Forecasting battery state of charge for robot missions

Abstract

Due to limited power onboard, a significant factor for success of distributed teams of robots is energy-awareness. The ability to predict when power will be depleted beyond a certain point is necessary for recharging or returning to a base station. This paper presents a framework for forecasting state of charge (SOC) of a robot's battery for a given mission. A generalized and customizable mission description is formulated as a sequence of parametrized tasks defined for the robot; the missions are then mapped to expected change in SOC by training neural networks on experimental data. We present results from experiments on the Turtlebot 2 to establish the efficacy of this framework. The performance of the proposed framework is demonstrated for three distinct mission representations and compared to an existing method in the literature. Finally, we discuss the strengths and weaknesses of feedforward and recurrent neural network models in the context of this work.