On Path Regression with Extreme Learning and the Linear Configuration Space

This paper studies the path regression problem, that is learning motion planning functions that render trajectories from initial to end robot configurations in a single forward pass. To this end, we have studied the path regression problem using the linear transition in the configuration space and shallow neural schemes based on Extreme Learning Machines. Our computational experiments involving a relevant and diverse set of 6-DOF robot trajectories have shown path regression's feasibility and practical efficiency with attractive generalization performance in out-of-sample observations. In particular, we show that it is possible to learn neural policies for path regression in about 10 ms. - 31 ms. and achieving 10-3 - 10-6 Mean Squared Error on unseen out-of-sample scenarios. We believe our approach has the potential to explore efficient algorithms for learning-based motion planning.