QDTraj: Diverse Trajectory Primitives for Articulated Object Manipulation

A new method called QDTraj enables robots to manipulate a wide range of articulated objects by automatically generating diverse low-level trajectory primitives. The approach uses Quality-Diversity algorithms and sparse reward exploration to create multiple solutions for the same manipulation task, allowing robots to adapt to live constraints and unexpected changes in real-world environments. This addresses the challenge of autonomous manipulation in open-ended domestic settings, where robots currently struggle despite advances in learning and robotics.