| # Datasheet for PULSE |
|
|
| Following the *Datasheets for Datasets* template (Gebru et al., 2021). |
| Refer to `croissant.json` for the machine-readable equivalent. |
|
|
| --- |
|
|
| ## Motivation |
|
|
| **For what purpose was the dataset created?** |
| To enable research on **long-horizon, multi-modal daily activity understanding** |
| where complementary sensor channels (whole-body motion, muscle activation, gaze, |
| limb kinematics, contact forces) interact during compositional tasks. Existing |
| datasets typically cover only one or two of these channels and focus on short |
| isolated actions; PULSE is designed to fill that gap with hardware-time- |
| synchronized 100 Hz recordings. |
|
|
| **Who created the dataset?** |
| Anonymous Authors (NeurIPS 2026 Evaluations & Datasets submission, double-blind). |
|
|
| **Who funded the creation of the dataset?** |
| Not disclosed at submission time. |
|
|
| --- |
|
|
| ## Composition |
|
|
| **What do the instances represent?** |
| Each instance is one **recording** = one execution of one of 8 daily-activity |
| scenarios by one of 40 volunteers. A recording contains up to 5 time-synchronized |
| sensor streams plus annotations. |
|
|
| **How many instances are there in total?** |
| - 304 task recordings (S1–S8), totaling ~7.0 hours |
| - 33 isolated motion-primitive recordings (S9), totaling ~2.9 hours |
| - 282 of the 304 task recordings (from 36 volunteers) carry dense action-segment annotations totaling **7,789 segments** |
|
|
| **Does the dataset contain all possible instances or is it a sample?** |
| Sample. 40 volunteers were recruited from a single university campus; not every |
| volunteer completed every scenario, and equipment availability / sensor dropout |
| caused some recordings to lack one or more modalities. This is documented in |
| `modality_coverage.xlsx` and `batch_alignment_summary.json`. |
|
|
| **What data does each instance consist of?** |
| Raw and aligned features at 100 Hz: |
| - **MoCap**: 422-dim raw → 620-dim processed (hip-relative + per-joint velocity) |
| - **EMG**: 8-channel surface EMG (band-pass 20–450 Hz, rectified) |
| - **EyeTrack**: 24-dim core numeric features; v1/s1 and v14/s8 have one-eye channels unavailable and are excluded from EyeTrack benchmarks. The raw eye-image video is **not** included in the benchmark, only the scene-cam |
| - **IMU**: 160-dim from 10 wearable units (accel + gyro + mag + orientation quaternion per unit) |
| - **Pressure**: 50-channel fingertip pressure (25 per hand) |
| - Auxiliary: scene-cam video (1280×720, 25 fps, MP4) |
|
|
| Per-segment annotation: start/end time (1 s resolution), motor primitive (taxonomy of 18), hand (left/right/both), object (scene-specific whitelist), description + 4 paraphrases. |
|
|
| **Is there a label or target associated with each instance?** |
| Yes — at two levels: (L1) one of 8 scene labels per recording; (L2) dense action segments per recording. |
|
|
| **Is any information missing from individual instances?** |
| Yes. EMG ≈99.3%, IMU ≈98.4%, EyeTrack ≈92.8%, MoCap ≈80.9%, Pressure ≈69.7%. All five simultaneously available for ~65.1% of task recordings (198/304). 22 recordings lack dense annotations and are released as raw aligned streams. |
|
|
| **Are there recommended data splits?** |
| Yes — two subject-independent splits: |
| - **Headline split** (T1 scene + T2 action): 31 train / 5 test (`v14, v30, v34, v38, v41`) |
| - **Full-modality split** (T4 missing-modality robustness): same train pool / 4 test (`v3, v25, v26, v27`) |
|
|
| **Are there errors, sources of noise, or redundancies in the dataset?** |
| Action-segment boundaries are at 1 s semantic resolution. Annotation produced by a VLM-assisted pipeline; ~15% independently re-annotated by a human expert (action-primitive κ=0.713, object κ=0.916, hand-label accuracy=0.863, boundary mean IoU=0.906; reported in the paper appendix). v14 lacks MoCap and pressure for all sessions. Two recordings (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. |
|
|
| **Is the dataset self-contained?** |
| Yes. No external links or third-party content are required to use the released data. |
|
|
| --- |
|
|
| ## Collection process |
|
|
| **How was the data acquired?** |
| In an instrumented laboratory furnished to resemble common real-world environments (office desk, kitchen counter, dining table, wardrobe area, packing station). Volunteers performed each scenario in their own style and at their own pace after a brief verbal description of the task. |
|
|
| **What mechanisms or procedures were used to collect the data?** |
| - Qualisys optical motion-capture system (master clock, hardware trigger) |
| - Myo Armband (8-channel surface EMG) |
| - Dikablis Glasses 3 (binocular eye tracking + scene-cam) |
| - 10 wearable IMUs |
| - Custom 50-channel fingertip pressure array |
|
|
| **Over what time-frame was the data collected?** |
| 2025–2026. |
|
|
| **Were the individuals from whom the data was collected notified about the data collection?** |
| Yes. Each participant received a written informed-consent document covering the study purpose, procedure, risks, and data-use terms before each session. |
|
|
| **Did the individuals consent to the collection and use of their data?** |
| Yes — written informed consent, plus a separate written media-release agreement explicitly authorizing release of synchronized scene-camera and eye-tracking videos. The released videos contain no audio track and no visible faces. |
|
|
| **If consent was obtained, were the consenting individuals provided with a mechanism to revoke their consent in the future?** |
| Yes — participants were informed they could pause or withdraw at any point; participants who withdrew received the full session compensation. Post-release withdrawal: corresponding author contact (to be disclosed in the camera-ready) handles takedown requests. |
|
|
| **Has an analysis of the potential impact of the dataset and its use on data subjects been conducted?** |
| Yes — see Section "Broader Impact" of the paper. The released videos contain no audio track and no visible faces, but first-person video remains privacy-sensitive auxiliary material. Biometric-attribute inference risks are addressed through the license and use restrictions. |
|
|
| --- |
|
|
| ## Preprocessing / cleaning / labeling |
|
|
| **Was any preprocessing/cleaning/labeling of the data done?** |
| Yes: |
| - Hardware synchronization at session start; residual drift corrected via inertial × MoCap cross-correlation (<10 ms) |
| - All streams resampled to 100 Hz (anti-aliased decimation) |
| - MoCap → hip-relative coordinates + appended per-joint velocity (422 → 620 dim) |
| - EMG band-pass filtered (20–450 Hz) and rectified |
| - EyeTrack kept as the full 24-dim numeric feature vector; v1/s1 and v14/s8 are excluded from EyeTrack-based benchmark training and evaluation |
| - For benchmark modeling: further downsampled to 20 Hz; per-channel z-score normalization using training-split statistics only |
| - Variable-length sequences zero-padded and masked at batch level |
| - Action-segment annotation via VLM-assisted pipeline with scene-constrained prompt; 15% independently re-annotated by a human expert |
|
|
| **Was the "raw" data saved in addition to the preprocessed/cleaned data?** |
| The released aligned data are at 100 Hz with the preprocessing above already applied. The original per-modality manufacturer outputs (Qualisys TSV, Myo SDK, Dikablis exports) are retained internally but not part of the public release. |
|
|
| --- |
|
|
| ## Uses |
|
|
| **Has the dataset been used for any tasks already?** |
| Yes — six benchmark tasks (T1–T6) defined in the accompanying paper, evaluated with a baseline suite covering three backbone architectures, nine fusion strategies, seven published baselines, SyncFuse, and DailyActFormer. These results should be read as baselines and diagnostics for open challenges rather than as a primary model contribution. |
|
|
| **Is there a repository that links to any or all papers or systems that use the dataset?** |
| Initially the HuggingFace dataset page; a full citation list will be maintained post-release. |
|
|
| **What (other) tasks could the dataset be used for?** |
| Long-horizon multi-modal action segmentation, sub-task anticipation, hierarchical imitation learning, sensorimotor coordination analysis, myoelectric prosthetic control, fingertip-pressure-conditioned dexterous manipulation, ergonomic/cognitive workplace studies. |
|
|
| **Is there anything about the composition of the dataset or the way it was collected and preprocessed/cleaned/labeled that might impact future uses?** |
| - Single-site cohort limits demographic generalization |
| - 25 fps scene-cam cannot recover the sub-frame anticipatory coordination signals visible at 100 Hz — vision-only models will miss the dataset's distinguishing temporal signal |
| - 1 s annotation boundary resolution |
| - Downstream users of biometric-class modalities for sensitive-attribute inference must obtain their own ethics review |
|
|
| **Are there tasks for which the dataset should not be used?** |
| Yes — covert worker surveillance, employee monitoring, biometric re-identification, or any application without freely given and revocable consent of the new subjects. See `LICENSE` for the binding restrictions. |
|
|
| --- |
|
|
| ## Distribution |
|
|
| **Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created?** |
| Yes — public release on HuggingFace Datasets (URL pending; anonymized mirror used for review). |
|
|
| **How will the dataset be distributed?** |
| HuggingFace Datasets repository with Git LFS, plus a Croissant 1.0 metadata file. |
|
|
| **When will the dataset be distributed?** |
| Anonymous review release at submission time; permanent public release upon acceptance. |
|
|
| **Will the dataset be distributed under a copyright or other intellectual property (IP) license, and/or under applicable terms of use (ToU)?** |
| **CC BY-NC 4.0** (`LICENSE`) with additional use restrictions covering re-identification and surveillance applications. Code is released under MIT (`CODE_LICENSE`). |
|
|
| **Have any third parties imposed IP-based or other restrictions on the data associated with the instances?** |
| No. |
|
|
| **Do any export controls or other regulatory restrictions apply to the dataset or to individual instances?** |
| None known. |
|
|
| --- |
|
|
| ## Maintenance |
|
|
| **Who will be supporting/hosting/maintaining the dataset?** |
| The authors (institution to be disclosed in the camera-ready). HuggingFace Datasets provides the hosting infrastructure. |
|
|
| **How can the owner/curator/manager of the dataset be contacted?** |
| Corresponding author email — to be disclosed in the camera-ready. During review, contact via the OpenReview submission system. |
|
|
| **Is there an erratum?** |
| Will be maintained as a section of the HuggingFace dataset page after release. |
|
|
| **Will the dataset be updated?** |
| Bug-fix patches will be issued as new minor versions with a changelog. Major schema changes will be issued as new major versions. |
|
|
| **If the dataset relates to people, are there applicable limits on the retention of the data associated with the instances?** |
| No fixed retention limit beyond what the participant media-release agreement allows. Withdrawal requests handled by the corresponding author. |
|
|
| **Will older versions of the dataset continue to be supported/hosted/maintained?** |
| Yes — older versions remain available as Git LFS history for reproducibility. |
|
|
| **If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so?** |
| Yes — pull requests / discussions on the HuggingFace Datasets page after release. |
|
|
