SHIFT

SHIFT is a novel framework combining uncertainty modeling with informative priors derived from automated rule extraction. SHIFT reformulates trajectory prediction as a classification task and employs heteroscedastic spectral‑normalized Gaussian processes to disentangle epistemic and aleatoric uncertainties. Informative priors are learned from natural language driving rules using a retrieval‑augmented generation framework powered by a large language model. Extensive evaluations on the nuScenes dataset demonstrate that SHIFT outperforms state‑of‑the‑art methods, particularly in complex scenarios such as intersections.

SHIFT leverages a two‑stage training process: first, encoding traffic rules as synthetic labels; second, regularizing real‑world trajectory predictions with these learned priors. The model architecture integrates a spectral‑normalized CNN backbone with a heteroscedastic Gaussian process output layer, enabling robust uncertainty quantification.

Qualitative Comparison of Trajectory Predictions. The first row displays predictions from SNGP with Rules as baseline, while the second row presents results from SHIFT. Each column represents a different scene. The target agent is highlighted with a black rectangle. The past trajectory (shown only for SHIFT) is also included for reference.

SHIFT achieves substantial improvements in uncertainty calibration and displacement metrics compared to baseline methods. It excels in low-data and cross-location scenarios, demonstrating strong generalization capabilities.

Please use the following BibTeX entry to cite this work:

@inproceedings{manas2025shift,
  title     = {Uncertainty-Aware Trajectory Prediction via Rule-Regularized Heteroscedastic Deep Classification},
  author    = {Manas, Kumar and Schlauch, Christian and Paschke, Adrian and Wirth, Christian and Klein, Nadja},
  booktitle = {Proceedings of the Robotics: Science and Systems (RSS)},
  year      = {2025}
}