humanoidsdata.com

Search

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

← All glossary terms

Hardware & control

State estimation

State estimation is the process of inferring variables that are not known exactly—such as a robot’s base pose, velocity, contact state or sensor bias—from noisy measurements, control inputs and a model of how the system evolves. An estimator should provide both an estimate and, where possible, a representation of its uncertainty.

Also known as: robot state estimation, robotic state estimation

Updated

Sensors observe the state only indirectly

Probabilistic Robotics frames state estimation as recovering variables from partial, noisy sensor data. An encoder can measure a joint angle, but it does not directly provide a walking humanoid’s global position. An IMU measures angular velocity and specific force, not an error-free world-frame pose.

An estimator combines those observations with a process or motion model. Bayesian filters represent a belief over possible states; Kalman-family filters approximate that belief with a mean and covariance under particular assumptions. Optimisation and factor-graph methods can estimate a window or history of states instead of updating only the current one.

Humanoid estimation depends on changing contact

Leg kinematics can constrain base motion while a foot is stationary, but the constraint becomes wrong when the foot slips or lifts. Rotella and colleagues incorporate flat-foot contact constraints into a humanoid estimator without assuming a fixed gait.

Pronto combines high-rate IMU and leg odometry with lower-rate visual or lidar corrections on legged robots. This illustrates a common split: proprioceptive estimates support the fast control loop, while exteroceptive sensors reduce longer-term drift.

Estimated state is not ground truth

Every estimate depends on sensor calibration, timestamps, model assumptions, contact detection and noise settings. A smooth output can still be biased or overconfident. Covariance reports modelled uncertainty, not a guarantee that unmodelled errors are absent.

Datasets should keep raw measurements separate from filtered state and record the estimator name, version, parameters, reference frame, update rate and covariance. Simulator ground truth, motion-capture measurements and online estimates should use distinct field names so evaluation does not accidentally train on unavailable or circular targets.

Sources