Trajectory Prediction with a Conditional Variational Autoencoder

Abstract

Conventional motion planners do not rely on previous experience when presented with a new problem. Trajectory prediction algorithms solve this problem using a pre-existing dataset at runtime. We propose instead using a conditional variational autoencoder (CVAE) to learn the distribution of the motion dataset and hence to generate trajectories for use as priors within the traditional motion planning approaches. We demonstrate, through simulations and by using an industrial robot arm with six degrees of freedom, that our trajectory prediction algorithm generates more collision-free trajectories compared to the linear initialization, and reduces the computation time of optimization-based planners.