Adaptive Non-Maximal Suppression filtering for online exploration learning with Cost-Regularized Kernel Regression

In many robotics tasks involving impacts (e.g. grasping, hitting, kicking) the existence of complex interactions between the physics of the objects and of the robot makes it hard to create an analytical model of the interactions that can be used for prediction and planning. Exploration learning can enable a robot to autonomously learn such tasks and models simultaneously by trial and error. The Cost-Regularized Kernel Regression (CrKR) algorithm has been used successfully for learning robot skills such as table tennis, ball-in-a-cup, and dart-throwing. However, despite its effectiveness for fast exploration learning with few data samples, a matrix inversion is required for each new trial. This matrix inversion makes the method computationally expensive so that exploration must finish before the task is fully learned. We present a novel method that can learn continually, by selecting the best information for learning at each step. Using an Adaptive Non-Maximal Suppression (ANMS) filter, we select a sparse subset of points for the regression for each new trial. This enables the robot to continue acquiring samples without the need to discard potentially informative trials. We compare our algorithm with the original CrKR regression, showing a good trade-off between the performance and efficiency of the exploration.