license: cc-by-nc-4.0
language:
- en
pretty_name: PULSE-sample
size_categories:
- 100M<n<1B
task_categories:
- time-series-forecasting
- video-classification
- other
tags:
- multi-modal
- daily-activity
- wearable-sensors
- motion-capture
- electromyography
- eye-tracking
- inertial-measurement-unit
- tactile-sensing
- sample
PULSE-sample — Representative Single-Recording Subset
This is a small representative subset of the PULSE dataset, hosted as a separate Hugging Face repository so that NeurIPS 2026 reviewers can inspect data quality and schema without downloading the full 85 GB release.
Full dataset:
velvet-pine-22/PULSE
What this sample contains (~284 MB total)
A single complete recording — v1/s1 (volunteer 1, scenario S1 "Office desk
organization", ~101 s) — with all five non-visual sensor modalities plus the
synchronized scene-camera video, action-segment annotations, and the global metadata
files needed to interpret everything.
PULSE-sample/
├── data/v1/s1/
│ ├── aligned_emg_100hz.csv # 8-channel surface EMG @ 100 Hz
│ ├── aligned_eyetrack_100hz.csv # 24-dim binocular gaze @ 100 Hz
│ ├── aligned_imu_100hz.csv # 160-dim wearable IMU @ 100 Hz
│ ├── aligned_mocap_100hz.csv # 56-joint optical motion capture @ 100 Hz
│ ├── aligned_pressure_100hz.csv # 50-channel fingertip pressure @ 100 Hz
│ ├── aligned_myo_pose_100hz.csv # forearm pose (auxiliary)
│ ├── aligned_myo_quat_100hz.csv # forearm orientation (auxiliary)
│ ├── alignment_metadata.json # per-recording sync diagnostics
│ ├── raw/ # raw Qualisys MoCap stream (.tsv)
│ └── videos/ # scene-cam + gaze-overlay (.mp4, 25 fps)
├── annotations/v1/s1.json # dense segment annotations (action / hand / object / text)
├── annotations_flat/segments.csv # the 30 segments of v1/s1, flattened
├── metadata/recordings.csv # full 337-row recording manifest
├── metadata/modality_coverage.xlsx # per-recording modality availability
├── LICENSE # CC BY-NC 4.0 (data)
└── CODE_LICENSE # MIT (code in companion repo)
How to use this sample
import pandas as pd
# Load all five modalities for the single recording
ROOT = "data/v1/s1"
emg = pd.read_csv(f"{ROOT}/aligned_emg_100hz.csv")
eyetrack = pd.read_csv(f"{ROOT}/aligned_eyetrack_100hz.csv")
imu = pd.read_csv(f"{ROOT}/aligned_imu_100hz.csv")
mocap = pd.read_csv(f"{ROOT}/aligned_mocap_100hz.csv")
pressure = pd.read_csv(f"{ROOT}/aligned_pressure_100hz.csv")
print(f"Aligned shapes (T, D): {[x.shape for x in [emg, eyetrack, imu, mocap, pressure]]}")
# Load the dense segment annotations
import json
with open("annotations/v1/s1.json") as f: ann = json.load(f)
print(f"{len(ann['segments'])} action segments")
All time series are sub-frame aligned (<10 ms) on a shared 100 Hz timebase. The first
sample of every modality file corresponds to t = 0 of the trimmed scene-cam video;
total length matches metadata/recordings.csv row v1s1 (duration_sec,
n_samples_100hz).
How this sample was created
Selected by the dataset authors as a representative recording: scenario S1 "office desk organization" was chosen because it contains a typical mix of grasp / move / place / adjust primitives without unusually short or long sub-tasks; v1 was chosen because it has all five modalities present and full-length scene-cam video.
The full 337-row metadata/recordings.csv is included so reviewers can see exactly
where this recording sits in the train/test split scheme and which other recordings
exist; the global Croissant metadata is on the main repo.
License & attribution
Data is released under CC BY-NC 4.0. By accessing PULSE-sample you agree to the
license, including the prohibition on commercial redeployment, re-identification, and
worker-surveillance applications. See LICENSE for the full terms. Companion code is
released under MIT (see CODE_LICENSE).
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}
}