Trajectory Pruning with Neural Networks for Efficient Monte-Carlo based Quadrotor Flight Envelope Prediction

Abstract

Flight envelope prediction is a challenging task where one of the difficulties is that widely used methods, like the level set methods, are impractical for systems with more than four coupled state dimensions due to the “curse of dimensionality”. Monte-Carlo simulation based approach suffers less from this, however a large number of simulations is needed to predict a flight envelope, while not all simulations directly contribute towards estimating its boundary. This paper proposes a new approach to alleviate this with the use of machine learning techniques that can distinguish more valuable control sequences within the random samples; this knowledge could be used to reduce the number of simulations required to predict the boundary.
An artificial neural network containing a long short-term memory is trained to map a randomly sampled control sequence to the relative position of the resultant end state of the trajectory compared to a predetermined reference reachable set. This trained network is applied for Monte-Carlo based reachability analysis of a dynamic model with model parameter changes compared to the reference model, which is able to reject 50% of randomly sampled sequences while at most 95% of the rejected samples would not have contributed towards reachable set boundary estimation.