A ROS-Based Modular End-to-End Architecture: Building and Validating a Safe and Reliable Autonomous Driving Stack.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Classical modular autonomous driving pipelines lack adaptability and robust scene interpretation in complex urban environments.
• Current systems struggle with novel scenarios and highly interactive traffic situations.

Purpose of the Study:

• To develop a modular, end-to-end ROS-based autonomous driving architecture that integrates learning-based models into a classical pipeline.
• To enhance safety, reliability, and adaptability in autonomous driving systems for urban environments.

Main Methods:

• Injected learning-based models (GaussianCaR, CLIP) into processing layers for dense semantic Bird's-Eye View (BEV) perception.
• Expanded Hierarchical Petri Net state space for multi-agent reasoning and refined planning with continuous curve optimization.
• Implemented Adaptive Nonlinear Model Predictive Control (ANMPC) for superior trajectory tracking.

Main Results:

• The proposed architecture improved the Driving Score from 53.81% to 66.46% in the CARLA simulator, with a significant increase in Infraction Penalty from 0.59 to 0.79.
• Achieved a Driving Score of 73.9% and an Infraction Penalty of 0.913 against pure end-to-end approaches.
• Demonstrated a shift towards safer, more conservative driving behavior.

Conclusions:

• Modular interpretability and competitive end-to-end performance are achievable in autonomous driving systems.
• The enhanced architecture offers a superior balance of safety, performance, and adaptability in complex urban driving scenarios.
• Future work will involve public code release for broader accessibility and validation.