Datasets:

velvet-pine-22
/

PULSE

license: cc-by-nc-4.0
language:
  - en
pretty_name: PULSE
size_categories:
  - 10M<n<100M
task_categories:
  - time-series-forecasting
  - video-classification
  - other
tags:
  - multi-modal
  - human-activity-recognition
  - motion-capture
  - emg
  - eye-tracking
  - imu
  - fingertip-pressure
  - wearable-sensing
  - long-horizon
  - action-segmentation
  - human-robot-interaction
  - myoelectric
configs:
  - config_name: emg
    data_files:
      - split: train
        path: data/v[0-9]*/s[0-9]*/aligned_emg_100hz.csv
  - config_name: imu
    data_files:
      - split: train
        path: data/v[0-9]*/s[0-9]*/aligned_imu_100hz.csv
  - config_name: mocap
    data_files:
      - split: train
        path: data/v[0-9]*/s[0-9]*/aligned_mocap_100hz.csv
  - config_name: eyetrack
    data_files:
      - split: train
        path: data/v[0-9]*/s[0-9]*/aligned_eyetrack_100hz.csv
  - config_name: pressure
    data_files:
      - split: train
        path: data/v[0-9]*/s[0-9]*/aligned_pressure_100hz.csv
  - config_name: action_segments
    data_files:
      - split: train
        path: annotations_flat/segments.csv
extra_gated_prompt: |
  By accessing PULSE you agree to the CC BY-NC 4.0 license and the
  additional use restrictions in LICENSE: no commercial redeployment, no
  attempts at re-identification of participants, and no use for covert worker
  surveillance or biometric identification without freely given consent.

PULSE

A Synchronized Five-Modality Dataset for Multi-Modal Daily Activity Understanding

MoCap · EMG · Eye Tracking · IMU · Fingertip Pressure — all hardware-synced at 100 Hz

NeurIPS 2026 Evaluations & Datasets · under double-blind review · CC BY-NC 4.0

At a glance

40	9	5	7,789
volunteers	scenarios (S1–S8 + S9 motion primitives)	modalities @ 100 Hz	dense action segments
337 total recordings (304 task + 33 S9)	~9.7 h total (S1–S8: ~7.0 h)	<10 ms cross-modal drift	6 benchmark tasks
282 annotated	36 annotated volunteers	17 motor primitives observed	57 unique objects

Status — anonymous review release. Author names, institution, and the permanent dataset URL will be revealed in the camera-ready version. Submission-time artifacts are available: companion code at velvet-pine-22/PULSE-code, a small representative sample at velvet-pine-22/PULSE-sample, and Croissant 1.0 metadata with RAI fields in croissant.json.

Tip. The Dataset Viewer at the top of this page is fully interactive — switch between the emg / imu / mocap / eyetrack / pressure / action_segments subsets via the dropdown, and hover any column to see its per-value distribution. The action_segments subset is the flat per-segment table (7,789 rows): hover the primitive, hand, or object columns to browse the label distributions interactively. The static figures further down summarise cross-tab statistics (per-modality coverage, per-volunteer matrix) the column-wise viewer cannot compute.

Quick start

Load any single modality directly through datasets:

from datasets import load_dataset

emg = load_dataset("velvet-pine-22/PULSE", "emg", split="train")
# Available configs: emg, imu, mocap, eyetrack, pressure, action_segments.
# Each per-modality CSV has a consistent schema across volunteers.
# (volunteer, scenario) is recovered from the source file path; the
# per-modality CSVs themselves do not carry id columns.

# Flat action-segments table (also browsable in the Dataset Viewer above):
seg = load_dataset("velvet-pine-22/PULSE", "action_segments", split="train")
# 7,789 rows × 15 columns: volunteer, scenario, primitive, hand, object, ...

Inspect one recording on disk:

data/v1/s1/
├── aligned_emg_100hz.csv
├── aligned_eyetrack_100hz.csv
├── aligned_imu_100hz.csv
├── aligned_mocap_100hz.csv
├── aligned_pressure_100hz.csv
├── aligned_myo_pose_100hz.csv     # auxiliary
├── aligned_myo_quat_100hz.csv     # auxiliary
├── alignment_metadata.json
├── raw/                           # opt-in: original Qualisys TSV (~40 MB)
│   └── aligned_v1s1_s_Q.tsv
└── videos/                        # opt-in: scene cam + eye-tracking video (~200 MB; no audio or visible faces)
    ├── trimmed_v1s1_*Scene Cam.mp4
    └── trimmed_v1s1_*Eye Tracking Video.mp4

Want only the benchmark streams? Skip the heavy bits at download:

hf download velvet-pine-22/PULSE --repo-type dataset \
    --exclude 'data/v*/s*/raw/**' \
    --exclude 'data/v*/s*/videos/**'
# ~10 GB instead of ~86 GB

Dataset statistics

Modality availability

Realistic missing-modality pattern across the 304 S1–S8 task recordings: EMG (99.3%), IMU (98.4%), EyeTrack (92.8%), and Scene Cam (93.1%) are present on >90% of task recordings; MoCap (80.9%) and Pressure (69.7%) are the two limiting modalities. All five sensors are simultaneously available on 198 / 304 task recordings (65.1%) — drives the T4 missing-modality benchmark.

Recordings and durations

Scenarios are well-balanced (33–40 recordings each). Recording length is right-skewed with a long tail of sessions up to ~9 minutes; median 77 s, mean 104 s.

Recordings per scenario and recording-duration histogram

Per-volunteer modality coverage and label distributions

The matrix below shows the missing-modality structure is systematic, not random: early volunteers (v1–v22) have intermittent MoCap and Pressure coverage; from v23 onwards almost all five modalities are present on every recording. Bimanual segments dominate (45%), and the manipulated-object distribution has a heavy head (top 8 of 57 objects account for ~50% of all segments).

Per-volunteer modality matrix; segments by hand; top manipulated objects

Motor-primitive distribution

7,789 dense segments across 17 motor primitives (taxonomy of 18; stabilize is the rarest at 0.3%, roll does not appear in the released split). The top three primitives — grasp / move / place — together account for 63% of all segments, reflecting the manipulation-heavy nature of daily activities.

Distribution of motor primitives across 7,789 dense action segments

The five modalities

Modality	Sensor	Channels / dim	Notes
MoCap	Qualisys optical	56 joints incl. all 10 fingertips · 422 raw / 620 processed dim	Hardware-trigger master clock
EMG	Myo Armband	8 channels, 20–450 Hz BP	Dominant forearm
EyeTrack	Dikablis Glasses 3	24-dim	Binocular numeric features; v1/s1 and v14/s8 are excluded from EyeTrack benchmarks because one eye's channels are unavailable
IMU	10 wearable units	160 dim total (acc / gyro / mag / quat per unit)	Body-distributed
Pressure	Fingertip pressure array	50 channels (25/hand)	Quantitative grip force

Auxiliary — first-person scene-camera video (1280×720, 25 fps). Released videos contain no audio track and no visible faces. They remain privacy-sensitive auxiliary material and are not part of the T1–T6 benchmark protocol.

Synchronization — Qualisys hardware trigger as master; sub-frame (<10 ms) cross-modal residual drift, corrected by post-hoc cross-correlation.

Annotations

Two-level scheme produced by a VLM-assisted pipeline with a 15% subset independently re-annotated by a human expert. The validation subset achieves action-primitive κ=0.713, object κ=0.916, hand-label accuracy=0.863, and boundary mean IoU=0.906 (reported in the paper appendix).

L1 — scenario label (one of 8 + S9 motion-primitive protocol) per recording.
L2 — dense action segments (mean 2.5 s, 1 s semantic boundary resolution). Each segment carries:
- motor primitive (taxonomy of 18; 17 observed in this release)
- hand (left / right / both)
- manipulated object (scene-specific whitelist; 57 distinct objects observed)
- natural-language description with 4 paraphrased variants for language-grounded use

Annotations are released in two complementary forms:

annotations/v*/s*.json — one nested JSON per recording (canonical structure, referenced by croissant.json).
annotations_flat/segments.csv — flat table with one row per segment (used by the Dataset Viewer action_segments subset; 7,789 rows × 15 columns).

Summary tables at metadata/annotations/.

Six benchmark tasks

	Task	Headline modality use
T1	Scene recognition (8-way)	All five
T2	Fine-grained action recognition (motor primitive × object × hand)	All five
T3	Grasp onset anticipation	EMG + IMU lead
T4	Missing-modality robustness	Trained on all, evaluated under modality dropout
T5	Tactile-driven grasp state recognition	Pressure-centered sensorimotor windows
T6	Cross-modal pressure prediction	EMG / hand MoCap → pressure

DailyActFormer is provided as a strong T2 baseline: it has the highest mean headline score, but margins over the strongest published baselines are small and confidence intervals overlap. T3 and T6 are intended as open challenge tasks rather than solved benchmarks. Full results, ablations, and the grasp-phase timing analysis are in the accompanying NeurIPS submission (linked here after double-blind review).

Train / test splits (subject-independent)

Split	Train	Test held out	Notes
Headline (T1, T2)	31 volunteers · 242 recordings · 6,582 segments	`v14, v30, v34, v38, v41` · 40 recordings · 1,207 segments	`v14` contributes only EMG, EyeTrack, IMU
Full-modality (T4)	same training pool	`v3, v25, v26, v27`	Restricted to recordings with full MoCap+EMG+EyeTrack+IMU

40 volunteers, IDs v1–v41 with v7 not assigned. S9 motion-primitive recordings are present for a subset of volunteers and have no MoCap or scene-camera video.

Repository layout

PULSE/
├── README.md            ← this file
├── DATASHEET.md         ← Gebru-style data card
├── LICENSE              ← CC BY-NC 4.0 + use restrictions
├── CODE_LICENSE         ← MIT, applies to (separately released) helper code
├── croissant.json       ← Croissant 1.0 machine-readable card with RAI fields
├── .gitattributes       ← LFS rules (do not delete before re-upload)
├── assets/              ← static figures used by this README
├── annotations/
│   └── v*/s*.json                    ← per-recording nested JSON (canonical)
├── annotations_flat/
│   └── segments.csv                  ← flat per-segment table (Dataset Viewer subset `action_segments`)
├── data/
│   └── v*/s*/
│       ├── aligned_<modality>_100hz.csv  (5 core + 2 Myo aux)
│       ├── alignment_metadata.json
│       ├── raw/      → opt-in raw Qualisys TSV (~40 MB / file, ~12 GB total)
│       └── videos/   → opt-in scene-cam + eye-tracking MP4 (~200 MB / pair)
├── metadata/
│   ├── modality_coverage.xlsx
│   ├── batch_alignment_summary.json
│   └── annotations/
│       ├── _run_summary.json
│       ├── segment_counts.{csv,xlsx}
│       └── logs/
└── docs/
    └── upload_to_huggingface.md

Ethics, privacy, and intended use

IRB / ethics committee approval obtained from the authors' institution (institution name withheld for double-blind review).
Written informed consent + separate media-release agreement for the scene-cam and eye-tracking videos; released videos contain no audio track and no visible faces.
The release does not include participant demographic attributes (sex, age, height, weight) collected for screening.
Use restrictions: see LICENSE. Surveillance, employee monitoring, and biometric re-identification applications without freely given consent are explicitly disallowed.
Downstream use of biometric-class modalities (EMG, eye tracking, pressure) for inference of sensitive attributes (fatigue, stress, medical conditions) on new subjects requires the deploying institution's own ethics review.

Citation

@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},
  note      = {Under double-blind review}
}