In the current automotive ecosystem, advanced driver-assistance systems (ADAS) rely on trajectory datasets to train and validate their models. However, these datasets mostly reflect routine driving, leaving real vehicle-to-vehicle conflict events underrepresented. The rarer the incident, the higher the cost when an ADAS fails to handle it. This is where SevDiff comes in, a denoising diffusion probabilistic model (DDPM) that conditions trajectory generation on a target time-to-collision (TTC) value. This innovation not only allows the creation of synthetic scenarios with controlled severity, but also opens the door to a new way of evaluating the robustness of autonomous driving systems.
SevDiff is trained on 468 interaction windows extracted from the UTE SQM-W-1 dataset, which records over 822,000 observations after smoothing. The model accepts as input a scalar requested minimum TTC and generates interaction trajectories between two vehicles whose actual severity matches the request. Results are impressive: it achieves a 100% hit rate within a ±0.5-second margin for TTC targets between 0.5 and 1.5 seconds, and remains between 97% and 99% for values of 2.0 to 2.5 seconds. The degradation to 39% at TTC = 5.0 seconds is physically interpretable and constitutes a precise characterization of the generator, not a simple pass/fail result.
From a technical perspective, the ability to condition generation on a severity indicator such as TTC represents a qualitative leap over previous approaches that only considered scene-level properties like spatial goals or agent structures. SevDiff proves that it is possible to generate conflict events with measurable precision, which is critical for training ADAS in edge cases. The kinematic plausibility of generated trajectories is confirmed by a maximum out-of-range rate of 4.7% across twelve features, and no negative speed or gap values in more than 96.5% of samples.
For companies developing autonomous driving technology or active safety systems, adopting models like SevDiff is not just about innovation, but about competitive advantage. The ability to simulate conflict scenarios on demand drastically reduces reliance on costly and difficult-to-collect real-world data. At Q2BSTUDIO, a company specialized in software and technology development, we understand that creating such models requires a combination of expertise in artificial intelligence, large-scale data processing, and robust cloud infrastructure. That is why we offer applied AI services that enable our clients to design, train, and deploy conditioned trajectory generators tailored to their specific safety and validation needs.
Integrating SevDiff into ADAS validation pipelines involves handling massive data volumes and running real-time inferences. This is where the cloud comes in. At Q2BSTUDIO we work with cloud AWS and Azure to provide scalable environments that support everything from distributed training of diffusion models to parallel simulation of thousands of conflict scenarios. Moreover, cybersecurity is a fundamental pillar: ensuring the integrity and confidentiality of trajectory data and models is critical in an industry where a failure can have catastrophic consequences. Therefore, we incorporate cybersecurity practices in every development phase.
Beyond trajectory generation, analyzing the results requires Business Intelligence tools that transform validation data into actionable insights. Power BI dashboards, for instance, can visualize SevDiff's hit rate as a function of target TTC, identify model biases, or compare performance across different ADAS configurations. At Q2BSTUDIO, we design BI/Power BI solutions that extract the full potential of generated data, facilitating informed decision-making.
The application of models like SevDiff is not limited to the automotive sector. The same severity-conditioning logic could be applied to other domains where rare events are critical, such as collaborative robotics, air navigation, or urban traffic management. Any system interacting in dynamic environments can benefit from synthetic generators that allow testing responses to edge cases without risking people or resources.
At Q2BSTUDIO, we believe the key lies not just in adopting cutting-edge technologies, but in integrating them coherently with each organization's strategy. That is why we offer custom software that encapsulates AI models, connects to real and simulated data sources, and deploys on reliable cloud infrastructures. The combination of AI, cloud, cybersecurity, and BI gives companies full control over their validation and certification processes.
The future of autonomous driving lies in the ability to anticipate the unexpected. Models like SevDiff demonstrate that conditioned generation of conflict events is feasible and precise. At Q2BSTUDIO, we are ready to help our clients implement these solutions, adapting them to their specific needs in software development, data analysis, and productive deployment. Safety is not an extra; it is the fundamental requirement.
