Português English

Dissertação de Mestrado de Luíza Caetano Garaffa

Detalhes do Evento

Aluna: Luíza Caetano Garaffa
Orientador: Prof. Dr. Edison Pignaton de Freitas

Título: Evaluation of mapless navigation for unknown indoor environment exploration by single and multiple autonomous mobile robots using Deep Reinforcement Learning
Linha de Pesquisa: Planejamento, Sistemas Multiagentes e Robótica.

Data: 13/12/2022
Horário: 13h30min
Local: Esta banca ocorrerá em formato remoto através do link:  https://mconf.ufrgs.br/webconf/00138178

Banca Examinadora:
– Prof. Dr. Alexandre César Muniz de Oliveira (UFMA)
– Profa. Dra. Mariana Luderitz Kolberg Fernandes (UFRGS)
– Prof. Dr. Renan de Queiroz Maffei (UFRGS)

Presidente da Banca: Prof. Dr. Edison Pignaton de Freitas

Abstract: Efficient exploration of unknown environments is a fundamental precondition for modern autonomous mobile robot applications. The traditional exploration approach consists in sensing the world to build a map, and using the generated map to decide where to go next. In the specific case of collaborative exploration by multi-robot systems, map sharing and merging is often employed. Such methods tend to result in high computational costs, which can restrict their application in scenarios with limited memory and processing resources. An alternative to this is to employ mapless navigation when performing exploration. However, defining the functioning of an appropriate resilient exploration strategy is not a straightforward task, especially in a mapless fashion. Concurrently, the employment of Deep Reinforcement Learning has enabled optimal or near-optimal solutions for several complex problems with high-dimensional inputs. However, to the best of our knowledge, there are no works that investigate the application of DRL solutions for mapless exploration aiming at efficient area coverage, without pre-determined goal positions. In that context, this dissertation reviews recent research works that use RL to design unknown environment exploration strategies for single and multi-robots. Based on the gathered information, we propose an end-to-end mapless exploration framework based in DRL and suitable for single robots and teams of n robots. The exploration policy is trained and tested in different simulation environments, and the results are compared with other exploration methods. Although less efficient than methods that use map generation and DRL, our method enabled cooperation between agents in a mapless fashion, increasing the explored region rate with smaller path lengths per individual robot.

Keywords: Mobile robotics. Unknown environment exploration. Deep Reinforcement Learning. Single robot exploration. Cooperative exploration. Multi-Robot Systems.