Deep reinforcement learning for quadrotor path following with adaptive velocity

This paper proposes a solution for the path following problem of a quadrotor vehicle based on deep reinforcement learning theory. Three different approaches implementing the Deep Deterministic Policy Gradient algorithm are presented. Each approach emerges as an improved version of the preceding one....

ver descrição completa

Detalhes bibliográficos
Autores: Rubí Perelló, Bartomeu|||0000-0002-8822-2681, Morcego Seix, Bernardo|||0000-0002-6944-7519, Pérez Magrané, Ramon|||0000-0002-9216-4234
Formato: artículo
Fecha de publicación:2020
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/334076
Acesso em linha:https://hdl.handle.net/2117/334076
https://dx.doi.org/10.1007/s10514-020-09951-8
Access Level:acceso abierto
Palavra-chave:Machine learning
Drone aircraft
Unmanned aerial vehicles
Trajectory control
Path following
Deep reinforcement learning
Deep deterministic policy gradient
Quadrotor
Aprenentatge automàtic
Avions no tripulats
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Descrição
Resumo:This paper proposes a solution for the path following problem of a quadrotor vehicle based on deep reinforcement learning theory. Three different approaches implementing the Deep Deterministic Policy Gradient algorithm are presented. Each approach emerges as an improved version of the preceding one. The first approach uses only instantaneous information of the path for solving the problem. The second approach includes a structure that allows the agent to anticipate to the curves. The third agent is capable to compute the optimal velocity according to the path’s shape. A training framework that combines the tensorflow-python environment with Gazebo-ROS using the RotorS simulator is built. The three agents are tested in RotorS and experimentally with the Asctec Hummingbird quadrotor. Experimental results prove the validity of the agents, which are able to achieve a generalized solution for the path following problem.