Reinforcement Learning-Based Adaptive Control for Self-Driving Cars in Urban Traffic

Engineering notes

Key topics: autonomous driving, self-driving car, self-driving, autonomous vehicle, reinforcement learning, control. 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 autonomous vehicles which are driven in heavy traffic situations in cities require adaptive control systems capable of addressing dynamic, uncertain and non-linear environments. Offline-trained or rule-based reinforcement learning (RL) models cannot be generalized in real-time by all since they are grounded on fixed datasets and are not flexible. The problem presented to the reader of this paper is a framework of Reinforcement Learning Adaptive Control (RT-ORLAC) which relies on an on-policy actor-critic framework and a Model Predictive Controller (MPC) at the same time with a Control Barrier Function (CBF)-based safety layer. The system as well learns continuously, through utilising live sensor data thereby real time decision and no previous experience is possible. An urban rate simulated on RT-ORLAC demonstrates that the collision rate, the response time, and the control of the RT-ORLAC are also 32, 25, and 18 percent smoother than the PID and offline PPO controllers. The proposed framework will be the foundation of safe adaptive, and computationally efficient autonomous driving in the dynamic urban environment.

Links and sources

Comments