Advisor — Master Thesis "Dynamically Leveraging Sensorimotor Regularities by Integrating Perception and Behavior Generation"
Co-Advised with Aravind Battaje & Alexander Koenig; Supervised and Examined by Oliver Brock & Marc Toussaint
Student
Malte Bernhard
Abstract
This thesis presents a method for efficiently leveraging sensorimotor regularities (SMRs), which encode how an agent’s motor commands modulate sensory input in relation with its environment, offering a structured way to reduce perceptual uncertainty and guide purposeful behavior. We embed this interaction structure in the Active InterCONnect (AICON) framework, wherein state estimation and action generation are tightly coupled. By modeling SMRs as active interconnections among relevant state, motor, and sensor components, an agent is equipped with capabilities of interactively perceiving its environment. Consequently it is able to actively increase its robustness to disturbances through behavioral adaptation. Our approach provides the means to modularly aggregate SMRs to a behavior, without relying on learning methods. To demonstrate this, we implement the scenario of a drone following a moving target while exposed to various disturbances, in a simulated environment. Experimental comparisons to a baseline controller reveal that our AICON-based approach efficiently leverages SMRs to adapt its behavior to these disturbances, maintaining high estimation accuracy and task performance. These findings underscore the potential of coupling perception and action through sensorimotor regularities to achieve robust real-time control.
Related Publication
Submitted at RSS 26.
More Information
More information can be found here.