Learning reciprocal actions for cooperative collision avoidance in quadrotor unmanned aerial vehicles

The ability to avoid collisions with each other is one of the fundamental requirements for autonomous unmanned aerial vehicles (UAVs) to be safely integrated into the civilian airspace, and for the viability of multi-UAV operations. This paper introduces a new approach for online cooperative collision avoidance between quadcopters, involving reciprocal maneuvers, i.e., coherent maneuvers without requiring any real-time consensus. Two maneuver strategies are presented, where UAVs respectively change their speed or heading to avoid a collision. A learning-based framework that trains these reciprocal actions for collision evasion (called TRACE) is developed. The primary elements of this framework include: 1) designing simulated experiments that cover a variety of UAV–UAV approach scenarios; 2) performing optimization to identify speed/heading change actions that satisfy safety constraints while minimizing the energy cost of the maneuver; and 3) using the offline optimization outcomes to train classifier (via ensemble bagged tree) and function approximation (via neural networks and Kriging) models for respectively selecting and encoding the avoidance actions. Trajectory generation and dynamics/controls are incorporated in the simulation environment used for training and testing. Over 90% accuracy in action prediction and over 95% success in avoiding collisions is observed when the trained models are applied to simulated unseen test scenarios.