Conformal Prediction of Motion Control Performance for an Automated Vehicle in Presence of Actuator Degradations and Failures

TL;DR

The paper tackles ensuring safe automated driving under actuator degradations by building a data-driven self-representation model of the motion controller. It trains a neural predictor to estimate the maximum lateral deviation $\varepsilon_{lat,max}$ for lane-change maneuvers under nominal, degraded, and failed actuator states, and wraps predictions with conformalized quantile regression to produce prediction intervals with guaranteed coverage $C_\alpha$. A large simulation-based dataset (20,000 samples) spanning road segments, maneuvers, and degradation scenarios supports training and evaluation, demonstrating near-desired coverage and informative interval lengths. The approach yields a principled, uncertainty-aware tool to constrain the admissible action space during behavior generation and highlights avenues for real-world validation and refinement to reduce conservatism.

Abstract

Automated driving systems require monitoring mechanisms to ensure safe operation, especially if system components degrade or fail. Their runtime self-representation plays a key role as it provides a-priori knowledge about the system's capabilities and limitations. In this paper, we propose a data-driven approach for deriving such a self-representation model for the motion controller of an automated vehicle. A conformalized prediction model is learned and allows estimating how operational conditions as well as potential degradations and failures of the vehicle's actuators impact motion control performance. During runtime behavior generation, our predictor can provide a heuristic for determining the admissible action space.

Figures (4)

• Figure 1: The MOBILE research vehicle in IPG CarMaker, figure from stolte_toward_2023.
• Figure 2: Simplified functional architecture of an automated vehicle, figure taken from stolte_toward_2023, originally based on ulbrich_towards_2017. Our conformalized control performance prediction model relies on self-perception information and is located next to the behavior generation block.
• Figure 3: Distribution of $\varepsilon_{\mathrm{lat}, \mathrm{max}}$ in the dataset
• Figure 4: Nominal path (blue) for a lane change maneuver at $a_{\mathrm{max}} = 3\,\mathrm{m} / \mathrm{s}^2$ compared to the resulting path (red) in case of $\mathrm{D}3$ (see \ref{['tab:deg_examples']}). Blue area indicates lateral spaced covered by prediction $\pm\hat{\varepsilon}_{\mathrm{lat}, \mathrm{Hi}}$ with $C_{\alpha, 2} = 99\%$. Light red area indicates the additional space covered by the ego-vehicle $\pm w_{\mathrm{Veh}}/2$. Red shaded area indicates occupancy by an obstacle.