Robot activity adaptation for safe human-robot collaboration based on probabilistic risk modeling

Abstract

In this paper we describe an approach for robot activity adaptation in a cognitive human-robot collaboration system, based on user-related information. Core element is a Bayesian Network model, which is the basis for reasoning about the risk of a user working in a collaborative task with a partially autonomous robot in a shared workspace. For local and global inference, the probabilistic model thereby combines scene information on situation and activity classes of different body parts and the whole user. All compiled information about the user thereby relies on the predictions from the real-time human body posture tracking. In order to improve the quality and reliability of the risk estimation we therefore maintain a locally-resolved posture estimation quality model and inject it into early stages of the inference process. For the activity adaptation we employ an optimization based dynamic path planning approach, which processes the spatially resolved inferences from the Bayesian Network in order to find robot motions which are optimal for the current situation in the shared workspace.