robotics//navigation

Navigation, in robotics, is the problem of knowing where a vehicle is and where it is pointing, continuously and with a stated uncertainty, so that its controllers and planners can act on it. Every solution combines two kinds of information that fail in opposite ways. Integrating the vehicle's own motion is smooth, fast and available everywhere, and its error grows without bound; absolute fixes from outside are slow, noisy or intermittent, and do not drift.

Navigation, in robotics, is the problem of knowing where a vehicle is and where it is pointing, continuously and with a stated uncertainty, so that its controllers and planners can act on it. Every solution combines two kinds of information that fail in opposite ways. Integrating the vehicle's own motion is smooth, fast and available everywhere, and its error grows without bound; absolute fixes from outside are slow, noisy or intermittent, and do not drift.

The first kind is the family of dead reckoning, the integration of rates and accelerations from a known starting point, usually from an IMU; odometry does the same with measured displacements (wheel turns counted by a rotary encoder, successive lidar scans, camera features). The second kind brings an anchor: a GNSS receiver, a magnetometer for heading, a barometer for height, beacons, or a known map in which a particle filter places the robot.

Drift grows until an external reference anchors it.

Whatever integrates motion needs something absolute to correct it, and the art of navigation is choosing which references, how often, and what to do in the minutes when none is available.

Fusion is where the two meet. A Kalman-type filter predicts with the inertial sensors at hundreds of hertz and corrects with each slow fix when it arrives (sensor fusion); a simple attitude estimate does the same with a complementary filter.

When no absolute reference exists, the robot can build its own. Visual-inertial odometry couples a camera with an IMU and drifts slowly, and SLAM builds a map while localizing in it, recognising places already seen to pull the accumulated drift back.

What can be known is decided before the algorithm: with an IMU alone, heading about the vertical and absolute position leave no trace in the measurements (observability).