One of the hallmarks of human intelligence is the ability of predicting the consequences of actions and e ciently plan behaviors based on such predictions. This ability is supported by internal models that human babies acquire incrementally during development through sensorimotor experience: i.e. by interacting with objects in the environment while being exposed to sensori perception. An elegant and powerful concept to represent these internal models has been proposed in developmental psychology under the name of object a ordances: action possibilities that an object o ers to an agent. A ordances are learned ecologically by the agent and exploited for action planning. Clearly, endowing arti cial agents with such cognitive capabilities is a fundamental challenge both in arti cial intelligence and robotics. We propose a learning framework in which an embodied agent (i.e. in our case, the humanoid robot iCub) autonomously explores the environment, and learns object a ordances as probabilistic dependencies between actions, object visual properties and observed e ects; we use Bayesian Networks to encode this probabilistic model. By making inferences across the learned dependencies a number of cognitive skills are enabled: e.g. i) predicting the e ects of an action over an object, or ii) selecting the best action to obtain a desired e ect. By exploring object-object interactions the robot can develop the concept of tool (i.e. a handheld object that allows to obtain a desired e ect on another object), and eventually use the acquired knowledge to plan sequences of actions to attain a desired goal (i.e. problem solving).
Humans solve complex tasks on a routine basis, by choosing, amongst a vast
repertoire, the most proper actions to apply onto objects in order to obtain
certain e ects. According to developmental psychology [
A computational model of a ordances
We follow the framework of [
The a ordances of the intermediate hand-held object (i.e. the tool) can be
incorporated in a more complete model for cognitive tool use. Tools can be typically
described by three functional parts: a handle, an e ector, and a body of a
certain length L connecting the two (see right part of Fig. 2). These three parts
are related to three di erent motor behaviors humans have to perform in order
to successfully use a tool: grasping the handle, reaching for a desired pose with
the e ector and then executing an action over an a ected object. Each of those
behaviors requires some prior mental reasoning, rst to estimate whether the
behavior is feasible (e.g. is the handle graspable?) and then to plan the
correct motion to be executed (e.g. determine the target hand pose and the nger
movements to grasp the tool). We can therefore de ne three levels of tool
affordances: i) usage a ordances, ii) reach a ordances and iii) grasp a ordances
(see right part of Fig. 2). These a ordances relate to speci c problems: i) what
actions the tool a ords, because of its e ector properties, ii) what part of the
workspace the tool a ords to reach for, depending on its length, iii) what grasps
the tool a ords, based on the shape and size of the handle. The outcomes of
these three reasoning processes are based on internal models that the robot can
learn through motor exploration. The model of a ordances in the left part of
Fig. 2 represents the usage a ordances. In previous work we proposed a learning
framework that enables a robot to learn its own body schema [13{15], and to
update it when tools are included [
Since the early days of Arti cial Intelligence (AI), planning techniques have been employed to allow agents to achieve complex tasks in closed and deterministic worlds. Every action has clearly de ned pre-conditions, and generates deterministic post-conditions. However, these assumptions are not plausible if we consider a real robot acting in real unstructured environments, where the consequences of actions are not deterministic and the world is perceived through noisy sensing. The a ordance model (and more generally, the tool use model) depicted in Fig. 2 provide probabilistic predictions of actions consequences, that depend on the perceived visual features of the objects and on the robot sensorimotor abilities and previous experiences. Inspired by recent advances in AI, we can use these predictions within probabilistic planning algorithms, to achieve a grounding of the planning operators based on the robot sensorimotor knowledge. Through this computational machinery, the robot is able to plan the sequence of actions that has the higher probability to achieve the goals, given its own motor abilities and the perceived properties of the available objects.