Cloud-Based Scheduling Mechanism for Scalable and Resource-Efficient Centralized Controllers

TL;DR

The paper addresses the scalability and resource challenges of Centralized Nonlinear Model Predictive Controllers ($CNMPC$) for multi-agent systems in cloud robotics. It introduces a Kubernetes-based scheduling framework comprising a mission planner, a scheduler, a UDP proxy-based data flow, and dynamic per-CNMPC resource allocation to deploy CNMPCs across worker nodes as the agent count changes, with the number of CNMPCs adapting via empirical deployment strategies. Experiments on an Ericsson real-time cloud test bed with up to 21 UAVs demonstrate that the scheduler maintains bounded processing times and resource utilization while enabling seamless migration between CNMPC instances. This approach advances scalable, flexible cloud-controlled robotics, enabling efficient centralized control for large swarms and paving the way for real-world cloud robotics deployments.

Abstract

This paper proposes a novel approach to address the challenges of deploying complex robotic software in large-scale systems, i.e., Centralized Nonlinear Model Predictive Controllers (CNMPCs) for multi-agent systems. The proposed approach is based on a Kubernetes-based scheduling mechanism designed to monitor and optimize the operation of CNMPCs, while addressing the scalability limitation of centralized control schemes. By leveraging a cluster in a real-time cloud environment, the proposed mechanism effectively offloads the computational burden of CNMPCs. Through experiments, we have demonstrated the effectiveness and performance of our system, especially in scenarios where the number of robots is subject to change. Our work contributes to the advancement of cloud-based control strategies and lays the foundation for enhanced performance in cloud-controlled robotic systems.

Figures (7)

• Figure 1: High-level overview of the proposed framework in the real-time cloud with its core components.
• Figure 2: Overview of the block diagram. Real-time cloud includes the mission planner, the scheduler, the controllers, and the proxy server (UDP tunnel). Robots include the multi-agent system and the agent end of the UPD tunnel.
• Figure 3: System overview with the real-time cloud test bed operating the scheduling mechanism within a Kubernetes cluster.
• Figure 4: The Kubernetes cluster experimental setup with a snapshot of the external VM that hosts the simulator. Three CNMPCs are contributed to two worker nodes, and control the trajectory of 21 UAVs.
• Figure 5: Comparison of resources utilization with (right figure) and without (left figure) scheduling mechanism. The number of agents exceeds 10 (5 on CNMPC$_1$ - Worker node 1, and 5 on CNMPC$_2$ - Worker Node 2), and 15 (8 on CNMPC$_1$ - Worker node 1, and 7 on CNMPC$_2$ - Worker node 2), and scale up to 21 (7 on CNMPC$_1$ - Worker node 1, 7 on CNMPC$_2$ - Worker node 1, and 7 on CNMPC$_3$ - Worker node 2).