Feature Selection for Learning Versatile Manipulation Skills based on Observed and Desired Trajectories

Abstract

For a manipulation skill to be applicable to a wide range of scenarios, it must generalize between different objects and object configurations. Robots should therefore learn skills that adapt to features describing the objects being manipulated. Most of these object features will however be irrelevant for generalizing the skill and, hence, the robot should select a small set of relevant features for adapting the skill.

We use a framework for learning versatile manipulation skills that adapt to a sparse set of object features. Skills are initially learned from demonstrations and subsequently improved using reinforcement learning. The robot also learns a meta prior over the features’ relevances to guide the feature selection process. In this paper, we explore using either desired trajectories or observed trajectories for selecting the relevant features. The framework was evaluated on placing, tilting, and wiping tasks. The evaluations showed that using the desired trajectories to select the relevant features lead to better skill learning performance.