CN-122009240-A - Automatic driving control system and method based on world model and intelligent chassis

CN122009240ACN 122009240 ACN122009240 ACN 122009240ACN-122009240-A

Abstract

The invention provides an automatic driving control system and method based on a world model and an intelligent chassis, which belong to the field of automatic driving and comprise the steps of obtaining multisource data of an external environment, constructing a time sequence feature sequence according to the multisource data, carrying out explicit probability modeling according to the time sequence feature sequence to obtain scene features, generating a world query vector according to the scene features, carrying out prediction according to the world query and the time sequence feature sequence through the world model to obtain a future Gaussian scene state, generating a track point sequence according to the future Gaussian scene state, carrying out confidence assessment on the track point sequence to obtain the track point confidence, generating a control instruction according to the track point sequence and the track point confidence, carrying out coordinated control according to the control instruction to generate a final control vector, controlling the chassis through the final control vector, obtaining feedback data, carrying out assessment according to the feedback data and carrying out adjustment optimization on the final control vector and the world model to realize effective chassis control optimization.

Inventors

LIU QINGCHAO
YANG ANQI
CAI YINGFENG
WANG HAI
CHEN LONG

Assignees

江苏大学

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (8)

1. Automatic driving control system based on world model and intelligent chassis, characterized by comprising: the system comprises a multi-mode sensing module, a world model modeling and predicting module, a behavior decision and track generating module, a chassis control executing module and a feedback learning and state monitoring module; Acquiring multi-source data of an external environment through a multi-mode sensing module, and constructing a time sequence feature sequence according to the multi-source data; Generating a world query vector according to the scene characteristics, and predicting through a world model according to the world query and the time sequence characteristic sequences to obtain a future Gaussian scene state; Generating a track point sequence according to the future Gaussian scene state through a behavior decision and track generation module, and carrying out confidence assessment on the track point sequence to obtain a track point confidence; generating a control instruction according to the track point sequence and the track point confidence coefficient by a chassis control execution module, performing coordinated control according to the control instruction, generating a final control vector, and controlling the chassis by the final control vector; And acquiring feedback data through a feedback learning and state monitoring module, evaluating according to the feedback data, and adjusting and optimizing a final control vector and a world model.
2. The system of claim 1, wherein the system further comprises a controller configured to control the controller, In the multi-mode sensing module, the construction process of the time sequence feature sequence comprises the following steps: Calibrating the multi-source data, wherein the multi-source data comprises data acquired by a vision sensor, a laser radar, a millimeter wave radar and a global positioning system; performing feature extraction and cross-modal alignment on the calibrated data to obtain fusion features; and projecting the fusion features into a bird's eye view representation space, and combining adjacent frames to generate a time sequence feature sequence.
3. The system of claim 1, wherein the system further comprises a controller configured to control the controller, In the world model modeling and prediction module, the process of performing explicit probability modeling on the time sequence feature sequence comprises the following steps: carrying out display probability modeling on semantic units corresponding to the sequence of the sequential features by using Gaussian distribution to generate Gaussian scene representation, wherein the Gaussian scene representation comprises a position mean value, a covariance matrix and a semantic category label of the semantic units; dynamically updating Gaussian scene representation through a scene evolution mechanism to obtain scene characteristics, wherein the scene evolution mechanism comprises static alignment, dynamic modeling and region completion.
4. The system of claim 1, wherein the system further comprises a controller configured to control the controller, In the world model modeling and prediction module, the acquisition process of the future Gaussian scene state comprises the following steps: Acquiring semantic descriptions of Gaussian scenes, encoding the semantic descriptions into semantic vectors, and fusing the semantic vectors with scene features to generate world query vectors; And inputting the world query vector and the time sequence feature sequence as a world model, and carrying out recursive prediction of a future time sequence through the world model to obtain a future Gaussian scene state.
5. The system of claim 1, wherein the system further comprises a controller configured to control the controller, In the behavior decision and track generation module, the track point confidence coefficient generation process comprises the following steps: Reasoning is carried out through a semantic BEV-driven track intention modeling network according to the future Gaussian scene state and the time sequence feature sequence, and high-level driving intention is generated; and generating a track point sequence according to the high-level driving intention and the future Gaussian scene state, wherein the track point sequence comprises the position, the expected speed and the expected acceleration of the track point.
6. The system of claim 1, wherein the system further comprises a controller configured to control the controller, In the chassis control execution module, the generating process of the final control vector includes: And carrying out track analysis according to the track point sequence and the track point confidence, generating a control instruction, carrying out coordinated control through an intelligent chassis layered model prediction controller according to the control instruction, and generating a final control vector, wherein in the intelligent chassis layered model prediction controller, a corresponding cost function is calculated according to the error of the track point sequence and the control instruction, the cost function is minimized as a target, the control instruction is optimized, and the optimized control instruction is adjusted, so that the final control vector is generated.
7. The system of claim 1, wherein the system further comprises a controller configured to control the controller, In a feedback learning and state monitoring module, calculating errors of final control instruction execution and actual execution according to feedback data, wherein the errors of the control instruction execution adjust the final control instruction or re-plan a track, and the feedback data comprise actual steering angle, acceleration amount, braking force and real scene state of a vehicle; and adjusting and optimizing the world model according to the real scene state and the future Gaussian scene state.
8. The automatic driving control method based on the world model and the intelligent chassis is characterized by comprising the following steps of: acquiring multi-source data of an external environment, and constructing a time sequence feature sequence according to the multi-source data; generating world query vectors according to the scene features, and predicting through a world model according to the world query and the time sequence feature sequences to obtain future Gaussian scene states; Generating a track point sequence according to the future Gaussian scene state, and performing confidence assessment on the track point sequence to obtain a track point confidence; Generating a control instruction according to the track point sequence and the track point confidence, performing coordinated control according to the control instruction, generating a final control vector, and controlling the chassis through the final control vector; And acquiring feedback data, evaluating according to the feedback data, and adjusting and optimizing the final control vector and the world model.

Description

Automatic driving control system and method based on world model and intelligent chassis Technical Field The invention relates to the technical field of automatic driving, in particular to an automatic driving control system and method based on a world model and an intelligent chassis, and belongs to the technical direction of intelligent driving decision and control systems. Background Most of current automatic driving systems adopt modularized structures, and the functions of sensing, predicting, deciding, controlling and the like are designed in a scattered manner. Although the architecture has the definition of engineering realization, because the modules rely on fixed interfaces to transfer information, the system-level linkage and collaborative optimization are difficult to realize, and the problems of information lag, error amplification, inconsistent system response and the like are easy to occur in a dynamic traffic environment. Meanwhile, the modularized system is difficult to capture long-term time sequence dependency in the environment, and the generalization capability and the continuity understanding capability of the modularized system on complex driving tasks are limited. In order to improve the overall understanding capability of the system to the environment, the industry begins to explore a unified expression method with scene semantic modeling as a core, so that the system can construct an internal world representation based on a historical state and current observation and predict a future state. However, the existing method has the common problem of disconnection at the execution level, and the world modeling result is difficult to directly convert into high-precision and controllable bottom control behaviors. Meanwhile, the behavior decision module often lacks linkage feedback with the state of the actuator, and cannot perform online correction under complex working conditions, so that track deviation is enlarged and even safety risks are induced. In addition, the lack of a confidence evaluation mechanism for the track planning result also weakens the decision reliability of the system in unstructured scenes. In the aspect of executing control, the intelligent chassis system is used as a final bearing object of an automatic driving decision result, and the control precision, response timeliness and stability of the intelligent chassis system have decisive effects on driving safety. However, the existing chassis control system focuses on path tracking and unidirectional instruction execution, and lacks a joint control optimization mechanism facing the world state prediction result. Meanwhile, for the problems of fault accumulation and performance degradation in the long-term operation process of the actuator, the support of effective health monitoring and redundant fault tolerance strategies is lacking, and the stable and reliable operation of the system under the high-frequency control cycle is difficult to ensure. Therefore, a set of end-to-end control system from scene modeling to behavior reasoning and then to intelligent execution and feedback regulation is urgently needed to be constructed, and the cooperative modeling and closed-loop control of the world model and the intelligent chassis are truly realized. Disclosure of Invention In view of this, in order to solve the technical problems of control error amplification, world model semantic characterization and chassis execution parameter mismatch and executor health state non-reverse correction decision model caused by information transfer lag in the modularized automatic driving architecture, the invention provides an end-to-end automatic driving control system and method based on the collaborative modeling of the world model and the intelligent chassis, and the closed loop optimization for accurately executing the chassis understanding from the environment semantic is realized under the conditions of limited training data and limited calculation force by constructing a space-time aligned BEV-language-control joint coding system and a bidirectional dynamic feedback mechanism. In order to achieve the above object, the present invention provides an automatic driving control system and method based on a world model and an intelligent chassis, comprising: 1. Automatic driving control system based on world model and intelligent chassis, characterized by comprising: the system comprises a multi-mode sensing module, a world model modeling and predicting module, a behavior decision and track generating module, a chassis control executing module and a feedback learning and state monitoring module; Acquiring multi-source data of an external environment through a multi-mode sensing module, and constructing a time sequence feature sequence according to the multi-source data; Generating a world query vector according to the scene characteristics, and predicting through a world model according to the world query and the time sequence charact