Learning the Pareto Space of Multi-Objective Autonomous Driving: A Modular, Data-Driven Approach

TL;DR

The paper develops an empirical, data-driven framework to map multi-objective trade-offs in autonomous driving by learning a Pareto surface over safety, efficiency, and interaction from naturalistic trajectories. Each driving event is embedded as a triple $ (S,E,I) $, and Pareto dominance identifies non-dominated states; a Gaussian Process Regression (GPR) model then yields a smooth empirical frontier. Applied to the Foggy Bottom and I-395 TGSIM datasets, the study finds Pareto-optimal states are exceedingly rare ($0.23\%$), yet those states exhibit higher scores across all objectives, with interaction offering the greatest room for improvement. The resulting Pareto surface provides a transferable, modular basis for Pareto-informed learning and control in multi-objective autonomous driving.

Abstract

Balancing safety, efficiency, and interaction is fundamental to designing autonomous driving agents and to understanding autonomous vehicle (AV) behavior in real-world operation. This study introduces an empirical learning framework that derives these trade-offs directly from naturalistic trajectory data. A unified objective space represents each AV timestep through composite scores of safety, efficiency, and interaction. Pareto dominance is applied to identify non-dominated states, forming an empirical frontier that defines the attainable region of balanced performance. The proposed framework was demonstrated using the Third Generation Simulation (TGSIM) datasets from Foggy Bottom and I-395. Results showed that only 0.23\% of AV driving instances were Pareto-optimal, underscoring the rarity of simultaneous optimization across objectives. Pareto-optimal states showed notably higher mean scores for safety, efficiency, and interaction compared to non-optimal cases, with interaction showing the greatest potential for improvement. This minimally invasive and modular framework, which requires only kinematic and positional data, can be directly applied beyond the scope of this study to derive and visualize multi-objective learning surfaces

Figures (8)

• Figure 1: Conceptual overview of the overarching framework linking empirical data to learning-based control.
• Figure 2: Distribution of maximum interaction values ($M_{\text{max}}$). Dashed line: 97th percentile ($u=1.8536$).
• Figure 3: Distribution of time headways across all AVs. Dashed line: 4 s (relaxed following).
• Figure 4: Distribution of first follower string stability gain. Dashed line: Gain = 1.
• Figure 5: Distribution of jerk. Dashed line: $2.5~\mathrm{m/s^3}$.