Robust RL–SMC Hierarchical Control for Leader–Follower Quadrotor Swarms in 3D Obstacle-rich Environments

Abstract

Coordinated quadrotor swarms operating in cluttered environments must simultaneously achieve robust low-level stabilization and high-level, constraint-aware motion generation for safe formation flight. This paper proposes a hierarchical RL-SMC control framework for leader–follower quadrotor swarms navigating three-dimensional obstacle-rich workspaces that include explicit no-fly regions. At the low level, a boundary-layer sliding mode attitude controller provides disturbance-tolerant tracking of commanded Euler references, while a mid-level translational loop converts bounded velocity commands into physically feasible thrust and attitude setpoints. At the high level, a reinforcement learning (RL) policy learns to generate bounded planar acceleration references for the leader, using an observation and reward design that explicitly accounts for both the leader state and the worst-case clearance of the formation footprint to obstacles and restricted zones. The followers track leader-aligned V-shape offsets, enabling formation motion without direct learning-based actuation. Simulation results in a constrained 3D environment demonstrate goal-reaching while maintaining formation integrity and satisfying safety margins, with smooth bounded guidance actions and no collisions or no-fly violations. The proposed architecture combines the adaptability of RL with the robustness properties of variable-structure control in a modular, safety-oriented swarm control stack.