humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Data & collection

Visual-inertial odometry

Visual-inertial odometry estimates a moving sensor rig's motion by combining camera observations with inertial measurements from accelerometers and gyroscopes. Its core output is a relative motion or pose estimate over time, not an independently measured ground-truth trajectory or necessarily a reusable global map.

Also known as: VIO, visual inertial odometry

Updated

Images and inertial samples constrain the same motion

VIO combines visual evidence across images with the rig's inertial measurements. OpenVINS provides a concrete filter-based implementation: an extended Kalman filter fuses inertial information with sparse visual feature tracks through a sliding-window formulation. That is one implementation, not a definition requiring every VIO system to use a Kalman filter.

The relationship between sensors matters as much as their individual specifications. Camera intrinsics, the camera-to-IMU transform and their timing relationship are explicit calibration quantities in OpenVINS. Sensor calibration and data synchronisation should therefore be treated as part of the recording pipeline, not inferred from the mere presence of cameras and an IMU.

Odometry emphasises the rig's motion; SLAM jointly estimates pose and a map that can support later localisation. The ORB-SLAM3 paper distinguishes these goals while noting that odometry systems with loop-closing extensions make the boundary less sharp. A product labelled VIO may include additional mapping features, so inspect its actual outputs rather than deciding from the acronym alone.

Visual-inertial sensing also does not supply every global reference. OpenVINS's observability analysis identifies global translation and yaw about gravity as unobservable in its analysed setup. Choosing an initial origin produces coordinates; it does not turn those coordinates into externally referenced truth.

A headset trajectory is not a hand-action label

For egocentric robot-learning data, VIO can describe how the recording rig moved while a person performed a task. It does not by itself measure finger joints, object pose, contact forces or the robot command that would reproduce the action. Those require separate measurements or inference, with their own uncertainty.

Keep raw camera and IMU data distinct from the derived trajectory. Record the estimator and calibration versions, frame conventions, output timestamps, tracking failures and resets. If a later mapping stage revises the trajectory, label the corrected result separately from the online estimate available during capture. This prevents training or evaluation from silently treating future-corrected poses as causal sensor observations.

Sources