Life-long Learning and Testing for Automated Vehicles via Adaptive Scenario Sampling as A Continuous Optimization Process

TL;DR

This paper tackles the challenge of evaluating AV intelligence when the distribution of critical scenarios is unknown. It introduces a life-long learning and testing framework comprised of a bi-level loop: an outer loop minimizes the AV score by generating and evaluating scenarios, and an inner loop repositions newly generated samples to maximize coverage of the unknown subspace via an adaptive packing of representational spheres. The core methodological shift is modeling scenario coverage with sphere-like subspaces and using repulsive forces (with a KD-tree acceleration) to reduce overlap and quickly uncover challenging regions, coupled with updating the unknown-space distribution $f_{D_r}$ across rounds. Simulation results on a two-lane road show that the approach yields faster, more accurate AV evaluation by revealing more critical scenarios, albeit with a trade-off in total space coverage. Overall, the work enables continuous, scalable AV testing and learning by coupling adaptive scenario sampling with continual space knowledge accumulation.

Abstract

Sampling critical testing scenarios is an essential step in intelligence testing for Automated Vehicles (AVs). However, due to the lack of prior knowledge on the distribution of critical scenarios in sampling space, we can hardly efficiently find the critical scenarios or accurately evaluate the intelligence of AVs. To solve this problem, we formulate the testing as a continuous optimization process which iteratively generates potential critical scenarios and meanwhile evaluates these scenarios. A bi-level loop is proposed for such life-long learning and testing. In the outer loop, we iteratively learn space knowledge by evaluating AV in the already sampled scenarios and then sample new scenarios based on the retained knowledge. Outer loop stops when all generated samples cover the whole space. While to maximize the coverage of the space in each outer loop, we set an inner loop which receives newly generated samples in outer loop and outputs the updated positions of these samples. We assume that points in a small sphere-like subspace can be covered (or represented) by the point in the center of this sphere. Therefore, we can apply a multi-rounds heuristic strategy to move and pack these spheres in space to find the best covering solution. The simulation results show that faster and more accurate evaluation of AVs can be achieved with more critical scenarios.

Figures (12)

• Figure 1: Scenario generation using parameterized space. The static environment of scenario is constructed and the parameters of SVs models, like $\alpha^i$ and $\beta^i$ are extracted; Then the parameters are used to construct scenario sampling space and the point in space represents a scenario; The sampled point finally contributes to a detailed scenario for AV.
• Figure 2: Important processes in life-long learning and testing.
• Figure 3: Pipeline of the proposed scheme.
• Figure 4: Illustration for different situations on covering space under the two-dimensional projection.
• Figure 5: Illustration on the moving process step by step under the two-dimensional projection.