Inferring Strategies of Avoidance: Towards Socially Competent Navigation in Human Environments

We present a framework for online navigation planning in multi-agent environments, where no explicit communication takes place among agents, such as pedestrian scenes. Inspired by pedestrian navigation, our approach encodes the concept of coordination into agents’ decision making through an inference mechanism about joint strategies of avoidance. Strategies of avoidance represent avoidance protocols that agents engage in to avoid colliding with each other throughout the scene. In this work, we model such strategies topologically, by employing the formalism of braids. This model allows us to characterize the collective behavior of a set of agents but also to enumerate future scene outcomes for a multi-agent scene. Inspired by the mechanisms of human action interpretation, we design an inference mechanism that enables an agent to infer future strategies of avoidance, by observing agents’ past behaviors. We integrate this mechanism into the agent’s decision making towards generating intent-expressive and socially compliant behaviors that reduce the planning effort for everyone in the scene. Results of statistical significance, generated upon extensive simulation evaluations, indicate faster average uncertainty reduction and faster average destination arrival, compared to purely efficient agents.