Högskolan i Skövde

A dilemma arises when sequence generation is implemented on embodied autonomous agents. While achieving an individual action goal, the agent must be in a stable state to link to fluctuating and time-varying sensory inputs. To transition to the next goal, the previous state must be released from stability. A previous proposal of a neural dynamics solved this dilemma by inducing an instability when a "condition of satisfaction" signals that an action goal has been reached. The required structure of dynamic coupling limited the complexity and flexibility of sequence generation, however. We address this limitation by showing how the neural dynamics can be generalized to generate hierarchically structured behaviors. Directed couplings downward in the hierarchy initiate chunks of actions, directed couplings upward in the hierarchy signal their termination. We analyze the mathematical mechanisms and demonstrate the flexibility of the scheme in simulation. © 2012 Springer-Verlag.

The implementation of sequence learning in robotic platforms offers several challenges. Deciding when to stop one action and continue to the next requires a balance between stability of sensory information and, of course, the knowledge about what action is required next. The work presented here proposes a starting point for the successful execution and learning of dynamic sequences. Making use of the NAO humanoid platform we propose a mathematical model based on dynamic field theory and reinforcement learning methods for obtaining and performing a sequence of elementary motor behaviors. Results from the comparison of two reinforcement learning methods applied to sequence generation, for both simulation and implementation, are provided.

Behavioral studies on the activation of affordances by understanding observation and action sentences on graspable objects show a direct relationship between the canonical orientation of graspable objects, their dimension and the kind of grip required by those objects to be grasped. The present work introduces the concepts of Dynamic Field Theory for modeling the results observed in the behavioral studies previously mentioned. The model was not only able to replicate qualitatively similar results regarding reaction times, but also the identification of same versus different object and the distinction between observable versus action sentences. The model shows the potential of dynamic field theory for the design and implementation of brain inspired cognitive systems. © 2012 Springer-Verlag.

In this article, a generic CPG architecture is used to model infant crawling gaits and is implemented on the NAO robot platform. The CPG architecture is chosen via a systematic approach to designing CPG networks on the basis of group theory and dynamic systems theory. The NAO robot performance is compared to the iCub robot which has a different anatomical structure. Finally, the comparison of performance and NAO whole-body stability are assessed to show the adaptive property of the CPG architecture and the extent of its ability to transfer to different robot morphologies. © 2011 IEEE.

The work presented in this paper builds on previous research which analysed human action segmentation in the case of simple object manipulations with the hand (rather than larger-scale actions). When designing algorithms to segment observed actions, for instance to train robots by imitation, the typical approach involves non-linear models but it is less clear whether human action segmentation is also based on such analyses. In the present paper, we therefore explore (1) whether linear models built from observed kinematic variables of a human hand can accurately predict human action segmentation and (2) what kinematic variables are the most important in such a task. In previous work, we recorded speed, acceleration and change in direction for the wrist and the tip of each of the five fingers during the execution of actions as well as the segmentation of these actions into individual components by humans. Here, we use this data to train a large number of models based on every possible training set available and find that, amongst others, the speed of the wrist as well as the change in direction of the index finger were preferred in models with good performance. Overall, the best models achieved R2 values over 0.5 on novel test data but the average performance of trained models was modest. We suggest that this is due to a suboptimal training set (which was not specifically designed for the present task) and that further work be carried out to identify better training sets as our initial results indicate that linear models may indeed be a viable approach to predicting human action segmentation.

Choice behaviour where outcome-contingencies vary or are prohabilistic has been the focus of many benchmark tasks of infant to adult development in the psychology literature. Dynamic field theoretic (DFT) investigations of cognitive and behavioural competencies have been used in order to identify parameters critical to infant development. In this paper we report the findings of a DFT model that is able to replicate normal functioning adult  performance on the Iowa gambling task (IGT).  The model offers a simple demonstration proof of the parsimonious reversal learning alternative to Damasio’s somatic marker  explanation of IGT performance. Our simple model demonstrates a potentially important role for reinforcement/reward learning to generating behaviour that allows for advantageous performance. We compare our DFT modelling approach to one used on the A-not-B infant paradigm and suggest that a critical aspect of development lies in the ability to flexibly trade off perseverative versus exploratory behaviour in order to capture statistical choice-outcome contingencies. Finally, we discuss the importance of an investigation of the IGT in an embodied setting where reward prediction learning may provide critical means by which adaptive behavioural reversals can be enacted.