Learning to Swarm with Knowledge-Based Neural Ordinary Differential Equations

Abstract

Understanding single-agent dynamics from collective behaviors in natural swarms is crucial for informing robot controller designs in artificial swarms and multiagent robotic systems. However, the complexity in agent-to-agent interactions and the decentralized nature of most swarms pose a significant challenge to the extraction of single-robot control laws from global behavior. In this work, we consider the important task of learning decentralized single-robot controllers based solely on the state observations of a swarm's trajectory. We present a general framework by adopting knowledge-based neural ordinary differential equations (KNODE) -a hybrid machine learning method capable of combining artificial neural networks with known agent dynamics. Our approach distinguishes itself from most prior works in that we do not require action data for learning. We apply our framework to two different flocking swarms in 2D and 3D respectively, and demonstrate efficient training by leveraging the graphical structure of the swarms' information network. We further show that the learnt singlerobot controllers can not only reproduce flocking behavior in the original swarm but also scale to swarms with more robots.

Distributed Event-Triggered Flocking Control of Lagrangian Systems

“…We evaluate flocking behavior through the metrics described in [38]. The average minimum distance to a neighbor measures the cohesion between the agents.…”

Learning Rigidity-based Flocking Control with Gaussian Processes

“…The first two Figures show how the same external force has different behaviours over agents 1 and 3, due to the different role they play in the overall motion of the system, while Figure 6b shows the unknown force introduced is different and although the y-coordinate is not constant, the proposed learningbased control law does well managing both components. In addition, inspired by [19], we evaluate the performance with some metrics. In particular, we study the average velocity to provide a measure of the alignment of the velocities of the agents.…”

“…Decentralized machine learning control algorithms for multi-agent systems has been recently studied in [15], [16] using graph neural networks and also in [17], [18]. In particular recent learning-based methods for flocking control of second-order agents can be found in [19], [14], [20], but none of them includes 3D flocking control with stability guarantee in the performance. To the best of the authors' knowledge, there are no available results for the design of a learningbased flocking control law for double integrator agents under partially unknown dynamics, based on online learning datadriven models with exponential stability guarantees.…”