Robotic Arm Contact Operation Task Using Reinforcement Learning Assistance

EasyChair Preprint 13553

Abstract

Robotic contact manipulations of objects are central elements in many applications. Insuch tasks, a robotic arm applies forces to an object at certain contact points to movethe object to a desired position or along a trajectory. The aim of this work is to deviseand implement an algorithm based on Reinforcement Learning (RL) to move a rectangular objectto a desired configuration along a designed trajectory by employing a three-link robotic arm. Traditional control methods arechallenging to implement due to the complex geometry of the object and the unknowncontact dynamics, particularly friction. Reinforcement Learning overcomes this by eliminating the need forsophisticated control gain tuning, allowing the robot to adapt to randomized scenarios. Residual policy learning \cite{silver2018residual} is utilized as a joint torque augmentation to ensure the end effector has the appropriate contact points and forces throughout the contact operation. The results indicate that residual policy learning significantly improves the accuracy of the object’s translation and rotation during the pushing process.

Keyphrases: Contact dynamics, Reinforcement Learning, Robotics Control

@booklet{EasyChair:13553,
  author    = {Chen Chen and Xu Dai and Jozsef Kovecses},
  title     = {Robotic Arm Contact Operation Task Using Reinforcement Learning Assistance},
  howpublished = {EasyChair Preprint 13553},
  year      = {EasyChair, 2024}}