Short answer
Body-worn IMUs add inertial motion data (acceleration and angular velocity) to workplace task video, giving head, torso or limb movement that cameras miss when views are occluded. They need defined placement, a stated sample rate, camera-IMU calibration and a shared clock. They do not give forces or robot actions, and drift and soft-tissue motion limit absolute position accuracy.
When an IMU channel earns its place
Video alone struggles with whole-body tasks: a worker bends behind a shelf, a cart blocks the camera, or a ladder climb leaves the frame. A body-worn inertial measurement unit keeps recording orientation and acceleration through those moments. It is also small and does not need line of sight.
Choose this channel when your question concerns body movement: stepping, turning, lifting posture, head motion during manipulation or the timing of whole-body transitions. Skip it when your question concerns hands and objects on a bench; there, camera quality, hand visibility and annotation matter more. For tasks that mostly happen at arm's length, see head-mounted camera capture.
What published collections show
Project Aria glasses, used for egocentric research, carry two IMUs, a magnetometer and a barometer, along with RGB, SLAM and eye-tracking cameras and a seven-microphone array (Engel et al., 2023). So a head-worn device can supply an inertial stream without extra equipment, but the stream tracks head motion rather than hands or legs.
Nymeria pairs full-body motion ground truth with Aria egocentric recordings and a third-person observer device. It covers 300 hours of daily activities in 1,200 recordings by 264 participants across 50 locations, with hierarchical language descriptions of motion (Ma et al., 2024). It shows how a body-worn motion suit, a head-worn recorder and an observer device can be combined at scale, though in daily life rather than industrial work.
Illustrative collection specification
All values below are examples the buyer sets for their own task. They are synthetic.
# Illustrative only - synthetic values, not an available dataset
imu_channel:
placements:
- {id: imu_head, location: "head-mounted device, rigid with scene camera"}
- {id: imu_chest, location: "sternum strap"}
- {id: imu_wrist_L, location: "left wrist band"}
- {id: imu_wrist_R, location: "right wrist band"}
sample_rate_hz: "buyer sets; record at least twice the fastest motion band of interest"
fields: [t_device, accel_xyz, gyro_xyz, mag_xyz (optional)]
units: {accel: "m/s^2", gyro: "rad/s"}
clock: "hardware-synchronized with scene camera; offset logged per session"
calibration:
camera_imu: "spatial and temporal, before and after each session"
static_bias_check: "30 s still at start and end"
orientation_frame: "defined and documented per sensor"
delivered: [raw_streams, calibration_report, per_session_sync_log]
metadata: [wearer_id (pseudonymous), sensor_serials, firmware, site_id, task_id]
Parameters that depend on your use
| Parameter | Decision | Consequence |
|---|---|---|
| Placement count | Head only, or head plus torso and wrists | More placements add setup time and wearer burden |
| Sample rate | Slow posture tracking versus fast hand motion | Higher rates raise storage and battery use |
| Axes and ranges | Accelerometer and gyro range | Ranges set too low clip fast motions |
| Magnetometer | Include or exclude | Metal structures distort it in plants |
| Synchronization | Shared hardware clock or post-hoc alignment | Software alignment error limits any frame-level labeling |
| Ground truth | Optical or suit-based body pose | Needed to validate IMU-derived pose |
| Session length | Short sessions or full shifts | Longer sessions increase drift and battery risk |
Calibration and synchronization
Cameras and IMUs must share a geometry and a clock. Kalibr is an open toolbox for multi-camera calibration and for spatial and temporal camera-IMU calibration, with outputs including a PDF report and YAML files (ETH Zurich ASL). Whatever tool a collector uses, ask for the same artifacts: intrinsics, extrinsics, the time offset between sensors and the residual error.
Plan for these failure modes: a strap that slips between calibration and task, a wearer who adjusts the device mid-session, magnetic disturbance near machines, and clock drift that grows over a long shift. Check each in the delivered logs rather than assuming it away.
Acceptance tests
- Completeness. Every session has all declared streams for its full length, with dropout counts below a buyer-set limit.
- Timing. Sync offset between camera and IMU is logged and under a buyer-set limit; check with a dedicated sync gesture such as a sharp tap or jump recorded at the start of each session.
- Bias and noise. Static periods show gyro bias and noise inside the sensor's stated range.
- Placement. Photos or video confirm strap location for sampled sessions.
- Plausibility. Integrated orientation matches camera-observed head motion on a random sample of sessions.
- Labeling alignment. Annotated events, such as step starts, fall at consistent offsets from IMU peaks.
Wire these into in-collection quality control so that bad sessions are caught on site.
What the channel cannot support
- Robot actions. IMU streams from a person are not joint commands for a humanoid.
- Contact forces. Acceleration is not force at the hand unless the rig is instrumented for it.
- Precise absolute position. Integrating acceleration drifts; pose estimates need fusion with cameras or an external reference.
- Fine finger motion. Hand-worn IMUs track wrist or segment orientation, not individual finger joints; for hands see 3D hand pose data.
- Embodiment transfer. Human limb timing and strength differ from a humanoid's; see the embodiment gap.
When IMU is combined with video-based estimation, full-body pose from workplace video covers the complementary camera-only route.
Pairing with other channels
IMU streams are most useful alongside other signals. With a multi-camera rig, they provide a motion reference when views are blocked; see multi-view synchronized task recording. With head-mounted gaze, they separate head motion from eye motion; see eye-gaze in demonstrations. For load-carrying tasks, wrist and torso sensors help describe posture and gait changes under load; see manual load handling and carrying. In each case, decide in advance which stream is the reference for timing, so disagreements have a rule.
Consent and workplace considerations
Wearable sensors record more than movement: pace, rest periods and gait can reveal health or performance. Consent should name every stream, retention period and permitted use, and should separate training use from any monitoring of the individual. Pseudonymous IDs reduce exposure but do not remove the need for permission. This is not legal advice; ask counsel before recording. See worker consent and identifiable people.
Questions to ask suppliers
- Which IMU models will be used, with what sample rate, range and noise specification?
- Where will sensors be placed, and how is placement verified in each session?
- How are camera and IMU clocks synchronized, and what offset is reported?
- What calibration is performed before and after each session, and which tool produces the report?
- How are drift, dropout and magnetic disturbance detected and flagged?
- Is any optical or suit-based ground truth recorded, and for what share of sessions?
- What do wearers do between tasks, and does the sensor set interfere with normal work or PPE?
- Are raw streams delivered with timestamps, or only filtered orientation?
- Does the consent cover inertial data specifically, including retention and secondary use?
Describe your needs
If your collection plan includes body-worn motion channels, describe the tasks, sensors and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, and can scope a commissioned collection of new recordings if no existing asset fits; a request does not guarantee a match. For wider navigation, see the workplace task demonstration hub and the humanoid training data guide. The catalog entry on sensor and IoT data gives related context.
Describe the workplace demonstration data your robotics project needs
Sources
- Project Aria: A New Tool for Egocentric Multi-Modal AI Research — Aria sensor set including IMUs.
- Nymeria: A Massive Collection of Egocentric Multi-modal Human Motion in the Wild — full-body motion with egocentric recordings at scale.
- Kalibr — camera and camera-IMU calibration toolbox.