Deep Q-Learning-Based Dynamic Management of a Robotic Cluster

Abstract

The ever-increasing demands for autonomy and precision have led to the development of heavily computational multi-robot system (MRS). However, numerous missions exclude the use of robotic cloud. Another solution is to use the robotic cluster to locally distribute the computational load. This complex distribution requires adaptability to come up with a dynamic and uncertain environment. Classical approaches are too limited to solve this problem, but recent advances in reinforcement learning and deep learning offer new opportunities. In this paper we propose a new Deep Q-Network (DQN) based approaches where the MRS learns to distribute tasks directly from experience. Since the problem complexity leads to a curse of dimensionality, we use two specific methods, a new branching architecture, called Branching Dueling Q-Network (BDQ), and our own optimized multi-agent solution and we compare them with classical Market-based approaches as well as with non-distributed and purely local solutions. Our study shows the relevancy of learning-based methods for task mapping and also highlight the BDQ architecture capacity to solve high dimensional state space problems. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —A lot of applications in industry like area exploration and monitoring can be efficiently delegated to a group of small-size robots or autonomous vehicles with advantages like reliability and cost in respect of single-robot solutions. But autonomy requires high and increasing compute-intensive tasks such as computer-vision. On the other hand small robots have energy constraints, limited embedded computing capacities and usually restricted and/or unreliable communications that limit the use of cloud resources. An alternative solution to cope with this problem consists in sharing the computing resources of the group of robots. Previous work was a proof of concept limited to the parallelisation of a single specific task. In this paper we formalize a general method that allows the group of robots to learn on the field how to efficiently distribute tasks in order to optimize the execution time of a mission under energy constraint. We demonstrate the relevancy of our solution over market-based and non-distributed approaches by means of intensive simulations. This successful study is a necessary first step towards distribution and parallelisation of computation tasks over a robotic cluster. The next steps, not tested yet, will address hardware in the loop simulation and finally a real-life mission with a group of robots.