⚡ electrical engineering

SevDiff: Severity-Conditioned Diffusion for Long-Tail Conflict Trajectory Generation

This paper introduces SevDiff, a severity-conditioned diffusion model that generates physically plausible, long-tail conflict trajectories by accepting a target Time-to-Collision (TTC) value as input, achieving high hit-rates for critical low-TTC scenarios while providing a physically interpretable metric for generator precision.

Original authors: Eni Solomon Laughter

Published 2026-07-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Eni Solomon Laughter

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to drive a car. You show it millions of hours of video from real drivers, but there's a catch: almost all of that video shows boring, safe driving. The robot learns to handle traffic jams and merging lanes perfectly, but it has never seen a near-crash. If you ask the robot to predict what happens when two cars almost hit each other, it will likely guess wrong because it has never practiced that specific, scary moment. This is the "long-tail" problem in artificial intelligence: the rare, dangerous events are the most important to get right, but they are too rare in real data to learn from naturally.

To fix this, scientists use "generative models," which are like digital artists that can invent new traffic scenes. Usually, these artists are told to draw a "lane change" or a "merge," but they can't be told exactly how dangerous the scene should be. It's like asking a painter, "Draw me a storm," but having no way to specify if you want a gentle breeze or a hurricane. This paper introduces a new tool called SevDiff that solves this by letting engineers specify the exact level of danger they want to see, measured by a concept called Time-to-Collision (TTC). Think of TTC as a countdown clock: if two cars are heading toward a crash, the clock tells you how many seconds until they hit. A low number (like 0.5 seconds) means a crash is imminent; a high number (like 5.0 seconds) means they are just driving safely. The big question this paper asks is: Can we build a robot artist that can be handed a specific number on that clock—say, "0.8 seconds"—and reliably draw a traffic scene that matches that exact level of danger?

The authors built SevDiff, a special kind of AI that acts like a "severity dial" for traffic accidents. Instead of just guessing what a dangerous scene looks like, SevDiff takes a specific number (the target TTC) as an instruction and generates a pair of car trajectories that match that danger level. The researchers tested this by asking the AI to create 300 different scenarios for various danger levels, ranging from a near-miss (0.5 seconds) to a very safe distance (5.0 seconds).

The results were surprisingly precise for the most critical moments. When the researchers asked for a near-crash with a TTC between 0.5 and 1.5 seconds, the AI got it right 100% of the time. Every single generated scenario landed within a tiny margin of error (±0.5 seconds) of the requested danger level. Even for slightly less urgent targets like 2.0 to 2.5 seconds, the AI was correct 97% to 99% of the time. However, as the requested danger level became more routine (like a safe 5.0-second gap), the AI's precision dropped, hitting the target only 39% of the time.

This drop in accuracy isn't a bug; it's a feature that the authors explain makes perfect sense. When you ask for a super-dangerous scenario (a tiny TTC), the AI is being asked to do something very different from the boring, safe driving it saw in its training data. The "danger signal" is so loud and unique that the AI follows it perfectly. But when you ask for a safe, routine scenario (a large TTC), the AI is being asked to do something that looks just like the thousands of safe drives it already knows. In those cases, the AI's own memory of "normal driving" competes with your specific instruction, making it slightly less precise.

The paper also checked if the AI was making up impossible physics, like cars driving backward or passing through each other. They found that the generated scenes were mostly realistic, with fewer than 5% of the features (like speed or distance) falling outside the realm of possibility. The authors note that while the AI is excellent at creating the statistics of a dangerous moment, it doesn't yet draw the full, frame-by-frame movie of the crash; it creates a summary of the event that can be turned into a visual story.

In short, SevDiff proves that we can teach an AI to generate traffic conflicts on demand with a high degree of control. It offers a new way to test safety systems by creating exactly the kind of rare, dangerous situations that real-world data is too scarce to provide. The authors suggest that while this tool is currently limited to a specific type of highway weaving section, it opens the door for a future where we can systematically stress-test self-driving cars against any level of danger we choose, rather than just hoping to stumble upon them on the road.

Technical Summary: SevDiff – Severity-Conditioned Diffusion for Long-Tail Conflict Trajectory Generation

1. Problem Statement

Advanced Driver Assistance Systems (ADAS) and autonomous vehicles are most likely to fail in rare, safety-critical scenarios (the "long tail"), yet these events are severely underrepresented in naturalistic driving datasets. While existing generative models (GANs, VAEs, diffusion models) can produce rare-looking scenarios by conditioning on spatial goals, agent structures, or natural-language objectives, they suffer from a severity-controllability gap. Current methods cannot accept a specific numeric target for conflict severity—such as a Time-to-Collision (TTC) value—and guarantee that the generated trajectory will realize that severity within a measurable error margin. Consequently, test engineers lack a tool to systematically generate scenarios at graded levels of danger (e.g., "generate 100 scenarios with a minimum TTC of 1.5s") for rigorous system validation.

2. Methodology: SevDiff

The paper introduces SevDiff, a Severity-Conditioned Denoising Diffusion Probabilistic Model (DDPM) designed to bridge this gap.

Conditioning Mechanism: Unlike latent-space conditioning where a signal must be decoded from an embedding, SevDiff conditions the diffusion process directly on a scalar Inverse Time-to-Collision (ITTC) value. ITTC is used because standard TTC is undefined when vehicles are not closing the gap; ITTC is defined as $\max(v_f - v_l, 0) / d$ , where $v$ is speed and $d$ is headway. This ensures every interaction window yields a valid conditioning label.
Architecture: The model operates in a 12-dimensional kinematic feature space (including mean/std speeds, headways, relative speeds, and ITTC statistics). The conditioning signal (target ITTC) is injected as a sinusoidal embedding at every denoising step, alongside the timestep embedding. The network learns to reverse the noise process while steering the output distribution toward the requested severity.
Data & Preprocessing: The model is trained on the UTE SQM-W-1 expressway weaving-section dataset (1,041 vehicles, ~822k observations).
- Smoothing: Savitzky-Golay filters were applied to remove noise in lateral distance and speed channels.
- Labeling: Interaction windows were extracted, and severity was determined by the peak ITTC within the window.
- Binning: Windows were assigned to 10 severity bins (TTC = 0.5s to 5.0s). Due to the natural long-tail distribution (only ~14% of windows fell in the 0–5s range), inverse-frequency weighted sampling was employed during training to ensure all severity levels contributed equally to gradient updates.
Evaluation Metric: The primary metric is Hit-Rate, defined as the fraction of generated samples where the realized minimum TTC falls within a defined error band ( $\pm \delta$ ) of the requested target $r$ . This contrasts with prior work that relies on aggregate distributional statistics or collision rates.

3. Key Contributions

Scalar SSM Conditioning: SevDiff is the first framework to condition a diffusion model directly on a numeric Surrogate Safety Measure (SSM) target (ITTC/TTC) and evaluate generation precision via a hit-rate metric.
Validated Preprocessing Pipeline: The authors present a robust pipeline for naturalistic weaving-section data, utilizing internal consistency diagnostics, Savitzky-Golay smoothing, and Peak-Over-Threshold (POT) severity binning to produce clean, always-defined ITTC labels.
Precision-vs-Severity Characterization: The paper provides a detailed characterization of generator performance across the severity spectrum, demonstrating that hit-rate degradation is physically interpretable based on the strength of the conditioning signal relative to the training prior.

4. Results

Evaluated on 300 samples per target across 10 TTC levels (0.5s to 5.0s):

High-Severity Precision: SevDiff achieves a 100% hit-rate within $\pm 0.5$ s for TTC targets of 0.5s to 1.5s. For targets of 2.0s–2.5s, the hit-rate remains 97–99%.
Degradation Pattern: Performance gracefully degrades for routine scenarios (TTC > 3.0s), dropping to 85% at 3.0s, 53% at 4.0s, and 39% at 5.0s. The authors interpret this not as a failure, but as a physical reality: at high TTC (low ITTC), the conditioning signal is close to the training prior's center of mass, causing the model to be pulled between the target and the natural distribution.
Physical Plausibility: Generated kinematic features are largely physically plausible. The maximum out-of-range (OOR) rate across 12 features is 4.7%. However, 3.5% of samples exhibited negative minimum headways, and 5.4% had mean ITTC exceeding max ITTC, indicating limitations in the unconstrained continuous feature space.
Realism: Nearest-neighbor analysis confirms that generated feature vectors are not statistical outliers; a generated TTC=3.0s scenario was closer to a real same-severity interaction than 94.6% of real-to-real pairs.

5. Significance and Claims

The paper claims that SevDiff provides a precision characterization of a generative model rather than a simple pass/fail result. By demonstrating that a diffusion model can be conditioned on a specific safety metric (TTC) with high fidelity in the critical long-tail region, SevDiff offers a tool for the structured, scenario-based testing required by emerging regulatory frameworks (e.g., ISO 34502, GB/T 44719).

The authors emphasize that the hit-rate degradation curve is a diagnostic tool: it quantifies exactly how much control is achievable at each severity level. This allows test engineers to understand the limits of synthetic data generation for specific safety cases. The work establishes that conditioning on a scalar SSM is feasible and effective, particularly for the rare, high-severity events that are most critical for ADAS validation, provided the training data is appropriately rebalanced.

6. Limitations and Future Directions

The authors explicitly note three limitations:

Generalization: The model is trained on a single weaving-section geometry; results may not transfer to other road types without retraining.
Physical Constraints: The unconstrained generation space leads to small percentages of physically impossible outputs (negative headways), suggesting a need for inequality-constrained objectives.
Temporal Structure: The model generates 12 summary statistics, not per-frame trajectories. Visualization requires a deterministic reconstruction step, which is not a learned capability.

Future work is proposed to address these by implementing constrained generation, training on multi-dataset/HD-map scenarios (e.g., nuPlan, Waymo), and adopting scene-encoding architectures like VectorNet for full trajectory synthesis.