Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Reinforcement learning-based mapless navigation with fail-safe localisation

Lin, Feiqiang, Ji, Ze ORCID: https://orcid.org/0000-0002-8968-9902, Wei, Changyun and Niu, Hanlin 2021. Reinforcement learning-based mapless navigation with fail-safe localisation. Presented at: 21st Towards Autonomous Robotic Systems Conference (TAROS 2021), Virtual and Lincoln, UK, 8-10 September 2021. Published in: Fox, Charles, Gao, Junfeng, Ghalamzan Esfahani, Amir, Saaj, Mini, Hanheide, Marc and Parsons, Simon eds. Towards Autonomous Robotic Systems: 22nd Annual Conference, TAROS 2021, Lincoln, UK, September 8–10, 2021, Proceedings. Lecture Notes in Computer Science; Lecture Notes in Artificial Intelligence Springer, pp. 100-111. 10.1007/978-3-030-89177-0_10

[thumbnail of TAROS_Feiqiang.pdf] PDF - Accepted Post-Print Version
Download (5MB)

Abstract

Mapless navigation is the capability of a robot to navigate without knowing the map. Previous works assume the availability of accurate self-localisation, which is, however, usually unrealistic. In our work, we deploy simultaneous localisation and mapping (SLAM)-based self-localisation for mapless navigation. SLAM performance is prone to the quality of perceived features of the surroundings. This work presents a Reinforcement Learning (RL)-based mapless navigation algorithm, aiming to improve the robustness of robot localisation by encouraging the robot to learn to be aware of the quality of its surrounding features and avoid feature-poor environment, where localisation is less reliable. Particle filter (PF) is deployed for pose estimation in our work, although, in principle, any localisation algorithm should work with this framework. The aim of the work is two-fold: to train a robot to learn 1) to avoid collisions and also 2) to identify paths that optimise PF-based localisation, such that the robot will be unlikely to fail to localise itself, hence fail-safe SLAM. A simulation environment is tested in this work with different maps and randomised training conditions. The trained policy has demonstrated superior performance compared with standard mapless navigation without this optimised policy.

Item Type: Conference or Workshop Item (Paper)
Date Type: Published Online
Status: Published
Schools: Engineering
Publisher: Springer
ISBN: 9783030891763
Date of First Compliant Deposit: 20 July 2021
Date of Acceptance: 1 July 2021
Last Modified: 28 Nov 2022 13:18
URI: https://orca.cardiff.ac.uk/id/eprint/142726

Actions (repository staff only)

Edit Item Edit Item

Downloads

Downloads per month over past year

View more statistics