Objects Relocation in Clutter with Robot Manipulators via Tree-based Q-Learning Algorithm: Analysis and Experiments

This work addresses the problem of retrieving a target object from cluttered environment using a robot manipulator. In the details, the proposed solution relies on a Task and Motion Planning approach based on a two-level architecture: the high-level is a Task Planner aimed at finding the optimal objects sequence to relocate, according to a metric based on the objects weight; the low-level is a Motion Planner in charge of planning the end-effector path for reaching the specific objects taking into account the robot physical constraints. The high-level task planner is a Reinforcement Learning agent, trained using the information coming from the low-level Motion Planner. In this work we consider the Q-Tree algorithm, which is based on a dynamic tree structure inspired by the Q-learning technique. Three different RL-policies with two kinds of tree exploration techniques (Breadth and Depth) are compared in simulation scenarios with different complexity. Moreover, the proposed learning methods are experimentally validated in a real scenario by adopting a KINOVA Jaco2 7-DoFs robot manipulator.