A stable method for task priority adaptation in quadratic programming via reinforcement learning

In emerging manufacturing facilities, robots must enhance their flexibility. They are expected to perform complex jobs, showing different behaviors on the need, all within unstructured environments, and without requiring reprogramming or setup adjustments. To address this challenge, we introduce the...

Full description

Saved in:

Bibliographic Details
Published in:	Robotics and computer-integrated manufacturing Vol. 91; p. 102857
Main Authors:	Testa, Andrea, Laghi, Marco, Bianco, Edoardo Del, Raiola, Gennaro, Hoffman, Enrico Mingo, Ajoudani, Arash
Format:	Journal Article
Language:	English
Published:	Elsevier Ltd 01.02.2025 Elsevier
Subjects:	Computer Science Machine Learning Machine learning for robot control Optimization and optimal control Reinforcement learning Robotics Machine learning for robot control Optimization and optimal control Reinforcement learning Reinforcement Learning Machine Learning for Robot Control Optimization and Optimal Control Optimization and Optimal Control Reinforcement Learning Machine Learning for Robot Control
ISSN:	0736-5845, 1879-2537
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	In emerging manufacturing facilities, robots must enhance their flexibility. They are expected to perform complex jobs, showing different behaviors on the need, all within unstructured environments, and without requiring reprogramming or setup adjustments. To address this challenge, we introduce the A3CQP, a non-strict hierarchical Quadratic Programming (QP) controller. It seamlessly combines both motion and interaction functionalities, with priorities dynamically and autonomously adapted through a Reinforcement Learning-based adaptation module. This module utilizes the Asynchronous Advantage Actor–Critic algorithm (A3C) to ensure rapid convergence and stable training within continuous action and observation spaces. The experimental validation, involving a collaborative peg-in-hole assembly and the polishing of a wooden plate, demonstrates the effectiveness of the proposed solution in terms of its automatic adaptability, responsiveness, flexibility, and safety. •Attainment of multiple tasks in robot control using Quadratic Programming.•Usage of a Reinforcement Learning strategy for online adaptation of task priorities.•Implementation of the Asynchronous Advantage Actor–Critic algorithm.•Demonstration of the stability of the developed controller.•Validation on a Franka through collaborative peg-in-hole and polishing tasks.
ISSN:	0736-5845 1879-2537
DOI:	10.1016/j.rcim.2024.102857