Map-less Navigation in Novel Environments using Deep Reinforcement Learning: A Curiosity-driven approach for Mobile Robotics
Author
Rasmussen, Anders Frederik Nørby
Term
4. term
Publication year
2019
Submitted on
2019-06-03
Pages
51
Abstract
Dette projekt undersøger, om en autonom robotstyrealgoritme baseret på deep reinforcement learning kan navigere i nye miljøer uden at være afhængig af et forudopbygget kort. Vi opretter en Unity-baseret simulator og anvender domain randomization—variation i udseende, belysning og layout fra episode til episode—for at gøre den lærte adfærd mere robust. For at fremme udforskning af ukendte områder tilføjer vi et Intrinsic Curiosity Module (ICM), som opmuntrer agenten til at opsøge nye situationer. Vi træner to agenter: et Deep Recurrent Q-Network (DRQN) og et DRQN kombineret med ICM, på en forenklet version af miljøet for at vurdere systemets potentiale. Under træningen bruger vi Learning from Easy Missions (vi starter med lettere opgaver og øger sværhedsgraden) samt den foreslåede metode Naive Visual Hindsight Experience Replay. Tilsammen har disse komponenter til formål at opbygge og afprøve et grundlag for kortfri robotnavigation i ukendte omgivelser.
This project explores whether an autonomous robot controller based on deep reinforcement learning can navigate in new environments without relying on a pre-built map. We create a Unity-based simulator and apply domain randomization—varying appearance, lighting, and layouts between training episodes—to make learned behavior more robust. To encourage exploration of unseen areas, we add an Intrinsic Curiosity Module (ICM) that rewards novelty. We train two agents: a Deep Recurrent Q-Network (DRQN) and a DRQN combined with ICM, using a simplified version of the environment to gauge the system’s potential. During training, we use Learning from Easy Missions (starting with simpler tasks and gradually increasing difficulty) and the proposed Naive Visual Hindsight Experience Replay strategy. Together, these components aim to build and assess a foundation for map-free robot navigation in unfamiliar settings.
[This abstract was generated with the help of AI]
Documents