Reinforcement Learning Approach for Co-Management of Computational and Mechanical Power Consumption in Mobile Robots

Abstract

Optimizing energy consumption remains a critical challenge in autonomous mobile robotics and is essential for extending robot battery life. Two significant sources of energy consumption are the mechanical and computational components, both of which contribute to battery usage. Recent studies indicate that dynamically coordinating mechanical and computational power usage, such as tuning processor frequency based on the robot's mechanical speed, can greatly enhance energy efficiency across diverse mobile robot platforms. This improvement is primarily due to the relationship between decision-making processes based on mechanical speed and computational workload. In this thesis, we begin by presenting a simulation model that estimates the instantaneous power consumption of a mobile robot by taking into account both its mechanical and computational components while also analyzing how its computational aspects impact path planning. The simulation model is adaptable to be tuned based on the level of accuracy needed for estimating the power consumption for the robot and the simulation time penalty. This makes a multi-fidelity power estimation tool for the robot with the capability to run-time changing the fidelity according to environmental conditions and internal computational capabilities. Subsequently, we propose an agile reinforcement learning algorithm for dynamic co-management. Using a rover capable of manipulating the speed of its brushless motor, adjusting the frequency of the Jetson TX2 processing unit, and utilizing an event-based camera as the perceptual sensor, we demonstrate that our method effectively addresses the scalability and accuracy issues found in previously proposed techniques. As a result, we achieve significantly greater efficiency in overall energy reduction and mission duration. Our experimental results show energy efficiency improvements ranging from a minimum of 16.98% to a maximum of 60.86% compared to the most efficient approaches documented in the literature.