Control of Agent Swarms Using Generalized Centroidal Cyclic Pursuit Laws

Arpita Sinha, Debasish GhoseIJCAI 2007

Introduction

In the field of swarm robotics, achieving coordinated collective behavior—such as rendezvous or formation control—through decentralized rules is a primary challenge. A traditional approach is cyclic pursuit, where agents follow their neighbors in a fixed sequence (e.g., agent 1 follows agent 2, agent 2 follows agent 3). While effective, standard cyclic pursuit is often rigid, assuming uniform gains and fixed network topologies. This paper introduces Generalized Centroidal Cyclic Pursuit, a more flexible framework where agents follow a weighted average (centroid) of their peers.

Generalized Centroidal Cyclic Pursuit

The proposed method generalizes the pursuit strategy in two key ways. First, it utilizes a "pursuit sequence" ($\eta$) that allows an agent to weigh the positions of multiple neighbors rather than just one. Second, it incorporates heterogeneous gains ($k_i$), allowing individual agents to have different sensitivities to their neighbors' positions.

A swarm of agents connected via a pursuit network, illustrating the centroidal pursuit topology.

The kinematics of each agent $i$ are modeled such that they move toward a point defined by the weighted average of other agents' positions. This system is represented as a linear differential equation $\dot{X} = AX$, where the properties of matrix $A$ determine the swarm's collective behavior.

Stability and Linear System Analysis

A critical aspect of this research is determining the conditions under which the swarm remains stable and converges to a single point. The authors employ Gershgorin’s Disc Theorem to analyze the eigenvalues of the system matrix $A$. Since $A$ is singular, the swarm's center of mass remains stationary or follows a predictable path while agents converge relative to one another.

Gershgorin discs in the complex plane representing stability regions for different controller gain configurations.

The analysis reveals that stability is surprisingly robust. The swarm can converge even if some gains are zero or negative, provided the overall system satisfies specific algebraic conditions derived from the disc locations in the complex plane.

Key Invariance Properties

The most significant contribution of this work is the discovery of two invariance properties:

  1. Reachable Point Invariance: The final rendezvous location depends solely on the agents' initial positions and their individual gains ($k_i$). It is entirely independent of the pursuit sequence ($\eta$). This allows for dynamic switching of the network topology mid-operation without altering the swarm's target destination.
  2. Stability Invariance: The stability of the rendezvous is also independent of the pursuit sequence. As long as the gains are set correctly, the specific distribution of weights among neighbors does not affect the convergence.

Results and Future Directions

Theoretical proofs and numerical simulations with 5 to 20 agents confirm these findings. The authors demonstrated that even with multiple switches in the pursuit network, the swarm reaches the same destination. Furthermore, if the system is tuned for specific instabilities, the agents can be made to "flock" along a straight-line trajectory.

While the current model assumes linear kinematics and full communication, future research will explore non-holonomic constraints, communication delays, and obstacle avoidance to bring these generalized pursuit laws closer to real-world deployment.

Summary updated

BibTeX
@inproceedings{sinha2007control,
  author = {Arpita Sinha and Debasish Ghose},
  title = {{Control of Agent Swarms Using Generalized Centroidal Cyclic Pursuit Laws}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure