CN-122024521-A - Automatic driving layered decision control method based on cloud vehicle coordination and digital twin closed loop

CN122024521ACN 122024521 ACN122024521 ACN 122024521ACN-122024521-A

Abstract

The invention discloses an automatic driving layered decision control method which utilizes a digital twin technology to carry out cloud training and verification and realizes strategy continuous optimization through a cloud vehicle collaborative architecture, and the method comprises the steps of constructing a driving primitive layered network architecture of a lane changing overtaking scene, constructing a cloud vehicle collaborative lane changing decision architecture which comprises a vehicle end sensing module, a cloud incremental global guiding IGG module, a vehicle end lane changing expert strategy model and a cloud continuous learning execution CLE engine.

Inventors

An Mande
LIU CHANG
YANG HAO
HAN YIMEI
LIU SHIYU
SHI JUNLING
SUN YUNHE
ZHAO LIANG

Assignees

沈阳航空航天大学

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (5)

1. The automatic driving lane-changing overtaking control method based on cloud vehicle cooperation and digital twin closed loop is characterized by comprising the following steps: Constructing a driving primitive layered network architecture of a lane-changing overtaking scene; The cloud vehicle collaborative lane change decision framework is constructed, and comprises a vehicle end sensing module, a cloud end increment global guiding IGG module, a vehicle end lane change expert strategy model and a cloud end continuous learning execution CLE engine; Constructing a lane change overtaking expert strategy model BERL Agent based on the driving primitive layered network architecture and the cloud vehicle cooperative lane change decision architecture; Training and verifying the lane change overtaking expert strategy model BERL Agent, deploying the BERL Agent strategy passing verification to a real vehicle, monitoring the running state in real time, and automatically intercepting data to return to the cloud when a lane change failure or a manual takeover of the equal-length tail scene is encountered, so as to trigger a new round of training closed loop.
2. The automatic driving lane-changing overtaking control method based on the cloud vehicle cooperation and the digital twin closed loop according to claim 1, wherein the driving primitive layered network architecture for constructing the lane-changing overtaking scene comprises the following steps: the primitive dividing principle taking lane change as a core is established, namely, the continuous driving process is decoupled into independent driving primitives according to the frequency and the safety characteristics of driving tasks, and the lane change/overtaking is established as independent tactical layer Tier 2 driving primitives, so that the driving primitives are separated from steady driving and emergency driving; according to the primitive dividing principle, a hierarchical primitive system is constructed, and the priorities are arranged as follows from high to low: Tier 0-System status primitive, covering non-driving status failure, static; Tier 1-safety covering primitive, which covers emergency braking and danger avoiding steering and is used as a safety bottom when lane change fails or is in danger; Tier 2-tactical transition primitives, namely covering lane changes, overtaking and remittance; tier 3-steady state driving primitive, which covers lane keeping and following as a default driving state; and establishing a preemption mechanism of the high-priority primitive to the low-priority primitive, wherein the Tier 1 safety primitive can immediately interrupt the execution of the Tier 2 lane change primitive if collision risk is detected in the lane change process.
3. The automatic driving lane-changing overtaking control method based on the cloud car cooperation and the digital twin closed loop according to claim 1, wherein, The vehicle end perception module is used for extracting a state perception representation special for lane change and taking the state perception representation as the input of a lane change expert model, wherein the state perception representation special for lane change comprises a dynamics state of the vehicle and feature vectors of surrounding key area interaction vehicle information; Deploying an incremental global guidance IGG module at the cloud end, which is used for aggregating macroscopic traffic flow data uploaded by road sides and motorcades, calculating a lane-level global state index and generating a global guidance signal for correcting a vehicle-end strategy; Constructing an expert model specially responsible for lane changing and overtaking tasks based on a reinforcement learning network, wherein the model receives state perception characterization and outputs a transverse lane changing and longitudinal speed control instruction; And deploying a cloud end continuous learning execution CLE engine as an admission controller of a strategy version at the cloud end, integrating a hard security rule checking and statistical confidence verification function, and managing updating and release of a lane change expert strategy.
4. The automatic driving lane-changing overtaking control method based on the cloud car cooperation and the digital twin closed loop as claimed in claim 1, wherein the lane-changing overtaking expert policy model BERL Agent training process comprises: Constructing a high-fidelity lane change simulation environment; And executing collaborative training based on IGG (intelligent gateway) guidance, namely dynamically correcting the rewarding function by introducing a global guidance signal generated by an IGG module, and taking the individual efficiency and the global traffic flow stability into consideration when the guidance model explores the lane change strategy.
5. The cloud vehicle cooperation and digital twin closed loop-based automatic driving lane-changing overtaking control method according to claim 4, wherein the trained lane-changing overtaking expert strategy model BERL Agent is verified in two stages by using a CLE engine, wherein the security rule hard constraint check is firstly carried out, the performance contrast test based on a confidence interval is then carried out, and the strategy update is only allowed when the performance lower bound of the new strategy is better than the performance upper bound of the old strategy.

Description

Automatic driving layered decision control method based on cloud vehicle coordination and digital twin closed loop Technical Field The invention belongs to the technical field of automatic driving and artificial intelligence, and particularly relates to an automatic driving layered decision control method based on cloud vehicle cooperation and digital twin closed loop. Background The core goal of an autopilot system is to achieve safe and efficient autonomous navigation in complex dynamic traffic environments. However, the prior art still faces significant challenges in architecture design and evolution model, mainly in the following aspects: 1. limitations of existing architecture: Although the traditional modularized Approach (module Approach) has interpretability, each module is independently optimized, error accumulation is easy to cause, the system is fragile, and complex and changeable interaction scenes are difficult to deal with. The emerging End-to-End Approach (End-to-End Approach), while having powerful data fitting capabilities, has serious "long tail distribution" problems. 99% of actual driving is a conventional scene (such as following a car), and only 1% is a high-risk long-tail scene (such as lane-changing overtaking). A single model trains on extremely unbalanced data, often ignoring these low frequency high risk scenarios, and its "black box" nature makes security difficult formal verification. 2. Lack of synergy from a global traffic perspective: The existing decision algorithm is mostly optimized based on local observations of the vehicle (such as speed and TTC of the vehicle). However, roads are shared resources, individual optimality often results in global suboptimal (e.g., frequent lane changes for a bicycle may result in ghost traffic congestion behind). The existing method lacks a mechanism, can feed back Global traffic flow index (Global TRAFFIC METRICS) to a bicycle in real time, and guides the bicycle to achieve stability of system-level traffic flow while pursuing self efficiency. 3. Lack of closed loop sustained evolution mechanism: the traditional training, namely deployment, mode is static and cannot adapt to the dynamic change of the environment (such as road conditions), and the direct online learning on a real vehicle has huge potential safety hazards and faces the risk of catastrophic forgetting (Catastrophic Forgetting). In summary, aiming at the problem that the safety and the passing efficiency are difficult to be compatible in the high-risk and complex transitional driving task of the automatic driving vehicle, the decision optimization method based on the cloud vehicle cooperation and the digital twin closed loop is provided, and the problem to be solved is to be solved. Disclosure of Invention In view of the above, the invention aims to provide an automatic driving layered decision control method based on cloud vehicle cooperation and digital twin closed loop, so as to solve the problem of safety and efficiency balance of an automatic driving vehicle in transitional driving tasks such as highway overtaking, lane changing and the like. The technical scheme provided by the invention is that the automatic driving lane-changing overtaking control method based on cloud vehicle cooperation and digital twin closed loop comprises the following steps: Constructing a driving primitive layered network architecture of a lane-changing overtaking scene; The cloud vehicle collaborative lane change decision framework is constructed, and comprises a vehicle end sensing module, a cloud end increment global guiding IGG module, a vehicle end lane change expert strategy model and a cloud end continuous learning execution CLE engine; Constructing a lane change overtaking expert strategy model BERL Agent based on the driving primitive layered network architecture and the cloud vehicle cooperative lane change decision architecture; Training and verifying the lane change overtaking expert strategy model BERL Agent, deploying the BERL Agent strategy passing verification to a real vehicle, monitoring the running state in real time, and automatically intercepting data to return to the cloud when a lane change failure or a manual takeover of the equal-length tail scene is encountered, so as to trigger a new round of training closed loop. Preferably, the driving primitive layered network architecture for constructing the lane change overtaking scene includes: the primitive dividing principle taking lane change as a core is established, namely, the continuous driving process is decoupled into independent driving primitives according to the frequency and the safety characteristics of driving tasks, and the lane change/overtaking is established as independent tactical layer Tier 2 driving primitives, so that the driving primitives are separated from steady driving and emergency driving; according to the primitive dividing principle, a hierarchical primitive system is constructed, and the prioriti