Interactive Costmaps: Integrating Prediction and Planning with Counterfactual Reasoning

Robot navigation in dynamic environments is a challenging problem, especially in situations where the variability arises from other active agents. Predicting motion in such situations requires reciprocal and interactive models, which also need to be computationally efficient. This must be integrated within a motion synthesis framework that takes into account other aspects of costs associated with the task. I will present a novel framework to generate interactive costmaps that identifies the intention of agents, translates them to interactive motion predictions over time (using our Counterfactual Reasoning framework) and represent them as costs that may be added to other layers within a motion planning framework. Interactive costmaps produce smoother and more efficient navigation while decreasing number of near collisions overall at a comparable computational cost.