  min_action_model/                 # motion-only action baseline artifacts
  min_subtask_model/                # motion-only subtask baseline artifacts
  min_all_modalities_action_model/  # current all-feature action artifacts
  min_all_modalities_subtask_model/ # current all-feature subtask artifacts
  episode_task_suite/               # 12-task suite metrics and predictions
    neural_mlp/                     # optional neural baseline artifacts per task
    research_directions/            # four-track taxonomy, CSV, and summary
    research_direction_extensions/  # four extra direction probes + predictions
    task_walkthroughs/              # case-study walkthroughs for all 12 tasks
  omni_exploration/                 # H20/ModelScope smoke-test artifacts

docs/
  index.html                        # GitHub Pages dashboard
  data/summary_metrics.json         # website-readable metrics bundle
  data/evidence_contract.json       # machine-readable proof boundary
  data/publication_audit.json       # machine-readable publication hygiene check
  data/project_manifest.json        # machine-readable public-surface metadata
  data/research_directions.json     # four-track website data bundle
  data/research_direction_extensions.json # four extra probe data bundle
  data/task_walkthroughs.json       # beginner task explanation data bundle
  assets/task_suite_infographic.png # 12-task presentation graphic
  assets/pipeline_diagram.png       # verified episode pipeline graphic
  assets/task_architectures.png     # verified 12-task minimal architecture map
  assets/charts/*.svg               # regenerated visualizations

notes/
  min_action_model.md
  all_modalities_model.md
  episode_task_suite.md

Raw Xperience-10M data is not committed. Download it from the official Ropedia distribution and follow the dataset terms.

Data Expected

The scripts expect a workspace with the Ropedia HOMIE toolkit and the Xperience-10M sample episode:

<workspace>/
  HOMIE-toolkit/
  data/sample/xperience-10m-sample/
    annotation.hdf5
    fisheye_cam0.mp4
    fisheye_cam1.mp4
    fisheye_cam2.mp4
    fisheye_cam3.mp4
    stereo_left.mp4
    stereo_right.mp4

The public sample dataset identifier is:

ropedia-ai/xperience-10m-sample

Hugging Face URL:

https://huggingface.co/datasets/ropedia-ai/xperience-10m-sample

Quickstart

From a workspace folder:

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

Download the sample:

hf download ropedia-ai/xperience-10m-sample \
  --repo-type dataset \
  --local-dir data/sample/xperience-10m-sample

On mainland-China servers, use ModelScope instead:

python scripts/omni/download_sample_modelscope.py \
  --output-dir data/sample/xperience-10m-sample \
  --mode minimal

--mode minimal downloads annotation.hdf5, README.md, and fisheye_cam0.mp4. Use --mode all-training to add all six MP4 streams while still skipping visualization.rrd.

Clone and run this repo:

git clone https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite.git
cd ropedia-xperience-10m-task-suite
python scripts/episode_task_suite.py --workspace /path/to/workspace

Run the same 12-task suite with lightweight neural heads:

pip install torch
python scripts/episode_task_suite.py \
  --workspace /path/to/workspace \
  --include-neural

Run the smaller baselines:

python scripts/train_min_action_model.py --workspace /path/to/workspace
python scripts/train_all_modalities_model.py --workspace /path/to/workspace

Xperience-10M Fine-Tuning Exploration On H20

This repo now includes a concrete first step toward a Qwen3-Omni fine-tuning pipeline over Xperience-10M. The important separation is:

• direct Qwen3-Omni inputs: RGB/fisheye video, embedded MP4 audio, and language prompts,
• adapter-required Xperience-10M sensor inputs: depth, pose/SLAM, hand/body mocap, contacts, and IMU.

The H20 work now has two separate evidence levels:

• an adapter-side smoke test over one Xperience-10M sample episode, useful for checking sensor feature extraction and label plumbing,
• a technical Qwen3-Omni LoRA smoke run that loaded the local Qwen/Qwen3-Omni-30B-A3B-Instruct weights and trained LoRA parameters on 128 windows from the single locally available episode.

Neither is a 32-episode result. The full pilot is still gated on raw Xperience-10M access and a held-out episode split.

python scripts/omni/build_episode_manifest.py \
  --data-root /home/cy/Ropedia/modelscope_data \
  --output outputs/omni_exploration/modelscope_manifest.json

python scripts/omni/qwen3_omni_adapter_smoke.py \
  --workspace /home/cy/Ropedia/ropedia-xperience-10m-task-suite \
  --episode-root /home/cy/Ropedia/modelscope_data/xperience-10m-sample \
  --target action \
  --window-frames 20 \
  --stride-frames 100 \
  --max-windows-per-episode 64 \
  --epochs 2 \
  --skip-video-features

Verified H20 run:

Item	Value
Server	8 x NVIDIA H20, 96GB each
Free storage checked	about 1.5TB under /home/cy
Data source	ModelScope ropedia-ai/xperience-10m-sample
Downloaded minimal data	1.93GB annotation.hdf5 + 85.7MB fisheye_cam0.mp4
Smoke windows	59
Split	single-episode chronological
Feature dim	4,262
Adapter soft-token blocks	11
Qwen3-Omni weights loaded	adapter smoke: no; LoRA smoke: yes
Result	0.0000 macro-F1, expected for this single-episode chronological smoke split

The zero score is not treated as a model claim. It is a useful signal that this split is not leaking labels across time: the train segment does not cover every action that appears in the held-out segment. The next real step is to add more episodes and split by held-out episode.

Sample Count Decision

The local Mac sample is only one episode. For H20 fine-tuning, decide sample count by storage and evaluation design, not by the local folder. The current H20 has about 1.5TB free under /home/cy; after reserving space for model weights, checkpoints, caches, and logs, a realistic first budget is:

Phase	Episodes/samples	Approx windows at stride 5	Purpose
Smoke	1-3	1k-3k	Verify loaders, token alignment, and task heads
Pilot	16-32	18k-37k	First held-out-episode evaluation
Useful LoRA run	64-128	74k-149k	Train sensor adapters plus selected Qwen3-Omni LoRA
Storage-heavy run	256+	297k+	Only after download layout and checkpoint size are stable

For the next run, use a 32-episode stratified pilot through the A100 relay, then scale to 128 episodes and later 512 episodes only after the download, transfer, manifest, train, and held-out evaluation path is stable. Do not treat "10M" as a reason to start with the entire dataset; the engineering unit that matters first is diverse held-out episodes, not adjacent windows from one session.

Use the budget helper before downloading:

python scripts/omni/plan_finetune_sample_budget.py \
  --storage-root /home/cy \
  --target-free-after-download-gb 800 \
  --all-training-per-episode-gb 2.4 \
  --full-preview-per-episode-gb 5.1

Refresh charts and the website data bundle:

python scripts/research_direction_taxonomy.py
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

32-Episode Readiness Gate

python scripts/omni/discover_xperience10m_sources.py \
  --workspace /home/cy/Ropedia/ropedia-xperience-10m-task-suite \
  --data-root /home/cy/Ropedia/modelscope_data \
  --output results/omni_finetune/source_discovery.json \
  --report-output results/omni_finetune/DATA_BLOCKER_REPORT.md

Current status in this repo:

• local_valid_episodes: 1 (degraded-valid: annotation + fisheye_cam0.mp4)
• local_complete_episodes: 0
• ready_for_32_episode_pilot: false
• A100 Hugging Face relay: active watcher, polling gated access every 15 minutes
• planned 32-episode pilot: stratified across 32 top-level session UUIDs
• HF full dataset blocker: ropedia-ai/xperience-10m returns 403 pending review
• source_discovery: results/omni_finetune/source_discovery.json
• blocker_report: results/omni_finetune/DATA_BLOCKER_REPORT.md
• relay_status: results/omni_finetune/A100_HF_RELAY_STATUS.md

Current H20-sourced evidence files in this repo:

• results/omni_finetune/episode_manifest.json
• results/omni_finetune/dataset_manifest.json
• results/omni_finetune/training_metadata.json
• results/omni_finetune/metrics.json
• results/omni_finetune/progress.jsonl
• results/omni_finetune/RUN_REPORT.md
• results/omni_finetune/DATA_BLOCKER_REPORT.md
• results/omni_finetune/A100_HF_RELAY_STATUS.md

Use this gate before scheduling any 32-episode full fine-tune run.

For the A100 Hugging Face relay, the 32-episode pilot should use stratified selection, not the first 32 paths in repository order. The current relay script scans 64 top-level session UUIDs, filters for complete leaf episodes, excludes visualization.rrd, applies a 0.25 GB minimum episode size, and selects 32 episodes from 32 different session UUIDs. This is still a pilot subset, but it is materially better for generalization checks than adjacent episodes from the same recording session.

Uploading the pilot Qwen3-Omni LoRA

A prepared upload package is available at results/omni_finetune/hf_upload.

python3 scripts/omni/upload_qwen3_omni_lora_to_hf.py \
  --repo-id cy0307/ropedia-qwen3-omni-lora-smoke \
  --source-dir results/omni_finetune/hf_upload \
  --message "Upload Xperience-10M Qwen3-Omni LoRA pilot"

This script requires a valid Hugging Face token via HF_TOKEN or --token. Network availability to huggingface.co is required.

Four Research Directions

The 12 tasks are now organized against the four Ropedia research directions in a generated artifact, not only in prose:

• research_direction_taxonomy.json
• research_direction_task_map.csv
• research_direction_summary.md
• docs/data/research_directions.json

The taxonomy uses two current baselines for every task:

Baseline	Role
Minimal interpretable heads	Softmax, logistic, ridge, and retrieval heads over the 8,378-d window feature vector. These expose the input/output contract cleanly.
Neural MLP heads	Small PyTorch MLP classifiers/regressors on the same features and splits. These check whether nonlinear heads help before moving to Qwen/Omni fine-tuning.

Current direction-level coverage:

Direction	Current status	Covered task evidence	What is not solved yet
A. Human Modeling & Motion Understanding	Partially implemented	hand_trajectory_forecast and contact_prediction are direct; timeline_action and object_relevance are proxies. Neural MLP improves hand forecasting from 0.8223 to 0.1116 MPJPE.	No full body/shape model, SMPL/MANO target, deformation prior, or multi-episode motion-generation evaluation yet.
B. 3D/4D Reconstruction & Neural Rendering	Proxy tasks only	cross_modal_retrieval, modality_reconstruction, and misalignment_detection test alignment/reconstruction prerequisites.	No NeRF, Gaussian Splatting, TSDF, mesh, novel-view synthesis, or calibrated 4D reconstruction model yet.
C. Egocentric Vision & Interaction	Strongest implemented track	6 direct tasks: action, subtask, transition, next-action, object relevance, and caption grounding, plus alignment/order diagnostics.	Single-episode chronological split limits generalization; audio and stronger video-language backbones still need to be added.
D. Scene Reconstruction & World Modeling	Early proxy tasks	Subtask/next-action, object relevance, retrieval, reconstruction, temporal order, and misalignment provide state/world-model probes.	No persistent scene graph, object permanence task, long-term map, or held-out-episode world model yet.

The important interpretation is that all four directions can be started from the Xperience-10M sample modalities, but only direction C is strongly represented by the current 12-task suite. Directions A, B, and D need additional targets and multi-episode training before they become full research deliverables.

Four Direction-Extension Probes

Beyond the original 12 core tasks, the repo now includes one extra data-backed probe for each research direction. These probes are computed from the same shared_windows.npz, windows.csv, and feature_manifest.json artifacts, so the reported numbers are real sample-derived metrics, not placeholder results.

• research_direction_extension_results.json
• research_direction_extension_summary.md
• docs/data/research_direction_extensions.json
• research_direction_extension_tasks.svg

Direction	New extension task	Input	Output	Minimal	Neural MLP	Why it matters
A. Human Modeling & Motion Understanding	body_motion_intensity	non-mocap video/depth/pose/IMU/SLAM/language features	high vs low body/hand motion	0.7827 macro-F1	0.7986 macro-F1	Starts a human-motion-energy target without leaking mocap input.
B. 3D/4D Reconstruction & Neural Rendering	multi_view_consistency_retrieval	fisheye camera feature query	synchronized stereo-left view rank	0.5534 MRR	0.3469 MRR	Tests whether multi-view features preserve synchronized 4D scene identity.
C. Egocentric Vision & Interaction	action_phase_progress	non-caption multimodal window	progress inside current action segment	0.3416 MAE	0.3038 MAE	Adds a task-structure/intent-style target beyond class labels.
D. Scene Reconstruction & World Modeling	ego_motion_forecast	current sensors excluding camera translation and captions	future camera-translation delta	0.1989 MAE	0.0989 MAE	Starts a short-horizon world-model target over wearer motion.

Run:

python scripts/research_direction_extension_tasks.py

These four probes make the four-direction mapping more concrete, but they are still single-episode extension baselines. Full research claims still require multi-episode training, held-out episode evaluation, and stronger task-specific models.

Task Walkthroughs For Juniors

Every task now has a beginner-facing explanation with:

• a concrete coffee-episode case study,
• exact input contract,
• middle process modules,
• output contract,
• minimal and neural metric,
• one important limitation.

Primary files:

• TASK_WALKTHROUGHS.md
• task_walkthroughs.json
• docs/data/task_walkthroughs.json

Compact map:

Task	Case study	Input -> process -> output
timeline_action	A pouring window should be named as the current action.	all-modality window -> action label builder + classifier -> action class
timeline_subtask	A fine action is grouped into a broader drink-preparation stage.	all-modality window -> subtask label builder + classifier -> subtask label
transition_detection	Detect the change from preparing to pouring.	window -> boundary builder + binary classifier -> boundary/steady
next_action	A preparing window predicts what happens 20 frames later.	current window -> future-label shift + classifier -> next action
hand_trajectory_forecast	A hand moving toward a cup becomes a future 3D hand path.	current window -> future mocap target + regressor -> hand trajectory
contact_prediction	Decide whether hand/body contact is happening.	non-contact features -> contact target + binary classifier -> contact label
object_relevance	Infer milk, cup, coffee, or related objects during pouring.	non-caption features -> multi-hot object target + sigmoid heads -> object set
caption_grounding	Query Pour milk into coffee and retrieve the matching moment.	text-like query + candidates -> projection + cosine ranker -> ranked windows
cross_modal_retrieval	Motion/IMU from pouring retrieves matching depth/video.	motion/IMU/camera -> projection + candidate index -> ranked depth/video windows
modality_reconstruction	Infer depth/video features from motion, IMU, and camera pose.	source modalities -> scaler + regressor -> target modality vector
temporal_order	Tell whether reaching then pouring was reversed.	adjacent window pair -> pair combiner + binary classifier -> correct/reversed
misalignment_detection	Catch motion paired with visual/depth features shifted in time.	motion side + visual side -> aligned/shifted pair builder + classifier -> aligned/shifted

Minimal 12-Task Architectures

These are deliberately minimal baselines. They are useful because every input/output contract is explicit, not because they are strong embodied-AI models.

Shared setup:

raw episode -> 20-frame windows, stride 5 -> 8,378-d current feature vector
chronological split: first 70% train, last 30% test
scalers are fit on train windows only

There are four reusable head families:

Head family	Used by	What it means
Linear softmax classifier	timeline_action, timeline_subtask, transition_detection, next_action, contact_prediction, temporal_order, misalignment_detection	z-score features, then XW+b, softmax, cross-entropy, L2
Dual ridge regression/projection	hand_trajectory_forecast, modality_reconstruction	z-score input/target, solve ridge regression with L2=10
Ridge + cosine ranking	caption_grounding, cross_modal_retrieval	project one modality into another feature space, then rank candidates by cosine
Multi-label logistic regression	object_relevance	z-score non-caption features, sigmoid object heads, threshold at 0.5

The optional neural run keeps the same feature vectors, leakage filters, chronological splits, and metrics, but replaces the task heads with small PyTorch MLP classifiers or regressors. Its outputs live under results/episode_task_suite/neural_mlp/, and the rollup is stored in the neural_tasks section of results/episode_task_suite/summary_report.json.

The task-specific heads are:

Task	Input	Minimal head	Output
timeline_action	all featurized modalities	linear softmax	current action class
timeline_subtask	all featurized modalities	linear softmax	current subtask class
transition_detection	all featurized modalities	linear softmax	steady vs action boundary
next_action	all featurized modalities at t	linear softmax	action at t+20 frames
hand_trajectory_forecast	all featurized modalities at t	ridge regression	future 10-frame left/right hand joints
contact_prediction	non-contact and non-caption feature blocks	linear softmax	any body contact
object_relevance	non-caption feature blocks	multi-label logistic	relevant object set
caption_grounding	sensor windows projected to text space	ridge projection + cosine ranking	matching time window for text query
cross_modal_retrieval	motion/IMU/camera projected to visual space	ridge projection + cosine ranking	matching depth/video window
modality_reconstruction	motion/IMU/camera	ridge regression	depth/video feature vector
temporal_order	[x_t, x_t+1, x_t+1-x_t]	binary linear softmax	correct vs reversed order
misalignment_detection	motion plus visual pair	binary linear softmax	aligned vs shifted by 8 windows

Key Results

Experiment	Main score	Accuracy	Notes
Motion-only action	0.9688 macro-F1	0.9828	Uses motion/IMU features only
Current all-feature action	0.9791 macro-F1	0.9828	8,378-dimensional feature vector
Motion-only subtask	0.9528 macro-F1	0.9759	Strong within-episode subtask signal
Current all-feature subtask	0.9308 macro-F1	0.9828	High accuracy, lower class-balanced score
Cross-modal retrieval	0.3764 top-5	n/a	Motion/IMU/camera retrieves matching depth/video
Transition detection	0.6552 macro-F1	0.9253	Boundary F1 is 0.2143
Hand trajectory forecast	0.8223 MPJPE	n/a	Predicts future hand-joint trajectory
Neural MLP hand forecast	0.1116 MPJPE	n/a	Same features/split, nonlinear regression head
Neural MLP temporal order	0.8718 F1	0.8707	Strong improvement on adjacent-window ordering
Neural MLP misalignment	0.7335 F1	0.7312	Detects shifted motion/visual pairs better than the linear head

Neural MLP Results

The neural baseline was run locally with --include-neural for all 12 tasks using 80 epochs, hidden size 128, batch size 128, and CPU execution. It is not a foundation model result; it is a controlled nonlinear-head comparison over the same 8,378-d handcrafted window features.

Task	Neural metric	Minimal metric	Readout
timeline_action	0.0263 macro-F1	0.0500 macro-F1	Still blocked by unseen future classes
timeline_subtask	0.0175 macro-F1	0.0495 macro-F1	Same single-episode split limitation
transition_detection	0.6485 macro-F1	0.6552 macro-F1	Similar to the linear baseline
next_action	0.0235 macro-F1	0.0593 macro-F1	Same unseen-label issue
hand_trajectory_forecast	0.1116 MPJPE	0.8223 MPJPE	Neural regression improves this target
contact_prediction	1.0000 macro-F1	1.0000 macro-F1	Degenerate one-class sample
object_relevance	0.1798 micro-F1	0.1839 micro-F1	Similar weak object signal
caption_grounding	0.0178 MRR	0.0172 MRR	Similar ranking behavior
cross_modal_retrieval	0.1530 MRR	0.2634 MRR	Linear ridge remains stronger here
modality_reconstruction	-0.0102 R2	-0.0160 R2	Small improvement but still weak
temporal_order	0.8718 F1	0.5487 F1	Neural head captures local temporal structure
misalignment_detection	0.7335 F1	0.4866 F1	Neural head improves alignment detection

The strongest single-episode self-supervised signal is cross-modal retrieval: motion/IMU/camera features retrieve matching depth/video windows substantially better than random.

Reproducibility Audit

I re-ran the full pipeline from the local raw public sample into /private/tmp/ropedia-audit and compared regenerated metrics with the committed artifacts. The baseline metrics, 12 task metrics, feature manifest, and available modality manifest matched exactly after float normalization.

See notes/reproducibility_audit.md for the commands and verification evidence.

Why Some Scores Are Low

The task suite intentionally uses a chronological split:

first 70% of the episode -> train
last 30% of the episode  -> test

The test segment contains some action/subtask labels never seen during training. Timeline and next-action classifiers therefore expose the core limitation of single-episode learning instead of hiding it behind random splits.

Feature Blocks Used

The current feature vector has 8,378 dimensions and includes:

• hand/body mocap joints and contact labels,
• camera translation and rotation,
• IMU acceleration and gyroscope traces,
• depth confidence features,
• six video streams,
• caption/object/interaction text features,
• SLAM point-cloud summary features,
• calibration parameters.

It does not yet include an audio feature block.

The exact feature block boundaries are stored in results/episode_task_suite/feature_manifest.json.

Data Notice

Xperience-10M data belongs to its original authors and is subject to the official Ropedia dataset license and access terms. This repo contains code and derived experiment artifacts only; it does not redistribute the raw videos or raw annotation dataset.