Autonomous Driving Simulation Using Reinforcement Learning in CARLA

Engineering notes

Key topics: autonomous driving, semantic segmentation, reinforcement learning, carla. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The objective of this project is to build up sophisticated autonomous driving simulation based on CARLA platform and reinforcement learning. The primary objective is to train a robust and reliable RL model for navigating complex urban scenes behaving in ways that are adherent with traffic rules as well as safe interactions with dynamic entities (e.g., vehicles, pedestrians). The implemented system relies on a custom Gym environment, connected to CARLA’s Python API, where we have RGB and semantic segmentation cameras as the main sources of input. Proximal Policy Optimization (PPO) model is used as decision-maker so that the vehicle has high-precision in lane keeping, traffic light obedience and obstacle avoidance. In order to address the typical RL problems (e.g., sparse reward, poor generalization), we follow curricular learning and reward shaping approaches. Instruction is by that stage, where training begins on the simpler roads and progresses to streets with much more traffic. It is designed to be computable on low-end hardware platforms without sacrificing the performance and it can be used by research labs which have a limited budget. Quantitative evaluations including collision rate, rule compliance and average episode rewards show the effectiveness of the proposed method.

Links and sources

• Official / arXiv page

Comments