humanoidsdata.com

Search

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

← All glossary terms

Hardware & control

Observability

Observability is the property that a system's internal state can be determined from its measured outputs over time, given the known inputs and the assumed model. If different states can generate indistinguishable observations, those differences are unobservable under that sensing setup.

Also known as: state observability, system observability

Updated

A property of the model and measurements

Caltech's output-feedback notes introduce observability as the ability to recover state from measurement history. A state need not be measured directly to be observable: the dynamics can make it inferable from how other measured quantities change.

This is distinct from the quality of a particular state-estimation algorithm. Observability asks whether the information exists under the model. An estimator asks how to use it. A better filter cannot recover information absent from its measurements without adding assumptions, prior information or another source of reference.

For a linear time-invariant model, an observability-matrix rank test gives a precise condition. Nonlinear robotic systems require more care: conclusions can depend on the operating trajectory and on whether the analysis concerns the nonlinear model or a local linearisation. OpenVINS illustrates that distinction using a linearised visual-inertial error-state system.

A smooth trajectory can still have an arbitrary origin

In the visual-inertial setup analysed by OpenVINS, global translation and rotation about gravity are unobservable. A camera and IMU can estimate useful relative motion without independently determining an absolute world origin or global yaw. Those freedoms are not the same as a failed pose estimate; they describe which reference information the sensors do not supply.

Consequently, two valid reconstructions can use different coordinate frames. Before comparing trajectories, specify which frame alignment is allowed and whether external position, heading or map information was available. Do not interpret an arbitrary initial origin as a measured global location.

Overconfidence can conceal missing information

OpenVINS documents how inconsistent linearisation can introduce spurious information along an unobservable direction, producing overconfident uncertainty estimates. A small reported covariance is therefore not sufficient evidence that the underlying state is accurately known.

For humanoid or wearable datasets, document the sensor configuration, reference information, initialisation procedure and motion conditions. Separate absolute measurements from estimates expressed in a locally chosen frame. When an evaluation reports pose accuracy, its alignment convention and treatment of unobservable directions are part of the result, not optional implementation details.

Sources