Grounding humanoid visually guided walking: From action-independent to action-oriented knowledge

Abstract

In the context of humanoid and service robotics, it is essential that the agent can be positioned with respect to objects of interest in the environment. By relying mostly on the cognitivist conception in artificial intelligence, the research on visually guided walking has tended to overlook the characteristics of the context in which behavior occurs. Consequently, considerable efforts have been directed to define action-independent explicit models of the solution, often resulting in high computational requirements. In this study, inspired by the embodied cognition research, our interest has focused on the analysis of the sensory-motor coupling. Notably, on the relation between embodiment, information, and action-oriented representation. Hence, by mimicking human walking, a behavior scheme is proposed and endowed the agent with the skill of approaching stimuli. A significant contribution to object discrimination was obtained by proposing an efficient visual attention mechanism, that exploits the redundancies and the statistical regularities induced in the sensory-motor coordination, thus the information flow is anticipated from the fusion of visual and proprioceptive features in a Bayesian network. The solution was implemented on the humanoid platform Nao, where the task was accomplished in an unstructured scenario.

Introduction

The automation of visually guided walking has arguably adopted in its infancy the so called cognitivist approach to artificial intelligence (AI), which, under the Cartesian dualist influence, has tended to look at physical and mental processes as belonging to different realms. Significant progress has been obtained from this view, though the gains are still distant from the sophistication observed in natural behavior.

Among the several challenges reported in the literature, one is undoubtedly to achieve reliable perception from noisy sensory data. Since the sensory input goes through a process of symbolization, and cognition involves computations over symbols, the physical context at which the latter emerged is no longer available to the cognitive process. In other words, in abstracting cognition from the context, information is inevitably lost.

To cope with the difficulties of perceiving the object while moving, several methodologies that flourished in the machine learning research have been employed (e.g. Markovian models, support vector machines, among others). These attempts have produced impressive results, although, by keeping intact the fundamental premise of decoupling between bodily and mental processes, they have relied on expensive resources in the form of context-free explicit models, knowledge databases, and intensive computation. Thus, the processing bottleneck has impacted the autonomy and the reactivity of the agent, and extraneous variables (i.e. unmodeled phenomena) have been controlled by adapting the scene to the task, which has compromised the generality of the solution.

In the last decades, a different perspective has been adopted to study natural behavior from the multi-disciplinary research on embodied cognition (EC), where knowledge representation is thought to be grounded in the physical interaction with the environment. Unlike the cognitivist approach, emergent behavior would not be explicitly represented nor planed in advance. The analysis of the sensory-motor coupling in natural tasks, from a dynamic system perspective, appears as a promising research direction that can provide more efficient, robust, and autonomous solutions.

However, adopting the EC methodology also poses important challenges to robotists, in particular, when fulfilling the requirements underlying the physical grounding hypothesis. Firstly, the autonomous evolution of the system, as it happens in natural beings, would ideally occur under a phylogenetic architecture that can modify itself. Secondly, the ontology of the system must ensure knowledge acquisition for diverse purposes, by fusing information from different sensory modalities. Lastly, enactment conditions the development of cognitive skills to the sensory-motor coupling and the interaction with the environment, thus knowledge acquisition is a slow process analogous to the natural one.

In view of the advantages and the challenges encountered in the aforementioned research paradigms, we have opted for an intermediate perspective, aiming at obtaining both generality and autonomy for applications in service robotics. Thereby, in this work we study the task of approaching and positioning in relation to visual stimuli. For this, we adopt the cognitivist assumption that human employs action-independent knowledge for localizing the object in the scene. But, as an embodied being, human can resort to contextual information to discriminate and perceive the object. Therefore, we explore action-oriented representations in the form of embodied features that capture bodily sensations emerging in the task. Furthermore, we examine the redundancies and the statistical regularities induced in the sensory-motor coordination for obtaining a more efficient perceptive processing. We describe a behavior scheme that includes a proposal for resources organization, and, through a case study, we consider the information fusion within a Bayesian network in charge of discriminating the object.

This document is organized as follows. In Section 2 some related contributions and challenges encountered from the cognitivist approach are discussed. Section 3 starts by the conceptualization of our research within the literature of embodied cognition, it then proceeds with the methodological analysis of the task where the behavior scheme is proposed and described. A case-study has been designed including simulations and experiments with the robot Nao, which is presented in Section 4. The results obtained are discussed in Section 5 along with the research perspectives. Finally, the conclusions of the study are given in Section 6.