A Mori-Zwanzig Formalism Based Estimation Approach for Dynamic Obstacle Avoidance

We present a novel learning-based approach for dynamic obstacle avoidance in multi-robot systems using the Mori-Zwanzig (M-Z) formalism. The key innovation lies in developing a method that enables an ego robot to predict the trajectory of a dynamic obstacle that is outside its sensing range, solely by observing the historical trajectory of an ally robot. We apply M-Z formalism to rigorously justify the use of sequential data in predicting obstacle behavior, and provide theoretical bounds on the prediction accuracy. Simulation results demonstrate that our method reduces collision rates compared to scenarios without obstacle trajectory prediction.