Programming by demonstration (PbD) allows matching the kinematic movements of a robot with those of a human. The presented Continuous Goal-Directed Actions (CGDA) is able to additionally encode the effects of a demonstrated action, which are not encoded in PbD. CGDA allows generalization, recognition and execution of action effects on the environment. In addition to analyzing kinematic parameters (joint positions/velocities, etc.), CGDA focuses on changes produced on the object due to an action (spatial, color, shape, etc.). By tracking object features during action execution, we create a trajectory in an n-dimensional feature space that represents object temporal states. Discretized action repetitions provide us with a cloud of points. Action generalization is accomplished by extracting the average point of each sequential temporal interval of the point cloud. These points are interpolated using Radial Basis Functions, obtaining a generalized multidimensional object feature trajectory. Action recognition is performed by comparing the trajectory of a query sample with the generalizations. The trajectories discrepancy score is obtained by using Dynamic Time Warping (DTW). Robot joint trajectories for execution are computed in a simulator through evolutionary computation. Object features are extracted from sensors, and each evolutionary individual fitness is measured using DTW, comparing the simulated action with the generalization.

In this paper, we study how human-robot interaction can be beneficial on the Continuous Goal-Directed Actions (CGDA) framework. Specifically, a system for robot discovery of motor primitives from random human-guided movements has been developed. These guided motor primitives (GMP) are used as scaffolds to reproduce a goal-directed actions. CGDA encodes goals as the changes produced on object features (color, area, etc) due to actions. This paper focuses on using motor primitives extracted from human-guided random robot movements to execute these goal-directed actions. The human guides the robot joints in random movements, which are later divided in small segments. These segments are compared in terms of joint positions and selected to be diverse. To perform goal-directed actions, the robot must discover an adequate sequence of GMP. To discover these sequences we organize the primitives as a tree with incremental depths (where each node represents a primitive) and use a breadth-first search. In one of the experiments performed, the robot executes a task based on spatial object features. In the other experiment, the goal is to paint a wall by following a color feature trajectory.

This paper presents a robot imagination system that generates models of objects prior to their perception. This is achieved through a feature inference algorithm that enables computing the fusion of keywords which have never been presented to the robot together previously. In this sense, robot imagination is defined as the robot's capability of generating feature parameter values of unknown objects by generalizing characteristics from previously presented objects. The system is first trained with visual information paired with semantic object descriptions from which keywords are extracted. Each keyword creates an instance of the learnt object in an n-dimensional feature space. The core concept behind the robot imagination system presented in this paper is the use of statistically fit hyperplanes in the feature space to represent and simultaneously extend the meaning of grounded words. The inference algorithm allows to determine complete solutions in the feature space. Finally, evolutionary algorithms are used to return these numeric values to the real world, completing an inverse semantic process.