Towards learning movement in dense crowds for a socially-aware mobile robot

Robots moving in a crowd occasionally reach situations where they need to decide whether to give way to a human or not, a situation we call a micro-conflict and model with a two player game. We collect data from a robot controlled by a human operator and use three different supervised learning algorithms (random forest, SVM and neuroevolution) to create a decision maker module which imitates the human operator’s behavior in micro-conflicts. Results show that the neuro-evolution based decision-maker gives the best performance under scenarios with various crowd density and urgency. In addition, we found that the neuroevolution method generalizes better to environments very different from those in the training set.