Automatic Parameter Tuning of Self-Driving Vehicles

TL;DR

This work tackles automatic tuning of trajectory-planning and controller parameters for automated driving by leveraging expert demonstrations. It introduces a demonstration-based cost function $J$ that quantifies the mismatch between closed-loop trajectories and demonstrations, and optimizes parameters using three local methods: Gradient Descent, Unscented Kalman Filter, and Maximum Likelihood Estimation. The methods are applied to a lateral trajectory planner within a linear time-varying MPC framework, tuning weights to reproduce real-world driving behavior despite noisy data. Results show that all three approaches improve manual parameter settings, with Maximum Likelihood Estimation delivering the best overall performance in the study.

Abstract

Modern automated driving solutions utilize trajectory planning and control components with numerous parameters that need to be tuned for different driving situations and vehicle types to achieve optimal performance. This paper proposes a method to automatically tune such parameters to resemble expert demonstrations. We utilize a cost function which captures deviations of the closed-loop operation of the controller from the recorded desired driving behavior. Parameter tuning is then accomplished by using local optimization techniques. Three optimization alternatives are compared in a case study, where a trajectory planner is tuned for lane following in a real-world driving scenario. The results suggest that the proposed approach improves manually tuned initial parameters significantly even with respect to noisy demonstration data.

Figures (4)

• Figure 1: Iterative parameter tuning approach with the simulated closed-loop trajectory $\boldsymbol{x}|_{[0,T]}$ of a trajectory planner or controller parameterized by $\boldsymbol{p}$, the input disturbance $\boldsymbol{z}_t$, the recorded trajectory of a demonstration $\boldsymbol{y}_{d}|_{[0,T]}$, and the tuning cost function $J$.
• Figure 2: Kinematic vehicle model with respect to a given reference curve $\Gamma$ based on gutjahr2016lateral.
• Figure 3: Bird's eye view of the relevant part of the BMW test track in Aschheim with the driving scenario path marked in red.
• Figure 4: Comparison of the resulting lateral deviation $d$ from the reference path and its curvature $\kappa$ using the tuned parameters with different optimization methods for the evaluation scenario.

Theorems & Definitions (1)

• Remark 1