Differentiable programming for the autonomous movement planning of a small vessel

Abstract

In this work we explore the use of differentiable programming to allow autonomous movement planning of a small vessel. We aim for an end to end architecture where the machine learning algorithm directly controls engine power and rudder movements of a simulated vessel to reach a defined goal. Differentiable programming is a novel machine learning paradigm, that allows to define a systems parameterized response to control commands in imperative computer code and to use automatic differentiation and analysis of the information flow from the controlling inputs and parameters to the resulting trajectory to compute derivatives to be used as search directions in an iterative algorithm to optimize a goal function. Initially the method does not know about any manoeuvring or the vessels response to control commands. The method autonomously learns the vessels behaviour from several simulation runs. Finally, we will show how the simulated vessel is able to fulfil some small missions, like crossing a flowing river while avoiding crossing traffic.