CN-122023589-A - Automatic driving scene generation method based on multi-objective antagonism optimization

Abstract

The invention provides an automatic driving scene generation method based on multi-objective antagonism optimization, which comprises the steps of 1, constructing a noise generator, wherein the inputs of the noise generator are a current noisy scene x t , a high-definition map code m and a time step t, the outputs of the noise generator are prediction noise epsilon θ (x t and t, 2, designing a multi-objective evaluator L total , wherein L total is a weighting function of a plurality of potential functions, and the independent variables of the potential functions are used for predicting a denoising scene Step 3, performing inverse sampling of the contrast guidance, step 4, outputting a final scene x 0 . On the premise of strict scene authenticity, the method and the device can efficiently and controllably generate the critical countermeasure scene with the highest test value for the automatic driving system, thereby greatly improving the test coverage rate and the efficiency.

Inventors

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

Assignees

• 同济大学

Dates

Publication Date

20260512

Application Date

20251229

Claims (4)

1. 1. An automatic driving scene generation method based on multi-objective antagonism optimization is characterized by comprising the following steps: step 1, constructing a noise generator: Constructing a noise generator based on a transducer architecture, which is used for learning joint space-time distribution of real traffic scenes, wherein the input of the noise generator is that a current noisy scene x t , a high-definition map code m and a time step t are taken as inputs, and the output of the noise generator is prediction noise epsilon θ (x t , t); Step 2, designing a multi-objective evaluator L total ; wherein L total is a weighted function of a plurality of potential functions, and independent variables of the potential functions are used for predicting a denoising scene Obtaining; Step 3, performing inverse sampling of the resistance guidance: Starting from the standard noisy scene x T , an iterative denoising loop is performed until time step t=0, at any intermediate time step t, the denoising loop performs the following sub-steps: Step 3A, denoising estimation, namely estimating the current estimated denoising scene through a back diffusion formula by utilizing the current noisy scene x t and the predicted noise epsilon θ (x t and t output by the noise generator Step 3B, gradient calculation, namely pre-estimating a denoising scene Inputting into a multi-objective evaluator, calculating a total energy value L total , and calculating a guidance gradient of the total energy with respect to the current noisy scene x t using a chain law Step 3C, noise correction, namely introducing a guide scale parameter gamma, and correcting the predicted noise epsilon θ (x t , t) output by the noise generator by utilizing the guide gradient g to obtain corrected noise epsilon θ ′(x t t); Step 3D, updating the state, namely updating the current noisy scene x t into a noisy scene x t-1 at the next moment based on the correction noise epsilon θ ′(x t t until a final scene x 0 is generated; And 4, outputting a final scene x 0 .
2. 2. The method for generating an automated driving scenario based on multi-objective resistance optimization of claim, wherein L total : L total ＝W real (L collision +L on - road )+w adv (L region +L attr +L adversarial ) Wherein, L collision is an anti-collision constraint function; L on-road -the in-transit constraint function; l region -a spatial region constraint function; L attr -agent attribute constraint functions; L adversarial -an antagonistic constraint function.
3. 3. The automated driving scenario generation method of claim based on multi-objective antagonism optimization, wherein the correction gradient g: -gradient of the potential function at time step t.
4. 4. A method of generating an autopilot scenario based on multi-objective antagonism optimization as defined in claim 3 wherein the noise epsilon θ ′(x t ,t)＝ε θ (x t , t) -gamma-g is modified.

Description

Automatic driving scene generation method based on multi-objective antagonism optimization Technical Field The invention belongs to the technical field of automatic driving automobiles, and particularly relates to an automatic driving scene generation method based on multi-objective antagonism optimization. Background Reliability verification of autopilot systems is the biggest challenge facing its large-scale commercial application. To achieve a level of safety comparable to human drivers, the automated driving system (autopilot DRIVING SYSTEM, ADS) needs to perform a test mileage equivalent to billions kilometers. However, in natural driving data, most of the time, it is a low-risk event-free scenario, while the low-probability, high-risk "corner cases" that lead to system failure are extremely sparse. Therefore, it is generally known in the industry that a technique of generating a key countermeasure scene in a highly efficient, controllable and directional manner must be adopted to greatly improve the coverage rate and efficiency of the simulation test. In the existing scene generation technology, the implementation scheme closest to the antagonism object of the invention mainly focuses on a method based on generating an antagonism network and deep reinforcement learning. The GANs-based scheme generally constructs a three-party game architecture, namely, a generator and a arbiter guarantee scene authenticity, and additionally introduces a criticality evaluator to maximize risk to ADS. However, the performance of this scheme is severely dependent on the stability of GANs training, and is very prone to pattern collapse, resulting in serious shortages of diversity of the generated scenario, and the whole risk space cannot be fully explored. More critical is that GANs, when processing high-dimensional, long-time-series trajectory data, it is difficult to maintain long-term kinetic constraints and causal relationships of the trajectory, making the resulting sequence lack of physical coherence. Another implementation is based on DRL, which models the generation process of multi-Agent scenarios as a markov decision process, learning strategies by Agent to maximize rewards that trigger SUT failure. While DRL is theoretically capable of handling timing dependencies, in practical applications, the reward signal is extremely sparse as SUT failure is a very low probability event. This rewarding sparsity makes the model training process extremely slow and difficult to converge to an effective key generation strategy in the face of high-dimensional, long-sequence scenes. Meanwhile, the high-dimensional state and action space bring great exploration difficulty to the DRL. In summary, no matter whether the pattern breakdown and the timing consistency of GANs are insufficient, or the rewarding sparsity and the convergence efficiency of the DRL are low, the prior art cannot provide a mechanism, which can efficiently explore the rare risk path at the same time, and precisely balance pursuing the high-risk antagonism and the reality conforming to the physical traffic specification. This is a key technical bottleneck that the present invention needs to address. Disclosure of Invention The invention aims to provide the automatic driving scene generation method based on the multi-objective antagonism optimization, which can efficiently and controllably generate the critical antagonism scene with the highest test value for the automatic driving system on the premise of strict scene authenticity, thereby greatly improving the test coverage rate and efficiency. The invention provides a diffusion model automatic driving scene driving generation method based on multi-objective antagonism optimization, and aims to solve the key bottleneck faced by the existing automatic driving scene generation technology (such as rule-based, VAE/GAN and the like). In the prior art, it is difficult to efficiently explore edge scenes with low probability and high risk, and it is difficult to balance pursuit of high risk antagonism and authenticity conforming to traffic rules, so that the generated scene sequence has poor continuity and low test value. The core solution of the invention is to combine the rich high-dimensional sequence generating capability of the diffusion model with the multi-objective antagonistic guide language: the method is characterized in that a multi-target comprehensive loss mechanism comprising antagonism, authenticity and diversity is embedded into each step of iteration mechanism of a diffusion model reverse denoising process through scene guidance. By the mode, the model can enlarge the targets of accurate balance and conflict, and high-efficiency orientation and control of the driving scene of multiple agents are realized. The technical scheme adopted is as follows: An automatic driving scene generation method based on multi-objective antagonism optimization comprises the following steps: step 1, constructing a noise ge