Citation Context

...g classifier (Pred) is trained for each action. Specifically, we use Support Vector Machine (SVM) classifiers with Radial Basis Function (RBF) kernel and optimized parameters to learn these predictors=-=[12]-=-. The multi-category classifier, after training, can predict the effect category given features and affordances as follows: εoai = Predai(features o, affordanceso\εoai) Here, affordances\εoai denotes ...

Citation Context

...ion accuracy of the corresponding action. The robot keeps learning progress of each action and in each time-step, it selects an action to explore based on the learning progress using ǫ-greedy strategy=-=[14]-=- where ǫ is set to 0.05. If an action (ai) and a number of objects are selected for exploration at time-step t, the robot first computes the effects predicted to be achieved on these objects using Pre...

Citation Context

...ots, enables efficient and effective learning in such environments by guiding the robot learning with intelligent exploration strategies[6]. Intrinsic Motivation (IM) approach in developmental robots =-=[7]-=- was inspired from curiosity based motivation mechanisms in human development, and has recently been effectively applied to cognitive robots where object knowledge is developed through selfexploration...

Citation Context

...d Sequentialfs (sequential features selection) method to select these features. The Sequentialfs method generates near-optimal relevant feature sets in a way similar to the one used in Schemata Search=-=[13]-=-. Starting from an empty relevance feature set, it selects one feature and adds it to the feature set of previous iteration. At each iteration, a candidate feature set for each not-yet-selected featur...

Citation Context

...ability in capturing object-action-effect dynamics, and in making predictions in different directions, for example in inferring the required actions to achieve desired effects given object properties =-=[22]-=-, [23]. We discuss that our ‘discriminative’ model still provides powerful mechanisms as it can effectively map the continuous object feature and behavior parameter spaces to the corresponding effects...

Citation Context

... Intrinsic Motivation mechanisms and distinctive feature selection approach. I. INTRODUCTION Human infants develop a set of action primitives such as grasping, hitting and dropping for single objects =-=[1]-=- by 9 months, and start inserting rods into circular holes in a box or stack up blocks into towers of two blocks from 13 months[2]. While in 13 months, the infants can only insert circular rodes into ...

Citation Context

... 1(a) with no assumption on the relative complexity of actions and predictions. In each learning step, the robot actively selects the most “interesting” action to explore based on Intrinsic Motivation=-=[9]-=-, and updates the prediction model of the corresponding action based on the observed effect. The robot also distinguishes “the most distinctive features” for prediction of each different affordance in...

Citation Context

... idea in [20]. Thus, in this paper we focused on the development of affordance learning structure assuming a number of developed sensorimotor structures. In this paper, the robot forms forward models =-=[21]-=- that enable it to predict the changes in the environment in terms of discrete effect categories. Recently generative models have been proved to be effective in their ability in capturing object-actio...

Citation Context

...epresentations from smaller pieces, nor the association of a response to a stimulus. Instead, she claimed, learning is “discovering distinctive features and invariant properties of things and events” =-=[4]-=-. Learning is not “enriching the input” but discovering the critical perceptual information in that input. We will argue that learning and prediction based on the most distinctive features not only pr...

Citation Context

...he effect predictions for all objects for the selected action 11: end for A. Learning of affordances Learning of affordances corresponds to learning the relations between objects, actions and effects =-=[11]-=-. In this study, object affordances are encoded as the list of effects achievable by executing different actions of the robot: affordanceso = (εoa1 , ε o a2 , ...) where εoa1 is the discrete effect cr...

Citation Context

...ctEffect(all-objects, Pred[action]) 12: ExpObjs += objs 13: end for A. Learning of affordances Learning of affordances corresponds to learning the relations between objects, actions and effects [14], =-=[15]-=-. In this study, object affordances are encoded as the list of effects achievable by executing different actions of the robot: affordanceso = (εoa1 , ε o a2 , ...) where εoa1 is the discrete effect cr...

Citation Context

...d the development order emerged from the learning dynamics that is guided by Intrinsic Motivation mechanisms and distinctive feature selection approach. I. INTRODUCTION Studies with infant chimpanzees=-=[1]-=- and human infants[2] revealed that there is a dramatic increase in exploration and success of object-object combinatory actions at around 1.5 years of age while such actions were at a very low freque...

Citation Context

... input” but discovering the critical perceptual information in that input. We will argue that learning and prediction based on the most distinctive features not only provide perceptual economy (as in =-=[5]-=-), but can be used to autonomously determine the structure of the learning problem. 4th International Conference on Development and Learning and on Epigenetic Robotics October 13-16, 2014. Palazzo Duc...

Citation Context

...y in capturing object-action-effect dynamics, and in making predictions in different directions, for example in inferring the required actions to achieve desired effects given object properties [22], =-=[23]-=-. We discuss that our ‘discriminative’ model still provides powerful mechanisms as it can effectively map the continuous object feature and behavior parameter spaces to the corresponding effects [24] ...

Citation Context

...sed on the observed effect. The robot also distinguishes “the most distinctive features” for prediction of each different affordance in order to “discover the information that specifies an affordance”=-=[10]-=- in training the prediction model. We expect these two mechanisms, namely (i) the Intrinsic Motivation based selection of actions to explore, and (ii) the use of the most distinctive features in affor...

Citation Context

... assumptions in the developmental setting of this paper, as we already showed that a set of basic primitive actions can be self-discovered through in interaction based on observed tactile profiles in =-=[15]-=-, and effect categories can be autonomously found for different actions, such as rolledout-of-table, pushed, no-change, grasped in [11]. ACKNOWLEDGEMENTS This research was supported by European Commun...

Citation Context

...riosity based motivation mechanisms in human development, and has recently been effectively applied to cognitive robots where object knowledge is developed through selfexploration and social guidance =-=[8]-=-. This approach adaptively partitions agent’s sensorimotor space into regions of exploration and guides the agent to select the regions that are in intermediate level of difficulty. This is achieved b...

Citation Context

... [23]. We discuss that our ‘discriminative’ model still provides powerful mechanisms as it can effectively map the continuous object feature and behavior parameter spaces to the corresponding effects =-=[24]-=- without any initial categorization of object properties as in [22], [23]. Furthermore, while bidirectional relations are not explicitly encoded in our system, we showed that our robot was able to mak...

Citation Context

...f age while such combinatory actions were at a very low frequency before that period for several months [4]. Such development patterns suggest non-gradual development changes in human infants as well =-=[5]-=-. This data suggests that the infants first develop basic skills (such as fine grasping and object rotation) that are precursors of combinatory manipulation actions. They also probably use the learned...

Citation Context

...ning complex action affordances, i.e. affordances that are provided by pairs of objects, we proposed a learning framework where a developmental robotic system learns object affordances1 in two-stages =-=[3]-=-. In the first stage, the robot learns predicting single-object affordances (such as pushability and rollability) by pushing single objects in different directions, and learning the relations between ...

Citation Context

... primitives such as grasping, hitting and dropping for single objects [1] by 9 months, and start inserting rods into circular holes in a box or stack up blocks into towers of two blocks from 13 months=-=[2]-=-. While in 13 months, the infants can only insert circular rodes into circular holes in a plate, by 18 months they can perceive the correspondence between different shaped blocks and they start insert...

Citation Context

...14], and (iii) learn complex actions such as bring object 1 over object 2 through imitation by sequencing the discovered behavior primitives and learned affordances using parental scaffolding idea in =-=[20]-=-. Thus, in this paper we focused on the development of affordance learning structure assuming a number of developed sensorimotor structures. In this paper, the robot forms forward models [21] that ena...

Citation Context

...er emerged from the learning dynamics that is guided by Intrinsic Motivation mechanisms and distinctive feature selection approach. I. INTRODUCTION Studies with infant chimpanzees[1] and human infants=-=[2]-=- revealed that there is a dramatic increase in exploration and success of object-object combinatory actions at around 1.5 years of age while such actions were at a very low frequency before that perio...

Citation Context

...ded as a set of active learning mechanisms for developmental robots, enables efficient and effective learning in such environments by guiding the robot learning with intelligent exploration strategies=-=[6]-=-. Intrinsic Motivation (IM) approach in developmental robots [7] was inspired from curiosity based motivation mechanisms in human development, and has recently been effectively applied to cognitive ro...