CUSP: Cross-domain Uncertainty-aware Driving Style Posterior

Inferring driving style from real-world trajectories is challenging under domain shift: datasets collected across regions or sites differ in speed distributions, congestion regimes, sensing noise, and annotation quality, causing learned ``styles'' to absorb domain-specific shortcuts such as absolute speed.
We propose CUSP, an uncertainty-aware framework that represents driving style with interpretable mechanism parameters of the Intelligent Driver Model (IDM), such as desired speed and time headway, and infers a posterior from short interaction windows under domain shift.
CUSP performs scalable amortised posterior inference and learns the parameters via a physics-grounded likelihood with variational regularisation.
To keep style semantics comparable across datasets, we stabilise the inferred parameters by discouraging domain leakage and absolute-speed shortcuts through adversarial training, optionally aided by masking speed-related channels and using simple neighbourhood context.
Experiments under cross-location and cross-site evaluations on highD and NGSIM show that CUSP preserves mechanistic fidelity while reducing speed and domain information in the inferred parameters, yielding more stable style semantics and uncertainty signals that support downstream interaction modelling.