EMOBOT: A Robot Control Architecture Based on Emotion-Like Internal Values
Nils Goerke · 2006
Getting autonomous robots to do the things we want them to do is still a challenge. Even the definition of parameters for a given controller type is very hard or realise for interesting robotic tasks. Designing controllers that are easily configured in an adequate way is far more complicated. The idea to make a controller learn an adequate set of parameters and functions for a given task is not completely new, but still not solved sufficiently. Biological entities demonstrate that it is in principle possible to realise such a system. Based on these ideas, a hierarchical robot control structure, with three main information processing branches: a sensory upstream, an actuatory downstream, and a controller was developed. The controller consists of two main units: an internal value system (IVS) and a learning action controller. The internal values (Drives, Emotions) are fed with the results from the sensory upstream. Several (17) primary and virtual sensor modules have been implemented within the different layers of the sensory upstream. The internal values are the decision basis for the learning action controller. The action controller activates the different action modules within the hierarchy of the actuatory downstream. We have implemented a total of 14 action modules that realise basic, simple and complex motor actions, including several senso-motoric closed loop controlled behaviours. The developed internal value system is aligned with psychological terms like "Drives" and "Emotions", without attempting to model these psychological concepts. The implemented internal values, like "Hunger", "Fatigue", "Fear", "Curiosity" or "Homesickness" are just meant as labels to indicate their specific functionality. The action selection unit has been implemented as a matrix implementation of a switching controller. The content of the matrix is a result of pre-design and subsequent learning. Training the action selection is performed using part of the internal values as a reinforcement signal, as well as human teacher generated reinforcement. The different realisation of the EMOBOT approach, with different kind of robots (grid world based, real value based and a real robot system) and with different set of internal values ("Drives" and "Emotions") show that the idea of controlling a robot with a set of metavalues, that are not directly connected to the outer world is a feasible approach to generate a complex behaviour. For a given configuration of the behaviour primitives and the action selection and the internal value system it is difficult to predict the overall behaviour of the robot system. The other way round, to construct a special set-up with the goal to create a specific behaviour is far more complicated, but on at the same time far more interesting. The results obtained, and partially presented within this work show clearly that the approach of learning hierarchical action selection based on internal values is a valuable way of finding robust robot controllers. Although a lot of scientific questions still remain unanswered we believe that the results presented are encouraging for the future.