Discovering and Manipulating Affordances

Omar Ricardo Chavez-Garcia (1, 2) , Mihai Andries (1, 2) , Pierre Luce-Vayrac (1, 2) , Raja Chatila (3, 1, 2)

1 ISIR - Institut des Systèmes Intelligents et de Robotique
2 AMAC
3 CNRS - Centre National de la Recherche Scientifique

Reasoning jointly on perception and action requires to interpret the scene in terms of the agent's own potential capabilities. We propose a Bayesian architecture for learning sensorimotor representations from the interaction between perception, action, and salient changes generated by robot actions. This connects these three elements in a common representation: affordances. In this paper, we are working towards a richer representation and formalization of affordances. Current experimental analysis shows the qualitative and quantitative aspects of affor-dances. In addition, our formalization motivates several experiments for exploring hypothetical operations between learned affordances. In particular , we infer affordances of composite objects, based on prior knowledge on the affordances of the elementary objects. The grounding of robotic knowledge [19] is the problem of creating links between the entities and events in the observable environment and their symbolic representations employed by a robot's reasoning algorithms. Solving this problem would allow robots to autonomously discover their environment, without the need of human intervention. Symbolic grounding cannot be achieved by a process of observation alone, and requires interaction between the agent and its environment. In this paper, we study, develop, and experimentally evaluate sensorimotor representations and scene interpretation processes based on visual and propri-oceptive inputs when the robot physically interacts with objects. This enables robots to understand their environment by interacting with it. Our architecture builds models of objects based on perceptual clues and effects of robot actions on them, which relate to the notion of affordance. We employ a Bayesian network that represents with continuous and discrete variables the objects, actions, and effects in the observable environment. We then perform structure learning to identify the most probable Bayesian network that best fits with the observed data. The discovered structure of the Bayesian network allows the robot to discover causal relationships in the environment using statistical data. The remainder of the paper is structured as follows. In Section 2 we discuss related work on the discovery of object affordances, and we introduce our specific

Mots clés: affordances; sensorimotor representations; developmental robotics 1 Motivation; Problem Statement

Domaines: Robotique [cs.RO]

Omar Ricardo Chavez-Garcia, Mihai Andries, Pierre Luce-Vayrac, Raja Chatila. Discovering and Manipulating Affordances. The 2016 International Symposium on Experimental Robotics (ISER 2016), Oct 2016, Tokyo, Japan. ⟨hal-01392826⟩