Safety analysis and deep reinforcement learning implementation for an automatic emergency braking system

Abstract

In advanced driver assistance systems, the Automatic Emergency Braking System (AEBS) plays a crucial role in reducing the risk of collisions. However, perception uncertainties, dynamic traffic interactions, and the risk that learning controllers adopt unexpected yet statistically rewarding strategies make the safety justification of AEBS more challenging. This thesis presents a DRL-based AEBS development and validation approach where safety analysis outputs are systematically displayed, controller design is performed using a deep reinforcement learning approach, and traceability is established between the problem types predicted in safety analyses and the observed behaviors during the design/training/testing processes. In this context, the study aims not only to achieve a high-performance braking policy but also to make visible the engineering steps that will support a learning controller to remain within safety-acceptable behavior boundaries. First, using System-Theoretical Process Analysis and Cause Tree Analysis, hazards, unsafe control actions, and causal scenarios were systematically identified in accordance with the ISO/PAS 8800 guidance. Subsequently, the safety measures derived from this analysis were evaluated in DRL design decisions, such as defining the action space, designing the reward function, determining the scope of scenarios to be used in training, and configuring operating limits, establishing a traceable link between safety outputs and the learning controller. The proposed approach was implemented in a MATLAB/Simulink-based scenario-oriented simulation infrastructure; rule-based emergency braking logic was replaced with braking commands generated by the DRL agent. Thus, a comparable experimental ground was created by changing only the decision-making mechanism while preserving the same perception and vehicle dynamics layers. DQN, representing the value-based discrete action approach, and TD3 algorithms, suitable for continuous control, were trained and tested comparatively under a common scenario set and consistent evaluation criteria. The results obtained show that, despite being able to achieve high reward values during the training process, the DQN-based approach converged to an undesirable strategy in the tests, such as early and hard braking, independent of the context. This indicates that deviations may occur between the safety-comfort balance targeted through reward components in training and the actual test behavior. In contrast, the TD3-based controller demonstrated a more successful establishment of the safety-comfort balance by exhibiting braking behaviors that were more sensitive to scenario dynamics, gradual according to the development of risk, and more consistent. In conclusion, this thesis demonstrates that STPA/CTA-based safety analysis can be directly integrated into the design of a learning AEBS controller, that the traceability approach is beneficial in explaining problematic behaviors that may arise in the learning controller, and that the choice of algorithm significantly affects the quality of test behavior.