A Simple Evolutionary Framework for Generating Robust Flight Manoeuvre in a Physically Simulated Flapping Wing Robot

This work demonstrates a successful evolutionary framework to generate the neural network controller of a physically simulated simple 3D flapping wing flying robot which shows robust flight manoeuvre and tolerance to external perturbations. In contrast to recent studies which dealt with the artificial evolution of neural network for steering flight [1,2], this work generates a robot which uses both asymmetric wingbeats and its tail for steering behaviours and retains its flight trajectory even in the presence of external perturbations. The simulated robot has two flat wings with 3 degrees of freedom (dihedral, sweep, and twist) and a tail with 2 degrees of freedom (bend and twist) (figure 1). Two bilaterally symmetric 8 node fully connected continuous-time recurrent neural networks (CTRNN) [3] were used for the controller. The controller has five sensors which are pitch, yaw, roll, altitude, and forward speed, which are connected to every neuron in the circuit. Roll and yaw sensors are connected separately at each side of the circuit to give inputs with opposite signs. In order to avoid unnecessary antiphase movement of the wings, the subnetworks are not connected. Setting the number of neurons to 8 ensures sufficient possibilies for searching two briefly distinct motor circuits which act as the pattern generating motorneurons and the cascaded neural circuits [4]. Figure 1. Overview of the system. During a certain period (Tps<t<Tpe), the robot receives wind gust. After perturbation, the robot should stabilise its attitude while keeping its position as close as possible to the target path. Shaded circles are sensors denoted by Y:yaw, R:roll, P:pitch, V:flight speed, and A:altitude. Lettered unshaded circles represent motorneurons (D:dihedral, S:sweep, T:twist, Ta:tail). Tail bend receives the average output of left and right neurons ((left+right)/2), and tail twist receives the difference (((left-right)/2) of them. The time constants, biases, and connection weights of the neural network are optimised using a geographically distributed genetic algorithm typically with a population of 100 (10×10 grids) for 50000-70000 evaluations. For each simulation, take-off by a single push was used as the starting strategy to prevent initial gliding or somersaulting of the robot. In order to gain maximum fitness, the robot should maintain its flight path which is a straight line of a certain altitude, over the entire simulation time. The evaluation strategy proceeded in two stages – take-off and perturbation. In the early time of the simulation, the robot only has to reach the desired altitude without receiving any perturbation. After successful take-off, a couple of wind gusts of a few seconds duration are presented in order to push the robot off the target path. The perturbed robot should return to the path line as soon as possible and should maintain its flight path continuously. The velocity of the wind and the time of its presentation varied over a small range of random values. In order to boost the evolutionary search, those individuals whose flight altitude was lower than 50cm after 5 seconds from the start of the simulation were regarded as failed robots which have non-repetitive flapping controllers, and those individuals were deleted from the population. This simple evaluation strategy effectively ensures a selection pressure encapsulating many aspects of sophisticated flight control. Because of the highly nonlinear nature of the robot-environment interaction and the tight couplings between the spatial axes for aerodynamic force generation, the robot is forced to capture a variety of capabilities through this simple perturbation method. The evolved robot showed successful compromises between stability and manoeuverability which are two conflicting characteristics of flapping flight, despite the simple and intuitive evolutionary system. Some movie clips are available at http://kucg.korea.ac.kr/~ysshim/flyer.html