Egocentric Representations for Autonomous Navigation of Humanoid Robots
Hendry Ferreira Chame · HAL (Le Centre pour la Communication Scientifique Directe) · 2016
The skill of visually approaching and positioning in relation to objects on the scene is of crucial importance for service robotics applications. Furthermore, the autonomy of the solution is essential, since human-centered scenarios, where these robots are expected to operate (e.g. at the office or home), are stochastic. Hence, it is important that the agent can react to unforeseen situations. The traditional approach of AI has not produced reliable results since it is based on extensive context-free models of the tasks, so action selection is a centralized and delayed process. Emergent models have in contrast produced fast response systems at the cost of poor generalization power, even to very similar scenarios. This research has taken an intermediate perspective between the cognitivist and the EC research. It employs simultaneously action-independent knowledge for visually recognizing the stimuli of interest, and local representations in the form of bodily sensations, in order to anticipate the consequences of action, to discriminate the object, to react to unexpected circumstances, and to assess the progress and success of the mission.