Bridging Semantic Intent and Physical Reality
Traditional anomaly generation methods often struggle to balance the need for complex, intent-driven behavior with the physical limitations of real-world mobility systems. This paper introduces a novel framework that uses Large Language Models (LLMs) to define high-level behavioral anomalies—such as erratic driving patterns or navigation errors—while simultaneously applying a kinematic constraint layer to ensure the resulting trajectories are physically executable.
The Two-Layer Generation Architecture
The proposed system operates through a dual-process pipeline:
- Semantic Layer (LLM): The LLM acts as the behavioral engine, generating natural language descriptions or intent-based sequences that define the 'what' and 'why' of the anomaly. This allows for the creation of diverse, non-repetitive, and contextually rich scenarios that simple stochastic noise models cannot replicate.
- Kinematic Constraint Layer: To prevent the LLM from suggesting physically impossible maneuvers (e.g., instantaneous velocity changes or impossible turning radii), a secondary constraint layer processes the LLM's output. This layer maps the high-level intent into valid coordinate-based trajectories that adhere to the specific physics of the mobility platform (e.g., car, drone, or robot).
Impact on Anomaly Detection
By generating synthetic data that is both semantically complex and physically valid, this approach significantly improves the training and evaluation of anomaly detection systems. It allows researchers to stress-test detection models against 'edge-case' behaviors that are rare in real-world datasets but critical for safety-critical systems. The framework effectively reduces the 'reality gap' often found in synthetic mobility data, providing a more robust foundation for training autonomous agents to recognize and respond to unexpected environmental or behavioral deviations.