Autonomous robot car with Deep Reinforcement Learning

Guangli Chen

AutorIn

Guangli Chen

Titel

Autonomous robot car with Deep Reinforcement Learning

Zitierfähige Url:

https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-945829

Erstveröffentlichung

2024

Datum der Einreichung

29.10.2024

Datum der Verteidigung

27.11.2024

Abstract (EN)

With the rapid development of Deep Reinforcement Learning (DRL), it has been widely applied in the field of autonomous driving (AD). This thesis presents an end-to-end approach based on the Proximal Policy Optimization (PPO) algorithm to achieve multi-tasks AD, including lane keeping and obstacle avoidance. To address the challenge of multi-tasks AD, a two-stage reinforcement learning (RL) strategy is adopted for both lane keeping and obstacle avoidance tasks, and its performance is compared with a RL baseline. Reward function is critical in RL. In this study, a suitable reward function is first identified through a non-end-to-end approach, as the non-end-to-end method converges faster. This reward function is then used to train the end-to-end lane keeping task. For obstacle avoidance, we propose three reward functions based on different principles. Each reward function is evaluated using both two-stage RL and normal RL, and the final comparison focuses on the obstacle avoidance performance of the three reward functions using the two-stage RL approach.

Freie Schlagwörter (EN)

Deep Reinforcement Learning, Autonomous driving, Proximal Policy Optimization, Lane keeping, Obstacle avoidance.

Klassifikation (DDC)

380

Klassifikation (RVK)

ZO 4660

GutachterIn