CN-122021877-A - Multi-target game test scene generation method for automatic driving field test

Abstract

The invention provides a multi-target game test scene generation method for automatic driving field test, which comprises the steps of 1, collecting an initial information set of a current scene, 2, obtaining a scene feature vector F of the current scene, normalizing the initial information set, then encoding the static element set, the dynamic track set and the interaction semantic set, finally fusing the encoded information into a scene feature vector F, 3, configuring a scene generation constraint condition, 4, dynamically generating a dynamic loss function of the current scene based on a large language model LLM, 5, generating a time-sequence track, 6, checking scene consistency and normalizing output. The invention solves the problems of poor coordination of a background vehicle, insufficient scene consistency, deficient interaction of multiple intelligent agents and fixed parameter configuration in the generation of the existing automatic driving closed field test scene.

Inventors

• BI XIN
• ZHOU HAITAO
• SHEN DEWEI
• QI TIANHAO
• WANG ZIHAO
• ZHANG QIANCHENG

Assignees

• 同济大学

Dates

Publication Date

20260512

Application Date

20251230

Claims (5)

1. 1. The multi-target game test scene generation method for the automatic driving field test is characterized by comprising the following steps of: Step 1, acquiring an initial information set of a current scene, wherein the initial information set comprises a static element subset, a dynamic track subset and an interaction semantic subset; Step 2, obtaining a scene feature vector F of the current scene, namely normalizing an initial information set, then encoding a static element subset, a dynamic track subset and an interaction semantic subset, and finally fusing the encoded information into the scene feature vector F; Step 3, configuring a scene generation constraint condition; step4, dynamically generating a dynamic loss function matched with the current scene based on a large language model LLM; step 5, generating a time-sequence track; And 6, checking scene consistency and outputting in a standardized way.
2. 2. The multi-objective gaming test scenario generation method for automatic driving yard testing according to claim 1, wherein the scenario generation constraints include a speed constraint and a target point configuration constraint.
3. 3. The method for generating a multi-objective game test scenario for automatic driving field testing according to claim 1, wherein step 4 specifically comprises: Step 4A, constructing LLM prompt words, which comprise semantic description of scene feature vectors F, scene generation constraint conditions and predefined indexes of loss functions; and 4B, inputting LLM prompt words into a large language model LLM to generate a dynamic loss function adapting to the current scene.
4. 4. The method for generating a multi-objective game test scenario for automatic driving field testing according to claim 1, wherein step 5 specifically comprises: step 5A, inputting a scene feature vector F, a dynamic loss function and a noisy track of a background vehicle into a conditional diffusion guide model CGDM to generate a basic control sequence of the background vehicle; and 5B, inputting a basic control sequence of the background vehicle into a multi-agent cooperative decoder to generate a time-sequence track.
5. 5. The method of claim 1, wherein the step 6 comprises a trajectory consistency check and an environment matching consistency check.

Description

Multi-target game test scene generation method for automatic driving field test Technical Field The invention belongs to the technical field of automatic driving tests, and particularly relates to a multi-target game test scene generation method for automatic driving field test. Background While the real road test can capture a large number of dynamic interaction scenes, is limited by uncontrollable of complex environments, and is difficult to accurately quantify and verify the functional boundary of an automatic driving system in a closed loop manner, on the other hand, the closed field test has the advantages of repeatability and controllability, but is often caused by the fact that scene design is disjointed with the real road environment, and data sources are limited, so that the test result is difficult to effectively reflect the actual performance of the system. The automatic driving system safety verification is required to rely on the controllability and the repeatability of a closed field test, but the conventional scene generation method has the core defects that the conventional regularization method is fixed in parameters and can only cover a simple scene, the search-based method lacks multi-agent cooperative logic and is poor in scene interactivity, the conventional generation model is prone to the problems of track logic fracture and insufficient environment suitability, the conventional scheme cannot dynamically adapt to the generation strategy according to the customized test requirements, the test scene is disjointed from the real road behavior, and the decision planning boundary of the automatic driving system is difficult to verify accurately. The existing automatic driving closed test site scene generation method is mainly divided into three types, and the core logic and limitation of the method are as follows (1) the traditional closed site test method comprises the implementation steps that ① experts define a scene template based on test requirements, ② set fixed parameters through a parameterized tool, ③ generate a static scene and guide the static scene into a cloud platform to execute a test. The core structure is that a scene template library and a parameter configuration module are depended on, and a template library stores predefined road structure and test subject association rules. (2) A scene generation method based on search optimization comprises the steps of enabling ① to define an objective function, determining scene variables (such as a background car cut-in angle and a background car cut-in speed), enabling ② to conduct iterative search in a high-dimensional variable space by adopting a heuristic algorithm (such as a genetic algorithm and particle swarm optimization), finding out parameter combinations meeting a key scene (such as TTC is less than or equal to 1.5 s), and outputting and verifying the searched scene by ③. The core structure comprises an objective function definition module, an optimization algorithm module and a scene verification module, and approximates to the boundary of the key scene through iterative search. (3) A scene generation method based on a traditional generation model comprises the implementation steps of ① collecting historical test data (such as tracks of a host vehicle, a background vehicle and road characteristics) of a closed field, ② training to generate a countermeasure network (GAN) or a Variation Automatic Encoder (VAE), learning data distribution, and ③ inputting random noise to generate a new scene track (such as a background vehicle lane change track). The core structure is that the reality of the scene is improved by the countermeasure training by relying on a data preprocessing module, a generating model (such as a generator/discriminator of GAN) and a track output module. In summary, the defects in the prior art (1) are mainly that the generated simulation scene has poor coordination of a background car, insufficient scene consistency, deficient interaction of multiple agents and fixed parameter configuration. The parameter configuration is fixed, the rule-based method depends on a predefined template, and once the parameter is set, the parameter cannot be dynamically adjusted, for example, a background car in an AEB scene can only cut in a fixed speed, and the cutting rhythm cannot be adjusted in real time according to the speed of a main car, so that the scene lacks flexibility, and multiple test scenes such as 'the main car is low-speed- & gtthe background car is fast-cut, & gtthe main car is high-speed- & gtthe background car is slow-cut' cannot be covered. The multi-agent interaction is deficient, namely, a multi-agent interaction logic is not constructed on the basis of the rule and the searching method, for example, in a multi-vehicle lane changing scene of a closed field, a background vehicle only independently executes lane changing actions, the running state of a neighboring vehicle is not considered, the 'interactio