humanoidsdata.com

Search

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

← All glossary terms

Hardware & control

Sensor fusion

Sensor fusion is the process of combining measurements from multiple sensors or information sources to estimate a robot's state or its environment. A fusion system must account for what each measurement observes, its coordinate frame, timing, uncertainty, and failure modes; merely placing sensor channels in the same record is not sensor fusion.

Also known as: multisensor fusion, multi-sensor fusion

Updated

Fusion estimates a common quantity

A humanoid may combine encoders, an IMU, cameras, force sensors and external tracking. Those devices do not automatically measure the same thing. Encoders describe joint configuration, an IMU measures inertial motion, and a camera constrains motion through observed scene structure. Fusion uses a model to relate such measurements to a common state or environmental quantity.

OpenVINS derives a measurement update by combining a prior state distribution with a noisy measurement model. The update depends on measurement uncertainty and on how the measurement is expected to change with the state. This is more specific than concatenating several inputs for a neural network, which may be multimodal learning without an explicit fused estimate.

Frames and timestamps are part of the measurement

Measurements must refer to compatible times and coordinate frames. A camera pose delayed by one frame can disagree with an IMU measurement even when both sensors are accurate. An incorrect sensor-to-body transform can appear as persistent motion or bias.

The ROS robot_localization documentation makes these assumptions visible through timestamped inputs, coordinate-frame parameters, covariances and selectable state variables. A dataset intended for fusion should retain timestamps, clocks, frame identifiers, calibration transforms, units and uncertainty information rather than only the final fused trajectory.

More sensors do not guarantee a better estimate

Two sensors can share a failure mode, provide correlated evidence or observe the same limited directions. Treating correlated measurements as independent can make an estimator overconfident. A sensor can also degrade under motion blur, magnetic interference, impacts, occlusion or contact.

Evaluation should therefore include ablations and failure cases, not only aggregate accuracy with every sensor enabled. Report which inputs were available, how they were synchronised, whether covariance was measured or tuned, and how missing or rejected measurements were handled. The fused output is an estimate; preserve the raw observations needed to audit it.

Sources