Humanoid Tactile Gesture Production using a Hierarchical SOM-based Encoding

Abstract

The existence of cortical hierarchies has long since been established and the advantages of hierarchical encoding of sensor-motor data for control, have long been recognized. Less well understood are the developmental processes whereby such hierarchies are constructed and subsequently used. This paper presents a new algorithm for encoding sequential sensor and actuator data in a dynamic, hierarchical neural network that can grow to accommodate the length of the observed interactions. The algorithm uses a developmental robotics methodology as it extends the Constructivist Learning Architecture, a computational theory of infant cognitive development. This paper presents experimental data demonstrating how the extended algorithm goes beyond the original theory by supporting goal oriented control. The domain studied is the encoding and reproduction of tactile gestures in humanoid robots. In particular, we present results from using a Programming by Demonstration approach to encode a stroke gesture. Our results demonstrate how the novel encoding enables a Nao humanoid robot with a touch sensitive fingertip to successfully encode and reproduce a stroke gesture in the presence of perturbations from internal and external forces.

Learning Robot Control Using a Hierarchical SOM-Based Encoding

“…A three-level Kohonen SOM has also been used in mobile robots to learn novel behavior plans [36], and also to recognize human gestures [37]. HSOME [9], [10] benefits from such a hierarchical representation increasing its power of representation exponentially with the number of layers. HSOME is a more general approach than its predecessor in that its structure is generic with no formal limitations on the height of the hierarchy and with no hand coded connections.…”

“…A node on every layer learns a compact representation of decayed activations from the level below. The resulting architecture is built on top of a biologically-inspired computational model of infant cognition [9], [10].…”

“…Cohen et al [8] introduced a computational model of infant cognition, i.e., the Constructivist Learning Architecture (CLA), an Information Processing approach suggesting that infant knowledge is the result of processing information in the environment. In previous work, we introduced a novel algorithm based on the CLA architecture that facilitated motor skill learning [9], [10]. The proposed Hierarchical Self-Organizing Map-based Encoding (HSOME) algorithm is structurally built as a hierarchy of temporally augmented layers of Kohonen Self-Organizing Maps (SOM).…”

“…Concretely, it was demonstrated that the HSOME algorithm (technical details are presented in Section III) can support learning in abstract domains (where robot observations and actions are represented by integers) and physical humanoid robots learning tactile gestures. HSOME successfully resolved ambiguity problems in hidden states during decision making and seamlessly recovered from perturbations internal (e.g., frictional forces on the skin and in the joints) and external (i.e., a human or artificial push) to the system [9], [10]. While a large part of the aforementioned research is focusing on humanoid robots and high level cognition, there is another research direction showing a great interest in biologically inspired robotics to approach the problem of learning motor skills from an evolutionary perspective.…”

Hierarchy supervised som SOM neural network applied for classification problem

“…Up to now, the conventional multivariate statistical techniques (cluster analysis, linear discriminant analysis) including unsupervised (Kohonen's network) and supervised (Bayesian network) artificial neural networks were compared for as general tools for the classification and identification problem [18][19][20][21][22]. One of those is the Self-Organizing Map (SOM) which was proposed by Tevou Kohonen [14].…”