Funded by: DFG

Runtime: 01.01.2022 -  28.02.2025

Project Coordinators: Prof. Dr.-Ing. Nele Rußwinkel (University of Lübeck) & Prof. Dr.-Ing. Stefan Kopp (Bielefeld University)

Research Associates: Nils Wendel Heinrich (M.Sc., University of Lübeck) & Annika Österdiekhoff (M.Sc., Bielefeld University)

Student Researchers: Martin Stuwe (B.Sc., University of Lübeck) & Carolin Klute (B.Sc., Bielefeld University)

The COMPAS project is a collaboration between the Social Cognitive Systems group (Universität Bielefeld) and the Human-Aware AI group (Universität zu Lübeck) and part of the DFG priority program 2134 The Active Self, which tackles questions like “how the agency experience affects cognitive processes” and “how a self-representation might be established in artificial agents”.

Motivation

Humans can learn new skills such as, e.g., inline skating through a process of gaining more and more control over how their actions will yield specific consequences. At first, it is necessary to put full attention to achieve a movement toward a goal a few meters away. After some training, one is able to skate and at the same time focus on a conversation with a friend. Even when going on coarse ground – with a lot of noise in haptic feedback – the skater is still confident that she has control over the situation. And, after taking off the skates, she will immediately switch back to the “normal-no-wheel” state of movement control. This example demonstrates that autonomous embodied agents need to be able to adapt how they control their situated action based on the current state of their body in a possibly uncertain environment and, importantly, an assessment of how they feel “in control” of acting in it.

Goals and Procedure

In COMPAS, we specifically investigate situated action control within environments of varying uncertainty and complexity and focus on the sense of control component of the agency experience. We hypothesize that the sense of control is grounded in sensorimotor experiences and strongly linked to predictive coding and higher-level perceptual processing. To test our hypotheses, we rely on a combination of human eye-tracking experiments and computational cognitive modeling using our custom-developed environment.

Innovation and Prospects

Robotic systems equipped with the investigated control architecture could make use of cognitive mechanisms that would enable new ways of collaborative human-robot interaction. A robot may inform its human partner of its low SoC, switch to a safer mode of operating, or ask for help. Further, we investigate mechanisms needed for collaborative situated interaction, like embodied Theory of Mind and mental (co-)simulation. For example, a robot equipped with our computational cognitive model may be able to anticipate not only its own action field, but also another one’s actions and their expected outcomes.

Project Partners

With the start of the second funding phase, we have initiated a collaboration with the Body Social Cognition and Action lab at LMU. The goal is to explore further psychophysiological markers of the sense of control that arise in our dynamic task. Here we apply the EEG measurement method in experiments in Munich.

Link to Project Details

https://activeself.de/cognitive-modelling-of-the-predictive-active-self-in-situated-action-compas/

Figure taken from Heinrich et al., 2023 (conference submission, MathPsych/ICCM/EMPG 2023). At the sensorimotor control layer, predicted sensory feedback is matched with the actual sensory feedback obtained from the environment. The degree to which both match each other is translated into the low-level Sense of Control which in turn is interpreted into a high-level Sense of Control. At the cognitive control layer, the high-level Sense of Control will directly affect which action possibility from the action field is selected and therefore pursued as the next action goal.