Representation and Experience-Based Learning of Explainable Models for Robot Action Execution

Ontology-Assisted Generalisation of Robot Action Execution Knowledge

“…For both actions, we learn dedicated execution models (using guided learning as in [8]) for each of the following objects in order to have a reasonable variety of models: apple, chips can, sugar box, mug, and tennis ball. Each model is learned using 25 executions of the action; the remaining objects are used for testing the generalisation.…”

“…To find the planar orientation of the object with respect to the robot, we use a RANSAC-like procedure [27] that finds the direction of 2D lines fitted to randomly selected points from the object's point cloud; the average of these line orientations is taken to represent the object's orientation. 18 The relational model of the grasping action is extracted from the following relations (similar to [8]):…”

“…This work builds upon [8], where a representation of action execution models was proposed based on which a robot jointly learns a relational model of success preconditions, or a collection of such models for multiple qualitative modes, as well as a continuous model that maps action parameters and execution constraints to predicted execution success. An execution model is learned from labelled execution data, such that the relations used for representing the relational model are defined per action, and the continuous model is represented by a Gaussian process so that prediction uncertainty can be encoded.…”

“…where depth : O → N is the depth of an object class in the hierarchy induced by the ontology and LCS : O × O → O is the least common subsumer of o and õ, namely the most specific class that is a common ancestor of o and õ. 8 The WUP similarity corresponds well to G as it considers (i) siblings to be more similar to each other than nodes among different levels of the hierarchy, which puts an intuitive prior preference on generalisation between sibling classes and (ii) classes with higher depths in the hierarchy to have higher similarity to each other than classes near the top of the hierarchy, which encodes the principle that class similarity increases with the specificity of the ontology model. c) Model selection: Given the posterior values P t+1 (õ|o, S) for each õ ∈ C o , an execution model M o * is selected for the object class o * that maximises the posterior:…”

“…We show that such a representation can be useful to determine when already learned models can be reused and when additional learning is needed. We analyse the method with learned execution models of parameterised skills as in [8] using a Toyota HSR [9] manipulating objects in two common domestic scenarios -grasping everyday objects and stowing them in a drawer. The results show that a robot can often reuse knowledge about objects it has manipulated before when dealing with new object classes, but also that feedback about failures can allow it to adapt its generalisation accordingly.…”