Abstract:
As mobile robots start operating in environments crowded with humans, human-aware navigation is required to make these robots navigate safely, efficiently and in socially compliant manner. People navigate in an interactive and cooperative fashion so that, they are able to find their path to a destination even if there is no clear route leading to it. There are significant efforts to solve this problem for mobile robots; however, they are not scalable to high human density and learning based approaches depend heavily on the context and configuration of the set they are trained with. We develop a method which infers initial trajectories from Gaussian processes and updates these trajectories jointly for all agents using a cost based interaction approach. We condition Gaussian processes online with the best hypothesis at each step of prediction horizon. The method is tested on a common public dataset and it is shown that it outperforms two state-of-the-art approaches in terms of human-likeness of predicted trajectories.
