Distributed Model Predictive Control on Nano-UAV Swarms Using ADMM and DAQP

Sammanfattning

Nano-UAV swarms operating in constrained environments require real-time collision avoidance on resource constrained hardware, which poses significant computational challenges. This thesis presents a distributed Model Predictive Control (MPC) framework addressing this challenge for nano-UAV swarms, combining the DAQP solver for embedded systems with an ADMM-based collision avoidance algorithm. This decentralized control approach enables the members of the swarm to solve their own local optimization problem while coordinating collision avoidance through a main computer. To solve these problems on resource-constrained hardware, the global optimization problem must be decoupled. This allows each drone to independently solve a reduced version depending only on its own states and limited neighbor information. This decoupling has been shown in this thesis and has proved to lead to robust collision avoidance behavior. The nano-UAVs are modeled as linear point-mass objects enabling 6 degrees of freedom using control signals for accelerations in world fixed x-, y- and z-directions. The control objective behind the MPC cost function includes state error as well as control effort costs. Collision avoidance is enforced through penalty terms in the MPC cost function rather than explicit inequality constraints, ensuring the problem remains feasible while keeping computational complexity low. A radio compression scheme is developed to enable transmission of full acceleration horizons within the hardware packet size limits of the Crazyradio PA. Move-blocking is investigated as a strategy to effectively expand the prediction horizon of each member of the swarm without increasing the number of decision variables. It is shown that well-designed move-blocked configurations enable earlier collision detection and avoidance, making the approach particularly suitable for calculations on resource-constrained hardware. The control framework is validated through Python-based software simulationsand Hardware-in-the-Loop experiments using the Crazyflie 2.1 platform. Demonstrating the feasibility of the approach within hardware constraints through physical flight validation remains as future work. Results demonstrate robust collision avoidance behavior across the evaluated ADMM configurations, with collision avoidance performance shown to depend on the tuning of the ADMM penalty parameter ρ and the number of iterations.