An Interpretable Monitoring Framework for Virtual Physics-Based Non-Interfering Robot Social Planning

A majority of collision avoidance and motion planning algorithms deployed on autonomous mobile robots tend to be reactive to the presence and motion of nearby dynamic actors. While these algorithms can produce collision free navigation, they do not necessarily follow social protocol and exhibit unnatural motions that force actors to behave very differently from their planned paths. As humans, we can reason about why and how we might interfere with others and use this reasoning to alter our motion proactively. We also adapt our motion based on different priorities, which affect how we accommodate and interact with each other. In this letter, we propose an approach for a mobile robot to generate similar non-interfering and priority-based behaviors that pertain to how the robot accommodates dynamic actors. We augment a very efficient but reactive virtual physics-based planner with Hidden Markov Models and a Decision Tree-based monitor that: i) predicts if the robot will interfere with a nearby actor, ii) explains what factors cause the prediction, iii) plans priority-based corrective actions, and iv) updates the prediction model at runtime to improve and refine robot behaviors. Our approach is validated with simulations and extensive experiments on ground and aerial vehicles in the presence of dynamic actors.