Serial order in an acting system: A multidimensional dynamic neural fields implementation

Learning and generating serially ordered sequential behavior in a real, embodied agent that is situated in a partially unknown environment requires that noisy sensory information is used both to control appropriate motor actions and to determine that a particular action has been successfully terminated. While most current models do not address these conditions of embodied sequence generation, we have earlier proposed a neurally inspired model based on Dynamic Field Theory that enables sequences in which each action may take unpredictable amounts of time. Here we extend this earlier work to accommodate heterogeneous sets of actions. We show that a set of matching conditions-of-satisfaction can be used to stably represent the terminal condition of each action and trigger the cascade of instabilities that switches the system from one stable state to the next. A robotic implementation on a vehicle with a camera and a simple robot arm demonstrates the stability of the resulting scheme.