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

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

💡 Research Summary

The paper presents a data‑driven framework for learning the Pareto frontier of autonomous driving performance across three core objectives: safety, efficiency, and interaction. By treating each vehicle‑time step as a point in a three‑dimensional objective space, the authors derive composite scores from naturalistic trajectory data and identify non‑dominated (Pareto‑optimal) states. The methodology is demonstrated on two high‑resolution datasets from the Third Generation Simulation (TGSIM) project—Foggy Bottom (urban intersections) and I‑395 (freeway segment).

Key methodological steps include: (1) defining a detection zone based on Tesla Vision to select relevant surrounding agents; (2) using an affine‑spacing policy to identify leaders and followers for headway and string‑stability calculations; (3) constructing three groups of surrogate metrics. Safety is quantified with a Generalized Surrogate Safety Measure (GSSM) that combines distance, angle, and relative speed, and models the upper tail of risk scores with a Generalized Pareto Distribution. Efficiency combines average forward headway and a string‑stability gain metric. Interaction is captured by jerk magnitude and sustained deceleration intensity. All raw metrics are min‑max normalized to the