Learning Reactive Robot Navigation Policies from Predictive Trajectory Planning

Humans are used to move in environments, where other humans are present. Due to this fact it is easier for humans to predict the path of a robot, if the robot navigates human-like. In this thesis we present a reactive navigation method to create human-like paths to navigate mobile robots. There are two basic approaches for social navigation of mobile robots. The first group are reactive approaches which create commands based on potentials. On the one hand reactive methods have low computational costs, but one the other hand they often do not create human-like paths. The second group are predictive approaches. Most of the predictive methods better reproduce human paths, but they tend to be computational more involved. In this thesis we aim to combine the benefits of both approaches. We propose a reactive method, which uses locally weighted regression and training data from a predictive path planning method to create acceleration commands to navigate a mobile robot. The use of a non-parametric regression method results in more flexibility to represent the training data compared to the use of a parametric representation. To use locally weighted regression efficiently we reduce the state space, which we use to represent the potentials. In order to do this, we first we represent the behavior, which we aim to learn, by to individual potentials, one for approaching the target and one which represents the cooperative collision avoidance between two agents. We further reduce the state space by taking advantage of symmetries and additional assumptions. In the experiments we demonstrate, that despite of these reductions, we are able to reproduce the paths generated by the predictive path planning method, which we use to create training data. In our experiments with a Pioneer III robot and multiple humans we demonstrate that this method can be used to drive a mobile robot successfully through an environment with pedestrians.