| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform", |
| "prov": "http://www.w3.org/ns/prov#" |
| }, |
| "@type": "sc:Dataset", |
| "name": "PULSE", |
| "description": "PULSE is a multi-modal dataset of daily activity recordings collected from 40 volunteers across 8 ecologically valid scenarios in an instrumented laboratory. It provides five hardware-time-synchronized modalities sampled at 100 Hz: full-body optical motion capture with finger-level hand articulation (56 joints), surface EMG (Myo Armband, 8 channels), binocular eye tracking (Dikablis Glasses 3, 24-dim numeric features), wearable IMU (10 units, 160 dim), and a 50-channel fingertip pressure array recording quantitative grip force. The release contains 304 task recordings (~7.0 hours) plus 33 isolated motion-primitive recordings (S9, ~2.9 hours). Of the 304 task recordings, 282 (from 36 volunteers) carry dense action-segment annotations totalling 7,789 segments tagged with motor primitive (taxonomy of 18), hand, manipulated object, and natural-language descriptions with four paraphrased variants. First-person scene-camera video (1280x720, 25 fps, from the Dikablis Glasses 3) and eye-tracking video are released alongside the five sensor streams as auxiliary material; released videos contain no audio track and no visible faces. Six benchmark tasks are defined: scene recognition, fine-grained action recognition, grasp onset anticipation, missing-modality robustness, tactile-driven grasp state recognition, and cross-modal pressure prediction.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "license": "https://creativecommons.org/licenses/by-nc/4.0/", |
| "url": "https://huggingface.co/datasets/velvet-pine-22/PULSE", |
| "version": "1.0.0", |
| "keywords": [ |
| "multi-modal", |
| "human activity recognition", |
| "motion capture", |
| "EMG", |
| "eye tracking", |
| "IMU", |
| "fingertip pressure", |
| "wearable sensing", |
| "long-horizon compositional tasks", |
| "action segmentation", |
| "human-robot interaction", |
| "myoelectric control" |
| ], |
| "creator": { |
| "@type": "Organization", |
| "name": "Anonymous Authors (NeurIPS 2026 Evaluations & Datasets submission, double-blind review)" |
| }, |
| "publisher": { |
| "@type": "Organization", |
| "name": "Anonymous (to be filled in camera-ready)" |
| }, |
| "citeAs": "@inproceedings{anonymous2026pulse, title={PULSE: A Synchronized Five-Modality Dataset for Multi-Modal Daily Activity Understanding}, author={Anonymous Authors}, booktitle={Submitted to NeurIPS 2026 Evaluations and Datasets Track}, year={2026} }", |
| "rai:dataCollection": "Forty healthy adult volunteers were recruited from a single university campus. Each volunteer performed a predefined set of daily activity scenarios inside an instrumented laboratory room furnished to resemble common real-world environments (office desk, kitchen counter, dining table, wardrobe area, packing station). Recordings were captured under hardware time synchronization (Qualisys MoCap controller as master trigger; sub-frame, <10 ms residual drift corrected via cross-correlation in post-processing). Each session lasted approximately 2-3 hours. Data were collected between 2025 and 2026 (see per-recording timestamps in alignment_metadata.json).", |
| "rai:dataCollectionType": "Lab-based instrumented recording with wearable + optical sensors. No web-scraped or third-party content.", |
| "rai:dataCollectionTimeframe": "2025-2026", |
| "rai:dataCollectionRawData": "Five hardware-synchronized sensor streams (Qualisys MoCap, Myo Armband EMG, Dikablis Glasses 3 eye tracking, 10 wearable IMUs, 50-channel fingertip pressure array) plus a first-person scene-camera video (1280x720, 25 fps).", |
| "rai:dataPreprocessingProtocol": "All five sensor streams resampled to a common 100 Hz timebase and aligned via Qualisys hardware trigger. MoCap converted to hip-relative coordinates with appended per-joint velocity features (422 -> 620 dimensions). EMG band-pass filtered (20-450 Hz) and rectified. Eye tracking is retained as the full 24-dimensional numeric feature vector; v1/s1 and v14/s8 have one eye's pupil-size and event channels unavailable and are excluded from EyeTrack-based benchmark training and evaluation. For benchmark modeling all modalities further downsampled to 20 Hz with anti-aliased decimation; per-channel z-score normalization uses training-split statistics only. Variable-length sequences zero-padded and masked at batch level. Action segments produced by a VLM-assisted annotation pipeline with a scene-constrained prompt enforcing the object vocabulary and motor-primitive taxonomy; ~15% subset independently re-annotated by a human expert (agreement reported in paper appendix).", |
| "rai:dataAnnotationProtocol": "Two-level annotation: (L1) one of 8 high-level scenario labels per recording; (L2) dense action segments (mean ~2.5 s) with 1 s start/end resolution, each carrying a motor primitive (taxonomy of 18), a hand label (left/right/both), a manipulated object from a scene-specific whitelist (50+ objects with Chinese-English mapping), and a natural-language description with four paraphrased variants for language-grounded use.", |
| "rai:dataAnnotationPlatform": "Custom annotation pipeline using a vision-language model on the first-person scene-camera video, with one human expert independently re-annotating a 15% sample.", |
| "rai:dataAnnotationAnalysis": "Human-VLM agreement is reported in the paper appendix: action-primitive kappa=0.713, object kappa=0.916, hand-label accuracy=0.863, boundary mean IoU=0.906, and segment-count ratio=1.151.", |
| "rai:dataAnnotationDemographics": "Annotators were members of the research team; demographic profile not collected.", |
| "rai:dataReleaseMaintenancePlan": "Hosted on HuggingFace Datasets at https://huggingface.co/datasets/velvet-pine-22/PULSE. Versioned releases via Git LFS; bug-fix patches issued as new minor versions with changelog. The release after acceptance will include a permanent DOI and a contact email for issues.", |
| "rai:personalSensitiveInformation": "The five benchmark sensor modalities (MoCap, EMG, EyeTrack numeric features, IMU, Pressure) do not contain direct personal identifiers. The auxiliary first-person scene-camera and eye-tracking videos contain no audio track and no visible faces, but they remain privacy-sensitive because they can reveal task context, hand appearance, and manipulation style. All participants signed a separate media-release agreement explicitly acknowledging that synchronized videos and biomechanical signals would be released as research data. Biometric-class modalities (EMG, eye tracking, pressure) can in principle be used to infer sensitive attributes such as fatigue, stress, or medical conditions; downstream deployment for such inference on new subjects requires its own ethics review.", |
| "rai:dataBiases": "Cohort is drawn from a single university campus and is not demographically representative of the general population: age, body size, handedness, and cultural manipulation-style diversity are limited. EMG is recorded only on the dominant forearm (single Myo Armband). Modality availability is non-uniform across task recordings: EMG ~99.3%, IMU ~98.4%, eye tracking ~92.8%, MoCap ~80.9%, pressure ~69.7%; all five simultaneously available for ~65.1% (198/304). Recordings vary in length (0.3-4.3 min, mean 1.4 min) and not all volunteers completed every scenario.", |
| "rai:dataLimitations": "(1) Single-site, single-cohort recruitment limits demographic generalization. (2) Realistic but non-uniform missing-modality pattern in 35% of task recordings. (3) Scene-camera at 25 fps cannot resolve the sub-frame (~20 ms) anticipatory coordination signals visible in the 100 Hz sensor streams, so vision-only models will miss the dataset's distinguishing temporal signal. (4) Action segment boundaries are at 1 s resolution. (5) 22 of the 304 task recordings lack dense annotations and are released as raw aligned streams.", |
| "rai:dataUseCases": "Intended for non-commercial research on: long-horizon multi-modal action segmentation, sub-task anticipation, missing-modality robustness, sensorimotor coordination analysis (EMG-motion timing), myoelectric prosthetic control, fingertip-pressure-conditioned dexterous manipulation, hierarchical imitation learning, and procedure-level evaluation. Six benchmark tasks (T1-T6) defined in the accompanying paper. NOT intended for: covert worker surveillance, employee monitoring, biometric re-identification, or any application without freely given and revocable consent of the new subjects.", |
| "rai:dataSocialImpact": "Positive: accelerates research on natural myoelectric prosthetic control, dexterous robot imitation learning, cognitive-ergonomic workplace studies, and assistive technology. Negative: fine-grained behavioral data could be misused for covert surveillance or biometric re-identification; the license explicitly disallows such uses, and downstream use for sensitive-attribute inference on new subjects requires fresh ethics review.", |
| "rai:dataSensitiveProtectedAttributes": "Released volunteer-level metadata is restricted to: anonymous volunteer ID (v1...v40), handedness (where applicable). Sex, age, height, weight were collected for screening per the consent form but are NOT included in the public release. The first-person scene-camera video may incidentally reveal physical appearance (skin tone, body morphology, dress).", |
| "rai:hasSyntheticData": false, |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "huggingface-repo", |
| "name": "huggingface-repo", |
| "description": "PULSE HuggingFace Datasets repository.", |
| "contentUrl": "https://huggingface.co/datasets/velvet-pine-22/PULSE", |
| "encodingFormat": "git+https", |
| "sha256": "TODO-after-data-upload-complete" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "modality-coverage-xlsx", |
| "name": "modality_coverage.xlsx", |
| "description": "Per-recording modality availability matrix across all 304 task recordings.", |
| "contentUrl": "metadata/modality_coverage.xlsx", |
| "encodingFormat": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet", |
| "sha256": "27af1ab8a8a5e9ce50c6f6d73544646ac55f308dbbd8a9178e5dbec956c5522e" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "batch-alignment-summary-json", |
| "name": "batch_alignment_summary.json", |
| "description": "Per-recording alignment metadata (detected modalities, common start/end timestamps, sample count, warnings).", |
| "contentUrl": "metadata/batch_alignment_summary.json", |
| "encodingFormat": "application/json", |
| "sha256": "485eec6f5c018c644441792d381c5fe6c1c2511ddf95b52d820666e174f43820" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "aligned-emg-files", |
| "name": "aligned-emg-files", |
| "description": "100 Hz aligned surface EMG (8 channels, Myo Armband on dominant forearm), one CSV per (volunteer, scenario) pair.", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "text/csv", |
| "includes": "data/v*/s*/aligned_emg_100hz.csv" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "aligned-imu-files", |
| "name": "aligned-imu-files", |
| "description": "100 Hz aligned IMU (10 units, 160 dim: triaxial acc/gyro/mag + orientation quaternions per unit).", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "text/csv", |
| "includes": "data/v*/s*/aligned_imu_100hz.csv" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "aligned-mocap-files", |
| "name": "aligned-mocap-files", |
| "description": "100 Hz aligned full-body optical motion capture (Qualisys, 56 joints, 422 raw / 620 processed dim).", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "text/csv", |
| "includes": "data/v*/s*/aligned_mocap_100hz.csv" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "aligned-eyetrack-files", |
| "name": "aligned-eyetrack-files", |
| "description": "100 Hz aligned binocular eye-tracking numeric features (Dikablis Glasses 3, 24 dim core; v1/s1 and v14/s8 excluded from EyeTrack benchmarks because one-eye channels are unavailable).", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "text/csv", |
| "includes": "data/v*/s*/aligned_eyetrack_100hz.csv" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "aligned-pressure-files", |
| "name": "aligned-pressure-files", |
| "description": "100 Hz aligned fingertip pressure (50 channels: 25 per hand).", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "text/csv", |
| "includes": "data/v*/s*/aligned_pressure_100hz.csv" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "scene-cam-video-files", |
| "name": "scene-cam-video-files", |
| "description": "First-person scene-camera video from Dikablis Glasses 3 (1280x720, 25 fps). Released videos contain no audio track and no visible faces. Released as privacy-sensitive auxiliary material; not part of the T1-T6 benchmark evaluation protocol.", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "video/mp4", |
| "includes": "data/v*/s*/videos/trimmed_*Scene Cam.mp4" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "annotation-files", |
| "name": "annotation-files", |
| "description": "Action-segment annotations (L1 scene label + L2 dense segments with primitive, hand, object, NL description with 4 paraphrases).", |
| "containedIn": { |
| "@id": "huggingface-repo" |
| }, |
| "encodingFormat": "application/json", |
| "includes": "annotations/v*/s*.json" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "recordings-csv", |
| "name": "recordings.csv", |
| "description": "Per-recording metadata table (337 rows): one row per recording with split labels and modality availability flags.", |
| "contentUrl": "https://huggingface.co/datasets/velvet-pine-22/PULSE/resolve/main/metadata/recordings.csv", |
| "encodingFormat": "text/csv", |
| "sha256": "96bc9df514f2da32c1bed9e293d25ecbe54300a83fef01655193a92ae6443aab" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "segments-csv", |
| "name": "segments.csv", |
| "description": "Flat CSV of all 7,789 dense action segments across 282 annotated recordings.", |
| "contentUrl": "https://huggingface.co/datasets/velvet-pine-22/PULSE/resolve/main/annotations_flat/segments.csv", |
| "encodingFormat": "text/csv", |
| "sha256": "61e49a28baf2fa5ab914b3ebc902d9cfe585ed770663789fe1e759ad49c5903c" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "recording", |
| "name": "recording", |
| "description": "One execution of one scenario by one volunteer, with all available aligned modalities.", |
| "key": { |
| "@id": "recording/recording_id" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "recording/recording_id", |
| "name": "recording_id", |
| "description": "Identifier of the form v{volunteer_id}s{scenario_id} (e.g., v1s1).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "recording_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/volunteer_id", |
| "name": "volunteer_id", |
| "description": "Anonymized volunteer identifier (v1..v40).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "volunteer_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/scenario_id", |
| "name": "scenario_id", |
| "description": "Scenario label: s1=office desk organization, s2=package shipping preparation, s3=kitchen spice organization, s4=post-meal table cleaning, s5=pre-meal table setting, s6=business travel luggage packing, s7=coffee/beverage preparation, s8=clothes hanging and folding, s9=isolated motion-primitive protocol.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "scenario_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/duration_sec", |
| "name": "duration_sec", |
| "description": "Recording duration in seconds (0.3-4.3 min for task scenarios; mean 1.4 min).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "duration_sec" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/n_samples_100hz", |
| "name": "n_samples_100hz", |
| "description": "Number of aligned samples at 100 Hz.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "n_samples_100hz" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/aligned_modalities", |
| "name": "aligned_modalities", |
| "description": "List of modalities successfully aligned for this recording (subset of {emg, imu, mocap, eyetrack, pressure}).", |
| "dataType": "sc:Text", |
| "repeated": true, |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "aligned_modalities" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/missing_modalities", |
| "name": "missing_modalities", |
| "description": "List of modalities not available for this recording.", |
| "dataType": "sc:Text", |
| "repeated": true, |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "missing_modalities" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/has_dense_annotations", |
| "name": "has_dense_annotations", |
| "description": "True if the recording carries dense L2 action-segment annotations (282 of 304 task recordings).", |
| "dataType": "sc:Boolean", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "has_dense_annotations" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "recording/split", |
| "name": "split", |
| "description": "Subject-independent split membership for the headline benchmark: 'train' (31 volunteers, 242 recordings) or 'test' (5 held-out volunteers v14, v30, v34, v38, v41; 40 recordings).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "recordings-csv" |
| }, |
| "extract": { |
| "column": "split" |
| } |
| } |
| } |
| ] |
| }, |
| { |
| "@type": "cr:RecordSet", |
| "@id": "action-segment", |
| "name": "action-segment", |
| "description": "One densely annotated action segment (mean ~2.5 s) within a recording. 7,789 segments across 282 annotated recordings.", |
| "key": { |
| "@id": "action-segment/segment_id" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/segment_id", |
| "name": "segment_id", |
| "description": "Globally unique segment identifier.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "segment_idx" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/recording_id", |
| "name": "recording_id", |
| "description": "Foreign key into recording.recording_id.", |
| "dataType": "sc:Text", |
| "references": { |
| "field": { |
| "@id": "recording/recording_id" |
| } |
| }, |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "recording_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/start_sec", |
| "name": "start_sec", |
| "description": "Segment start time in seconds (1 s resolution).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "start_sec" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/end_sec", |
| "name": "end_sec", |
| "description": "Segment end time in seconds (1 s resolution).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "end_sec" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/primitive", |
| "name": "primitive", |
| "description": "Motor primitive label drawn from a taxonomy of 18 (e.g., grasp, move, rotate, open, close, place, push, pull, ...).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "primitive" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/hand", |
| "name": "hand", |
| "description": "Hand involved: 'left', 'right', or 'both'.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "hand" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/object", |
| "name": "object", |
| "description": "Manipulated object from a scene-specific whitelist (50+ objects with Chinese-English mapping).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "object" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/description", |
| "name": "description", |
| "description": "Natural-language description of the segment.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "description" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "action-segment/paraphrases", |
| "name": "paraphrases", |
| "description": "Four paraphrased variants of the description for language-grounded downstream use.", |
| "dataType": "sc:Text", |
| "repeated": true, |
| "source": { |
| "fileObject": { |
| "@id": "segments-csv" |
| }, |
| "extract": { |
| "column": "description_variants" |
| } |
| } |
| } |
| ] |
| } |
| ], |
| "prov:wasDerivedFrom": [ |
| { |
| "@id": "https://huggingface.co/datasets/velvet-pine-22/PULSE", |
| "prov:label": "PULSE (primary collection)", |
| "sc:license": "CC BY-NC 4.0" |
| } |
| ], |
| "prov:wasGeneratedBy": [ |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q4929239" |
| }, |
| "prov:label": "Multi-modal sensor recording", |
| "sc:description": "Lab-based primary recording with 40 healthy adult volunteers performing 8 daily-activity scenarios in an instrumented laboratory. Five hardware-time-synchronized wearable sensor modalities were recorded at 100 Hz: Qualisys optical motion capture (acting as master trigger), Myo Armband surface EMG (8 channels), Dikablis Glasses 3 binocular eye tracking, 10 wearable IMUs, and a 50-channel fingertip pressure array. A first-person scene-camera video (1280x720, 25 fps) was recorded alongside; released videos have the audio track removed and contain no visible faces. Sub-frame residual drift (<10 ms) was corrected via cross-correlation in post-processing. Sessions lasted approximately 2-3 hours per volunteer; total release covers ~9.7 hours across 304 task recordings plus 33 motion-primitive recordings." |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q5227332" |
| }, |
| "prov:label": "Sensor stream alignment and feature extraction", |
| "sc:description": "All five sensor streams were resampled to a common 100 Hz timebase and aligned via the Qualisys hardware trigger. MoCap was converted to hip-relative coordinates with appended per-joint velocity features (422 to 620 dimensions). Surface EMG was band-pass filtered (20-450 Hz) and rectified. Eye tracking was reduced to a 14-dimensional subset that is common across all volunteers (one earlier participant lacks left-eye pupil-size and event channels). For benchmark modeling, all modalities were further downsampled to 20 Hz with anti-aliased decimation; per-channel z-score normalization uses training-split statistics only. Variable-length sequences were zero-padded with batch-level masking. The released video audio track was stripped to remove personally identifiable speech; volunteer faces are not visible in the first-person scene-camera footage." |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q109719325" |
| }, |
| "prov:label": "VLM-assisted action segmentation with human audit", |
| "sc:description": "Two-level annotation pipeline. (L1) One of 8 high-level scenario labels per recording, assigned during data collection. (L2) Dense action segments produced by a vision-language-model assisted pipeline applied to the first-person scene-camera video, using a scene-constrained prompt that enforces a fixed object vocabulary (~50 objects with Chinese-English mapping) and an 18-class motor-primitive taxonomy. Each segment carries a motor primitive, a hand label (left/right/both), a manipulated object, and a natural-language description with four paraphrased variants. A 15% audit subset was independently re-annotated by two human experts; inter-annotator agreement (VLM vs. human, human vs. human) is reported in the paper appendix. 22 of the 304 task recordings lack dense annotations and are released as raw aligned streams. No crowdworkers or external annotation platforms were used." |
| } |
| ] |
| } |
