The article proposes a novel paradigm for testing and improving the intelligence of automated vehicles (AVs). Unlike previous studies that aim to minimize the expectation of the failing frequency of an AV in various scenarios, this approach aims to explore all possible scenarios that could be sampled to ensure an AV works well in all such scenarios.
The key aspects of the proposed scheme are:
Formulating the testing as a continuous optimization process: The testing is formulated as an optimization problem that iteratively generates potential critical scenarios and evaluates the AV's performance. The goal is to minimize the AV's score by finding more critical scenarios.
Life-long learning and testing: The scheme involves a two-loop process. The outer loop generates new samples, evaluates the AV, and updates the space knowledge. The inner loop maximizes the coverage of the currently unknown subspace by sampling and rearranging the new samples.
Adaptive scenario sampling: The inner loop uses a heuristic strategy to pack the new sample spheres in the space. The spheres are assumed to have repulsive forces that push them away from the known subspaces and overlap with each other, in order to efficiently cover the unknown subspace.
Comprehensive evaluation: By continuously identifying critical scenarios through the life-long learning and testing process, the scheme aims to comprehensively assess the intelligence of the AV, going beyond just minimizing the expectation of failing frequency.
The simulation results demonstrate that the proposed scheme can achieve faster and more accurate evaluation of AVs by identifying more critical scenarios compared to standard Monte Carlo and Quasi-Monte Carlo methods.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Jingwei Ge,P... at arxiv.org 05-03-2024
https://arxiv.org/pdf/2405.00696.pdfDeeper Inquiries