- title
- Autonomous Navigation with Deep Reinforcement Learning in Duckietown Simulator
- Please use the following URL when quoting:
- https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-968228
- publication_date
- 2025
- Date of submission
- 27.03.2025
- Date of defense
- 23.04.2025
- Abstract (EN)
- With the rise of deep reinforcement learning (DRL), autonomous driving has achieved rapid developments. This thesis investigates the approach of DRL for lane-keeping function and automated navigation in Gym-Duckietown simulator, utilizing convolutional neural network (CNN) and proximal policy optimization (PPO). To address challenges on multi-tasks learn-ing, a three-stage reinforcement learning is adopted for implementing both lane-following and navigation task, In the second and third stage, CNN-weight transfer is employed to iter-ate the weights of the corresponding CNN from the previous stage training in order to com-pare with model performance under conventional training. Experimental results demonstrate a positive influence of weight transfer on training process accelerating and on better lane-keeping performance. In navigation task, A*-algorithm is employed to generate the shortest path from the start point to the destination. The reward function exhibits its critical role in the training, as an suboptimal reward for navigation results in failure of navigation imple-mentation and in interference of lane-keeping ability.
- Keywords (EN)
- Deep Reinforcement Learning, Autonomous driving, Proximal Policy Optimization, Gym-Duckietown, Lane keeping, Autonomous Navigation
- Classification (DDC)
- 380
- Classification (RVK)
- ZO 4660
- Examiner
- Prof. Dr. Ostap Okhrin
- Prof. Dr. Pascal Kerschke
- supervisor
- M.Sc. Dianzhao Li
- Awarding institution
- Technische Universität Dresden, Dresden
- version
- publizierte Version / Verlagsversion
- URN Qucosa
- urn:nbn:de:bsz:14-qucosa2-968228
- Qucosa date of publication
- 06.05.2025
- Document type
- master_thesis
- Document language
- English
- licence
CC BY 4.0