Navigation Among Movable Obstacles with Mobile Manipulator using Learned Robot-Obstacle Interaction Model

Taegeun Yang, Minsung Yoon, Jeil Jeong, Sung-eui Yoon

Keywords: Navigation Among Movable Obstacles (NAMO), Mobile Manipulator, RL based MPPI

TL;DR: This paper addresses the task of creating a drivable path by pushing obstacles out of the way using a mobile manipulator in a situation where obstacles are blocking the driving path.

Abstract: In this paper, we address the online Navigation Among Movable Obstacles (NAMO) problem by employing a mobile manipulator. Unlike mobile robots, mobile manipulators offer the advantage of effectively relocating obstacles out of the driving path while tracking a global path. However, the high degrees of freedom (DOF) of mobile manipulator complicates whole-body control. To address these challenges, we propose a Reinforcement Learning (RL) based Model Predictive Path Integral (MPPI) framework. This strategy includes identifying actions for stable pushing through RL, training robot-obstacle kinodynamic interaction model from policy-generated data, and applying this model in MPPI to maneuver obstacles while tracking the global path. In our experiment, we demonstrated that our method successfully pushes obstacles aside and maintains adherence to the global path when it is obstructed.

Submission Number: 18