CoWear

CoWear is an aligned multi-device wearable sensing dataset for studying motion, localization, and cross-device sensor fusion. It synchronizes on-device measurements from a mobile phone, smartwatch, and Rokid glasses with 6-DoF ground-truth trajectories. This release contains 363 aligned sessions, packaged as 15 downloadable tar.zst shards.

Participant-session mapping

See PARTICIPANT_ANNOTATION.md and participant_sessions.csv for the real collection participant-to-session mapping of the final released data.

Data preview

The following complete session is shown as a representative example. These plots are generated directly from the aligned files included in the release.

Aligned measure waveforms

Acceleration and angular-velocity magnitudes from mobile, watch, and Rokid IMU streams. The horizontal axis uses each record's final canonical alignedRelativeS value. This field is already materialized after manual alignment; users must not apply any additional watch/rokid/truth shift or perform a second synchronization pass.

Ground-truth trajectories

The horizontal X-Z position trajectories for all five retained ground-truth roles, alongside the four components of a representative Rokid orientation quaternion. Ground truth includes left/right foot trajectories.

Released data

Only aligned outputs derived from data/processed are included:

Source	Released roles	Contents
groundtruth/align.csv	mobile, watch, rokid, left, right when available	position in metres and orientation quaternion on the aligned timeline
measure/align/mobile/	mobile	IMU, BLE scans, markers, and metadata
measure/align/watch/	watch	direct IMU when available; relayed IMU only when direct IMU is absent; markers and metadata
measure/align/rokid/	rokid	IMU, markers, and metadata
alignment_info.json	session	alignment shift, anchor, and output-window metadata

Raw data and measure-side left/right streams are excluded. Device availability varies by session; use measure_roles in the manifest before selecting samples.

For watch IMU, prefer measure/align/watch/imu.jsonl when present. measure/align/watch/imu_relay.jsonl is published only for sessions without direct watch IMU.

Time and alignment contract

The files in this release are the final aligned outputs. In every IMU JSONL record, alignedRelativeS is the canonical session time in seconds, rebased to the released session window and already containing all reviewed manual alignment corrections. alignedTimestampMs is the corresponding absolute timestamp. In groundtruth/align.csv, time_s and timestamp_ms are on the same released canonical clock. Read these fields directly for synchronization. Do not read raw-data alignment plans, do not reapply values from alignment_info.json, and do not shift or cross-correlate the released streams again.

Archive layout

Each archive uses the following structure:

<date>/<session_id>/
|-- alignment_info.json
|-- extrinsic_calibration.json
|-- groundtruth/
|   `-- align.csv
`-- measure/
    `-- align/
        |-- mobile/
        |-- watch/
        `-- rokid/

Archive members start at <date>/<session_id>/. Extract a shard with:

tar --use-compress-program=unzstd -xf aligned-0606-001.tar.zst

Index and checksums

The Dataset Viewer uses manifest.csv as a session-level catalog. The aligned time-series files are stored in the downloadable tar.zst shards. See shards.csv for sizes and SHA-256 checksums.

Preview figures use session_20260706_151520, a session with complete mobile, watch, and Rokid measure coverage and all five ground-truth roles.