Autonomous mobile robots in the modern industrial environment are surrounded by numerous moving objects, which the robot is able to track using its sensors. Often the future position of such objects is needed, therefore we examined the usage of time series methods for trajectory prediction with an emphasis on neural network models. We showed that encoder-decoder LSTM model can successfully learn periodic patterns in the movement of a robot. Enhanced version of this architecture was used to predict short-term trajectories, which we implemented in practice as a ROS node for trajectory prediction.
|
