Researchers led by Mingxi Zhou have trained small neural networks to predict an autonomous underwater vehicle's velocity, so it can keep track of its position when its Doppler velocity log cannot see the seabed, according to a paper on arXiv. Positions estimated from inertial data alone drift strongly, and the cost of the sensors limits swarm operations.

The networks were trained on field data from an AUV and tested on three new missions recorded on two different days. Localisation with the predicted velocities gave lower average errors than dead reckoning aided by the Doppler log. The size of the gain depended on ocean currents.

Source: arXiv Robotics (cs.RO)

Background

Mingxi Zhou

Robotics researcher

Robotics researcher.