Connectivity-preserving cluster formation for heterogeneous multi-robot systems

This thesis addresses autonomous cluster formation in heterogeneous mobile robot teams subject to limited sensing range. Building on a passivity-based connectivity-maintenance framework, real multiplicative adjacency weights encode sensing constraints and collision avoidance within a graph Laplacian structure. Two alternative distributed spatial control laws are proposed. The first extends the real weights with complex unit-phasor cluster signatures, yielding a Hermitian Laplacian, and applies the resulting multi-partite consensus protocol directly to robot positions encoded as complex numbers, driving same-label robots to a common point through the graph algebraically. The second operates entirely on real quantities: a spatial-cohesion law provides local same-label attraction and inter-cluster repulsion, while a distributed-centroid estimator based on Proportional Dynamic Average Consensus enables graph wide aggregation of same-label robots even without direct edges. The framework is implemented on a heterogeneous testbed combining DJI RoboMaster S1 ground vehicles and Crazyflie quadrotors under ROS2 with motion-capture localization. The RoboMaster platform was converted from its closed factory configuration into a custom open architecture supporting low-level control. A Webots software-in-the-loop (SIL) environment was developed for both platforms, enabling simulation-to-real transfer with zero code changes. Both control laws are first validated on real hardware using a homogeneous swarm of six Crazyflie quadrotors, confirming successful four-cluster formation and connectivity preservation under real flight dynamics, sensor noise, and communication delays. The framework is then evaluated on a mixed swarm of four ground robots and four quadrotors in the Webots SIL environment, demonstrating successful multi-cluster formation, connectivity preservation and robustness under realistic sensing constraints in the heterogeneous case. The results lay the groundwork for future real heterogeneous deployment, dynamic cluster reassignment and opinion-dynamics-based emergent clustering.