Optimal and Reactive Control for Agile Drone Flight in Cluttered Environments

We present a control framework for agile and reactive collision avoidance with quadrotor drones, applied to the state-based tier of the DodgeDrone Challenge, a simulation competition where the goal is to reach a finish line as fast as possible without collisions. The approach consists of an optimal control scheme with a Log-Sum-Exponential (LSE) obstacle avoidance formulation for efficient motion planning through cluttered environments with a high number of obstacles. The framework copes with a lower update frequency than required in classical Model Predictive Control by combining low-frequency smooth on-trajectory motion planning updates with computationally inexpensive high-frequency linear feedback control. This control structure is augmented with a monitoring function to allow for emergency reactive control with artificial repulsive potential fields in case of an optimal control failure to find a feasible solution in time. The approach is evaluated in extensive simulation runs, which show (1) the effectiveness of the total approach, (2) the increased added value of the LSE formulation for high numbers of obstacles, and (3) the trade-off between the execution time and the success rate. We show that our performance is competitive with the top five teams, even though the approach is, due to its model-based design, not specifically tailored to or trained for this particular challenge but is generally extendable towards other drone applications. Copyright (c) 2023 The Authors.