Humanoid robot imitation through continuous goal-directed actions: an evolutionary approach

Abstract

Humanoids can learn motor skills through the programming by demonstration framework, which allows matching the kinematic movements of a robot with those of a human. Continuous goal-directed actions (CGDA) is a framework that can complement the paradigm of robot imitation. Instead of kinematic parameters, its encoding is centered on the changes an action produces on object features. The features can be any measurable characteristic of the object such as color, area, etc. The execution of actions encoded as CGDA allows a robot-configuration independent achievement of tasks, avoiding the correspondence problem. By tracking object features during action execution, we create a trajectory in an n-dimensional feature space that represents object temporal states, allowing generalization, recognition, and execution of action effects on the environment. Experiments have been performed, using a humanoid robot in a simulated environment. Evolutionary computation was used for joint parameter calculation of a humanoid robot. The objective is to generate a motor trajectory which leads to a feature trajectory equal to the objective one. In one of the experiments, the robot performs a spatial trajectory based on spatial object features. In a new experiment, the robot paints a wall by following a color feature encoding.

Real Evaluations Tractability using Continuous Goal-Directed Actions in Smart City Applications

“…These methods and constraints were integrated and open-sourced (Source code available at ) within the CGDA architecture. For all the experiments, the execution of the “wax” (known as “clean” in previous literature) and the “paint” actions [ 8 ] were performed, choosing IET as the evolutionary strategy. The “wax” action is a representative example of a geometrical action, with the particularity that intermediate positions are also relevant for action execution.…”

“…This interpolation is performed within the set of intermediate goals of fixed duration. A generalization example of the “wax” action is depicted in Figure 2 extracted from reference [ 8 ], where the scalar features are Cartesian coordinates that encode a geometrical action.…”

“…However, for the general case of not only geometrical features, e.g., the percentage of object or environment that has been painted, this is a not valid solution. To deal with this, several execution strategies based on Evolutionary Algorithms (EA) have been previously studied by the authors [ 8 ]. These strategies search for the optimal execution trajectory in the joint motor parameter space of the robot, using the recognition discrepancy described above as the fitness for evaluation.…”

Robot Imitation Through Vision, Kinesthetic and Force Features with Online Adaptation to Changing Environments

“…The "paint" action is a representative use case presented in previous work of the authors [8]. While in previous work the generalized "paint" action was generated synthetically as a linear growth from 0% to 100% of the painted portion of a tracked object (a wall), this feature trajectory was now generated from 4 user demonstrations.…”

“…1: procedure FTE(X) 2: individuals ←initialize 3: while not termination conditions do 4: for each individual do for j < n do 4: while not termination conditions do 5: for each individual do for j < n do 4: while not termination conditions do 5: for each individual do end for 13: motor execution(U ) 14: end procedure Experimental evidence from previous publications has determined FTE to be the strategy that requires most evaluations for fitness convergence [8]. The main intuition behind this large amount of required evaluations is that evolutionary algorithms are greatly affected by the size of the search space.…”