Autonomous navigation is critical for unlocking the full potential of Unmanned Surface Vehicles (USVs) in complex maritime environments. Deep Reinforcement Learning (DRL) has emerged as a powerful paradigm for developing self-learning control policies, yet the design of reward functions to balance conflicting objectives, particularly fast arrival at the target position and collision avoidance, remains a major challenge. The precise, quantitative impact of reward parameterization on a USV's maneuvering behavior and the inherent performance trade-offs have not been thoroughly investigated. Here we demonstrate that by systematically varying reward function weights within a framework relying on the Proximal Policy Optimization (PPO), it is possible to quantitatively map the trade-off between collision avoidance safety and mission time. Our results, derived from simulations, show that agents trained with balanced reward weights achieve target-reaching success rates exceeding 98\% in dynamic multi-obstacle scenarios. Conversely, configurations that disproportionately penalize obstacle proximity lead to overly cautious behavior and mission failure, with success rates dropping to 22\% due to workspace boundary violations. This work provides a data-driven methodological framework for reward function design and parameter selection in safety-critical robotic applications, moving beyond ad-hoc tuning towards a more structured parameter influence analysis.

Autonomous navigation is critical for unlocking the full potential of Unmanned Surface Vehicles (USVs) in complex maritime environments. Deep Reinforcement Learning (DRL) has emerged as a powerful paradigm for developing self-learning control policies, yet the design of reward functions to balance conflicting objectives, particularly fast arrival at the target position and collision avoidance, remains a major challenge. The precise, quantitative impact of reward parameterization on a USV’s maneuvering behavior and the inherent performance trade-offs have not been thoroughly investigated. Here, we demonstrate that by systematically varying reward function weights within a framework relying on the Proximal Policy Optimization (PPO), it is possible to quantitatively map the trade-off between collision avoidance safety and mission time. Our results, derived from simulations, show that agents trained with balanced reward weights achieve target-reaching success rates exceeding 98% in dynamic multi-obstacle scenarios. Conversely, configurations that disproportionately penalize obstacle proximity lead to overly cautious behavior and mission failure, with success rates dropping to 22% due to workspace boundary violations. This work provides a data-driven methodological framework for reward function design and parameter selection in safety-critical robotic applications, moving beyond ad-hoc tuning towards a more structured parameter influence analysis.

Development of AI-driven automated driving functions requires vast amounts of diverse, high-quality data to ensure road safety and reliability. However, manual collection of real-world data and creation of 3D environments is costly, time-consuming, and hard to scale. Most automatic environment generation methods still rely heavily on manual effort, and only a few are tailored for Advanced Driver Assistance Systems (ADAS) and Automated Driving (AD) training and validation. We propose an automated generative framework that learns real-world features to reconstruct realistic 3D environments from a road definition and two simple parameters for country and area type. Environment generation is structured into three modules - map-based data generation, semantic city generation, and final detailing. The overall framework is validated by training a perception network on a mixed set of real and synthetic data, validating it solely on real data, and comparing performance to assess the practical value of the environments we generated. By constructing a Pareto front over combinations of training set sizes and real-to-synthetic data ratios, we show that our synthetic data can replace up to 90% of real data without significant quality degradation. Our results demonstrate how multi-layered environment generation frameworks enable flexible and scalable data generation for perception tasks while incorporating ground-truth 3D environment data. This reduces reliance on costly field data and supports automated rapid scenario exploration for finding safety-critical edge cases.

The development of AI-driven automated driving functions requires vast amounts of diverse, high-quality data to ensure road safety and reliability. However, both the manual collection of real-world data and creation of 3D environments are costly, time-consuming, and hard to scale. Most automatic environment generation methods still rely heavily on manual effort, and only a few are tailored for Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS) training and validation. We propose an automated generative framework that learns ground-truth features to reconstruct 3D environments from a road definition and two simple parameters for country and area type. Environment generation is structured into three modules—map-based data generation, semantic city generation, and final detailing. The overall framework is validated by training a perception network on a mixed set of real and synthetic data, validating it solely on real data, and comparing performance to assess the practical value of the environments we generated. By constructing a Pareto front over combinations of training set sizes and real-to-synthetic data ratios, we show that our synthetic data can replace up to $85\%$ of real data without significant quality degradation. Our results demonstrate how multi-layered environment generation frameworks enable flexible and scalable data generation for perception tasks while incorporating ground-truth 3D environment data. This reduces reliance on costly field data and supports automated rapid scenario exploration for finding safety-critical edge cases.