Building Real-time Awareness of Out-of-distribution in Trajectory Prediction for Autonomous Vehicles

AAAI 2026 Workshop TrustAgent Submission29 Authors

Keywords: Autonomous Vehicles, Trajectory Prediction, Out-of-Distribution Detection, Uncertainty Quantification, Change-point Detection, Quickest Change Detection, Robustness, Sequential Forecasting, Deep Learning, Safety-Critical AI

Abstract:

Ensuring trust and control in autonomous systems requires
that they remain robust and reliable when facing the unpre-
dictable complexity of the real world. In the context of agen-
tic AI for autonomous driving, accurate trajectory prediction
is foundational to safety-critical decision-making. However,
due to discrepancies between training data and real-world
conditions encountered during inference, even well-trained
machine learning models may produce unreliable predictions.
Such sim-to-real gaps (also known as imperfect training data)
may be unavoidable due to the overwhelming complexity of
data annotation and environment uncertainties. To support
verifiable and trustworthy autonomy, we present a principled
and computationally efficient framework for detecting when
a model’s predictions deviate from expected, in-distribution
behavior. Leveraging the intuition that in-distribution (ID)
scenes exhibit error patterns similar to training data, while
out-of-distribution (OOD) scenes do not, we formulate OOD
detection as a quickest change-point detection problem, en-
abling timely recognition of subtle or deceptive shifts in driv-
ing scenes that may compromise reliability. We address the
challenging settings where the OOD scenes are deceptive,
meaning that they are not easily detectable by human intu-
itions. Our solutions can handle the occurrence of OOD at
any time during trajectory prediction inference. Experimental
results on multiple real-world datasets demonstrate the effec-
tiveness of our methods.

Submission Number: 29