Contact-aided state estimation

Contacts provide temporary references#

An inertial measurement unit measures angular velocity and specific force at high rate, but integrating those signals accumulates drift. When a robot believes that a foot is planted without slipping, joint encoders and the kinematic model relate that contact point to the moving body. An estimator can use this relation as a correction.

The contact is not a permanent world landmark. It is added when a foot or hand becomes load-bearing and removed when contact breaks. Bloesch and colleagues formulated an extended Kalman filter that includes foothold positions in its state and fuses leg kinematics with IMU measurements.

Use on humanoids and legged robots#

A humanoid controller needs body orientation, velocity, and position estimates to balance and place the next foot. Contact-aided estimation can operate without cameras in dark, textureless, or visually occluded settings. It can also supply a short-term motion estimate alongside vision or lidar in a wider sensor fusion system.

Hartley and colleagues derived a contact-aided invariant extended Kalman filter for a biped. Their model combines inertial propagation, forward-kinematic corrections, changing contacts, and IMU-bias estimation. Their analysis also shows an important boundary: absolute position and rotation about gravity are not observable from IMU and point contacts alone.

This is a source-specific form of state estimation, not a synonym for it. Visual-inertial odometry uses visual motion constraints. Contact-aided estimation instead gets its external correction from assumed stationary contacts, although one system can combine both.

A slipping foot is a bad landmark#

The stationary-contact assumption fails when a foot slides, rolls, sinks into a soft surface, or touches a moving object. A false contact decision can then turn an incorrect kinematic constraint into a confident state correction. Contact detection, slip rejection, and uncertainty tuning are therefore part of the estimator design.

Encoder offsets, link flexibility, impact vibration, IMU bias, and an inaccurate kinematic model also affect the result. Contacts constrain drift but do not create an absolute global reference. Long operation still benefits from vision, lidar, beacons, or another source that observes global position and heading.