A neuroevolution-based learning of reciprocal maneuver for collision avoidance in quadcopters under pose uncertainties

This paper focuses on the idea of energy efficient cooperative collision avoidance between two quadcopters. Two strategies for reciprocal online collision-avoiding actions (i.e., coherent maneuvers without requiring any real-time consensus) are proposed. In the first strategy, UAVs change their speed, while in the second strategy they change their heading to avoid a collision. The avoidance actions are parameterized in terms of the time difference between detecting the collision and starting the maneuver and the amount of speed/heading change. These action parameters are used to generate intermediate way-points, subsequently translated into a minimum snap trajectory, to be executed by a PD controller. For realism, the relative pose of the other UAV, estimated by each UAV (at the point of detection), is considered to be uncertain – thereby presenting substantial challenges to undertaking reciprocal actions. Performing supervised learning based on optimization derived labels (as done in prior work) becomes computationally burden-some under these uncertainties. Instead, an (unsupervised) neuroevolution algorithm, called AGENT, is employed to learn a neural network (NN) model that takes the initial (uncertain) pose as state inputs and maps it to a robust optimal action. In neuroevolution, the NN topology and weights are simultaneously optimized using a special evolutionary process, where the fitness of candidate NNs are evaluated over a set of sample (in this case, various collision) scenarios. For further computational tractability, a surrogate model is used to estimate the energy consumption and a classifier is used to identify trajectories where the controller fails. The trained neural network shows encouraging performance for collision avoidance over a large variety of unseen scenarios.