Curriculum-Based Deep Reinforcement Learning for Autonomous Driving: A Proximal Policy Optimization Approach in CARLA Simulation

Engineering notes

Key topics: autonomous driving system, autonomous driving, 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

This study goes into more detail about the application of deep reinforcement learning for training autonomous agents in exploring the environment. The PPO algorithm in addition to curriculum learning has been used to train the driving agent using the CARLA simulator. The curriculum learning method structures the learning procedure in a manner in which complex problems are reduced to sub-problems, making it easier for the agent to initially master simpler driving scenarios before moving onto the complex scenarios. In our implementation, the multi-modal observation from the environment has been processed by a unique neural network design utilizing the ShuffleNet v2 from the visual modality, along Gated Recurrent Units (GRUs) for temporal integration. Incorporation with curriculum learning resulted in successful implementation in several towns and under varying conditions to prove the effectiveness over traditional reinforcement learning techniques. Further advancement is made to apply structured learning techniques to create robust autonomous driving systems with capability.

Links and sources

• Official / arXiv page

Comments