SBOF: An End-to-End Framework for Simulative Black-Box Optimization of Hybrid Perceptive Functions

The adaptation of perceptive functions to new tasks or new domains is currently a major challenge in the field of autonomous driving. Since the creation of manually labeled datasets is generally time-consuming and cost-intensive, training approaches based on virtual simulation are becoming more relevant. Current approaches focus on the usage of virtual simulation to adapt machine learning models. However, in many automotive applications, perceptive functions consist of combined machine learning and traditional-rule-based approaches, called hybrid approaches. Rather than changing underlying machine learning models, we approach the task of domain adaptation of such hybrid perceptive functions by the optimization of often neglected and arbitrary black-box parameters. We therefore present a general methodology for end-to-end black-box optimization of hybrid perceptive functions for a specific target domain using virtual simulation. In a proof-of-concept, we carried out experiments to adapt a hybrid 3D object detector to the KITTI domain, using the simulator CARLA. Without the need for manually labeled data, our proposed methodology cannot only improve performance in the simulated target domain but can also improve performance in the real target domain. Furthermore, we compare the effectiveness of different black-box optimization algorithms for this specific task. Based on convergence behavior and the maximum achieved performance, we recommend the Covariance Matrix Adaptation Evolution Strategy (CMA-ES).