Learning and Recognizing Intentions of Agents from Contextual Cues

Project Leaders: Sven Wachsmuth, Herbert Janssen
PhD Student: Jan Moringen

Biological systems are driven by intentions that determine their behavior. They are often coupled to state changes in the world caused by the biological system. It is, therefore, bounded by its actuator capabilities, on the one side. On the other side, there are typically numerous ways to achieve the same intention. Thus, there is a large perceptual variability and intentions are typically not directly observable. Nevertheless, humans and most of the higher animals are very good at guessing intentions when they see someone acting or even see only a single picture of an action. While the movement itself is often not that informative, contextual cues must be exploited to an extensive degree. The aim of the project is to test the hypotheses that the outcome (goal state) of an ongoing action observed can be anticipated without analyzing the movement itself in detail. Instead various contextual factors are integrated, like scene category, surrounding objects, scene layout, action history, verbal comments, relation of agent and objects, sounds, etc. In this regard, intentions are understood in a more technical sense as low dimensional hidden variables that connect the observed actions, the contextual factors, and the resulting state changes in the environment.