A novel event-triggered control framework is proposed in this paper to realize safe reinforcement learning (SRL) for autonomous vehicle (AV) control. Safety is guaranteed by designing an additional safe controller to correct the unsafe actions proposed by the deep reinforcement learning (DRL) agent. Event-triggered control barrier functions (CBFs) are used to impose safety constraints on the actions in a discrete manner. Based on twin delayed deep deterministic policy gradient (TD3), an event-triggered safe TD3 (ET-STD3) is presented for safe AV control. Experiments are conducted to train and validate the proposed ET-STD3 in a simulated car-following scenario. Both RL-based and model-based baselines are also tested in the same scenario for comparison. Results show that ET-STD3 achieves better control and safety performance than other involved baselines at the cost of comparable triggering times with the event-triggered baseline.
Event-Triggered Control for Automated Vehicles Based on Safe Reinforcement Learning
24.09.2024
1372702 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Safe Reinforcement Learning for Automated Vehicles via Online Reachability Analysis
IEEE | 2024
|