Efficient deep reinforcement learning in robotic motion planning

Research output: ThesisThesis fully internal (DIV)

217 Downloads (Pure)

Abstract

Controlling robot behavior has been realized in the past by means of manual programming of behaviors. This requires too much human programming effort and is typically too specific for one particular task, without possibilities for adaptation. Furthermore, it is difficult to determine mathematical control models for every new robot type. Borrowing from biology, there is a learning paradigm, reinforcement learning, where the robot learns to perform a particular task by trial and error, being rewarded for appropriate actions and punished for making errors. However, reinforcement learning in robot motion planning requires solutions to several challenges: Imbalanced data distribution, multi-objective optimization, the curse of dimensionality, and the 'Sim2Real' gap, i.e., the differences between the simulated and the physical world. In this thesis, we address a number of these challenges and propose solutions. We investigate how curriculum learning, learning with planning, imitation learning, and a hierarchical setup of the learning process could help to improve the Reinforcement Learning (RL) training efficiency and robustness in robotic motion planning tasks. It was found that these methods all played a positive role in improving the sampling efficiency and task performance in the motion planning tasks. Using a guidance mechanism that improves the probability of positive experiences in task exploration appeared to contribute most to improved performance.
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • University of Groningen
Supervisors/Advisors
  • Schomaker, Lambert, Supervisor
  • Mohades Kasaei, Hamidreza, Co-supervisor
Award date12-Mar-2024
Place of Publication[Groningen]
Publisher
DOIs
Publication statusPublished - 2024

Cite this