World model and network slice closed-loop based automatic driving chassis control method

Abstract

The application provides an automatic driving chassis control method based on a world model and a network slice closed loop, and belongs to the field of intelligent driving, and comprises the following steps: acquiring multi-modal data in a traffic scene through perception, encoding the multi-modal data to obtain a hidden state, performing future state prediction and control generation on the hidden state through a world model to obtain a future state and a control instruction, optimizing the world model according to the future state, constructing a digital twin scene, receiving and controlling the automatic driving chassis according to the control instruction in the digital twin scene, and acquiring multi-modal data through perception, transmitting corresponding data flow of the perceived, predicted future state and control instruction to a network slice resource, and dynamically adjusting and configuring the network slice resource, and online optimizing the world model and the network slice resource according to a feedback mechanism to realize multi-step high-precision environment prediction and chassis dynamic response while meeting end-to-end low latency.

Description

Technical Field

[0001] This invention relates to the fields of intelligent driving and intelligent chassis control technology, and more specifically, to an autonomous driving chassis control method based on a world model and network slicing closed loop. Background Technology

[0002] With the synergistic development of autonomous driving technology and chassis intelligence, the requirements for the real-time performance, robustness, and safety of vehicle perception, prediction, decision-making, and chassis execution are increasing. However, existing autonomous driving systems often design perception, trajectory prediction, network communication, decision-making, and chassis execution modules in isolation, lacking a unified end-to-end closed loop and chassis digital twin integration. This makes it difficult to simultaneously achieve low latency, multi-step high-precision prediction, chassis dynamic response, and system robustness in complex traffic environments.

[0003] At the network communication level, although breakthroughs have been made in 6G autonomous network management and slicing technology in recent years, its deep integration with autonomous driving control has not yet achieved closed-loop simulation. "Network slicing" can pre-deploy dedicated resources on the same physical network to ensure ultra-low latency and high reliability: related research has proposed an AI-driven slicing self-optimization framework that can monitor indicators such as latency and bandwidth and reconfigure online; there are also studies using secure transfer learning in non-terrestrial networks to ensure quality of service. However, currently, it is not deeply coupled with chassis execution control in the same digital twin environment, making it impossible to quantify and optimize the impact of network jitter on chassis execution during the development phase.

[0004] In terms of data fusion and security, autonomous driving systems need to simultaneously process heterogeneous information from multiple sources, such as cameras, radar, and LiDAR, to obtain more comprehensive and robust scene cognition. However, centralized fusion training models pose significant privacy risks: sensitive data such as vehicle trajectories and surrounding environmental features may be misused or leaked in the cloud. Although federated learning and differential privacy frameworks have attempted to address these issues, such as differential privacy protection strategies that integrate "camera-LiDAR-radar" joint training, their performance and latency trade-offs in "edge-cloud" collaborative scenarios have not been fully studied. Furthermore, existing world models have limited generalization capabilities in visually limited environments such as rain, fog, and nighttime, or in high-density traffic scenarios. Prediction errors increase significantly in these extreme scenarios, affecting the robustness of control decisions and also limiting the stability of chassis control.

[0005] In summary, traditional autonomous driving systems face multiple challenges in terms of real-time performance, predictive continuity, system robustness, and data security: First, segmented design and single-step prediction lead to high latency and poor foresight in the perception-decision cycle; second, although network slicing technology and digital twin simulation have developed rapidly, there is a lack of a closed loop that deeply couples the two with autonomous driving control; finally, the privacy protection mechanism of multimodal fusion is still imperfect, and its performance in extreme scenarios needs to be strengthened. Summary of the Invention

[0006] In view of this, this invention proposes an autonomous driving chassis control method based on a world model and network slicing closed loop. It aims to achieve high-precision environmental prediction, chassis dynamic response and data security in multiple steps while meeting the requirements of low latency at the end-to-end autonomous driving architecture that integrates multi-step world model, network slicing and digital twin. This provides solid support for intelligent chassis autonomous driving in high-density urban roads and highway scenarios in the future.

[0007] To achieve the above objectives, this invention proposes an autonomous driving chassis control method based on a world model and network slicing closed loop, comprising: The system senses and acquires multimodal data in a traffic scenario, encodes the multimodal data, and obtains the hidden state. The world model is used to predict and generate future states and control commands based on the hidden states, thereby obtaining future states and control commands; the world model is then optimized based on the future states. A digital twin scenario is constructed, in which the autonomous driving chassis is controlled according to the control commands, and multimodal data is perceived and acquired. Specifically, the data streams corresponding to the perception, prediction of future states, and control command transmissions are mapped to network slice resources, and the network slice resources are dynamically adjusted and configured. The world model is optimized through federated training using a distributed federated training method; the world model and network slice resources are optimized online based on a feedback mechanism.

[0008] Optionally, the process of encoding the multimodal data includes: The multimodal data is encoded separately to obtain corresponding modal feature vectors, wherein the multimodal data includes image data, point cloud data, radar echo data, and map data; The hidden states are obtained by performing HERMES multimodal fusion on the modal feature vectors.

[0009] Optionally, the process of predicting and controlling the future state of the hidden state includes: The predicted hidden state at the current time step is obtained by using the world model to predict the hidden state at the previous time step based on the hidden state at the previous time step, the control instructions at the previous time step, and the auxiliary features. A joint objective function is constructed, which includes the hidden state prediction error and the chassis control cost. With the goal of minimizing the joint objective function, the future state is predicted and the optimal control command is generated by combining the world model, thus obtaining the future state and the control command.

[0010] Optionally, the process of optimizing the world model based on future states includes: The real-time observation state is obtained, and the error loss at the same time step and the corresponding next time step is applied based on the real state and the future state to update the model parameters of the world model.

[0011] Optionally, the digital twin scenario includes a traffic environment digital twin and an intelligent chassis digital twin. The traffic environment digital twin maps the real road environment and obtains multimodal data of the real road environment in real time. The intelligent chassis digital twin simulates the behavior of the chassis based on the vehicle dynamics model of the chassis and according to control commands.

[0012] Optionally, the process of dynamically adjusting and configuring the network slice resources includes: Obtain the end-to-end latency and bandwidth utilization of network slice resources, and update the resource share of different network slice resources based on the end-to-end latency and bandwidth utilization. Configure network slice resources using Oracle's secure migration learning method.

[0013] Optionally, the process of online optimization of the world model and network slice resources includes: Performance metrics are obtained through a feedback mechanism, including end-to-end latency throughout the entire process of sensing control execution, multi-step prediction error between future state and actual observation, and deviation between the actual motion trajectory of chassis control and the predicted control path. Monitor the overall convergence status based on the aforementioned performance indicators; The world model parameters are updated using a gradient method based on the multi-step prediction error and the time backpropagation method; the resource share of the network slice resources is adjusted based on the end-to-end latency throughout the process; until the performance index is within the safety threshold.

[0014] On the other hand, this invention proposes an autonomous driving chassis control system based on a world model and network slicing closed loop, for executing the above-mentioned method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) A deep integration of prediction, communication and control is proposed as an end-to-end closed-loop architecture that deeply couples high-fidelity digital twins with AI-driven network slice management. By real-time monitoring and decoupling feedback of end-to-end latency, prediction error and chassis motion deviation, the physical impact of network jitter on chassis execution can be quantified and optimized during the development stage. Combined with Oracle's secure transfer learning mechanism, the end-to-end latency of the entire link is reduced, and the robustness and control stability of the system in complex or mutating network environments are improved.

[0016] (2) An integrated prediction and decision-making method that embeds a world model into model predictive control (MPC) is proposed. Through a dual-loop mechanism of online planning and background self-supervision, the world model is used as a differentiable simulator to solve the prediction of future multi-step environmental evolution and the optimal control command sequence in one go. This improves the forward-looking nature of the system's decision-making and the high-precision trajectory tracking capability in complex dynamic scenarios. Furthermore, through continuous learning in the background, the model can achieve adaptive evolution to new environments.

[0017] (3) A multimodal federated learning framework (PHMS-Fed) with differential privacy is adopted. Under the premise of protecting the security of sensitive data such as vehicle driving trajectory and chassis condition, it realizes efficient distributed collaborative training of the global world model, effectively solves the data privacy leakage risk brought about by traditional centralized training, and significantly enhances the generalization ability and prediction accuracy of the world model in extreme scenarios such as rain, fog and night by using diverse data collected from massive edge nodes. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of the end-to-end autonomous driving intelligent chassis control system in an embodiment of the present invention; Figure 2 This is a diagram illustrating the internal working mechanism of the world model-driven integrated prediction and decision-making module in this embodiment of the invention. Figure 3 This is a schematic diagram of the workflow of the online self-optimization module for network slicing in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the distributed federated training and privacy protection framework in an embodiment of the present invention. Figure 5 This is a schematic diagram of the human-computer interaction interface of the digital twin closed-loop evaluation system in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] This invention proposes an end-to-end intelligent chassis control method for autonomous driving based on a world model and network slicing closed loop. By deeply integrating multimodal perception, integrated prediction and decision-making, dynamic network resource scheduling, and distributed federated learning, a complete technical solution is constructed that can simultaneously satisfy low latency, high accuracy, strong robustness, and data security. The method mainly includes the following steps: S1. Multimodal data preprocessing and latent state encoding: Using multi-source time-series data from front-end cameras, LiDAR, RADAR and high-definition maps, after data synchronization, standardization and noise reduction, a multimodal encoder in the HERMES architecture is used to generate a fused latent state vector, and a distillation and pruning quantization strategy is used to obtain a lightweight encoder group.

[0021] S2. Integrated Spatiotemporal Prediction and Decision-Making Based on World Model: This step employs the "Prompting with the Future MPC" (PWTF) framework, using the world model as a differentiable environmental dynamics model for decision-making and planning. It outputs the predicted hidden states of the environment and the optimal control command sequence for the next N frames in a single step. To ensure the long-term stability of the prediction model, a "dual autoregressive" self-supervised mechanism is further incorporated to enhance the model's robustness to predictions of complex stochastic dynamic environments.

[0022] S3. Digital Twin Scenarios and Slice Deployment: Real traffic scenarios are reproduced in a high-fidelity simulation platform (such as CARLA), that is, the digital twins of the traffic environment and the intelligent chassis are integrated, and dedicated network slice resources are pre-configured according to the business flow (sensing data, status data, control commands).

[0023] S4. Online self-optimization of slices: Drawing on the 6G autonomous network management framework, for four types of services—"sensing data flow, status data flow, control command flow, and chassis execution flow"—the end-to-end latency and bandwidth usage of each slice are monitored in real time, and resources are dynamically reconfigured using AI-driven scheduling strategies. At the same time, Oracle's secure migration learning is applied, that is, an optimal slice configuration strategy library for different network conditions is pre-built as an Oracle source, and when link quality changes suddenly, expert strategies are used to guide the current strategy to quickly generate new slice configuration vectors.

[0024] S5. Distributed Federated Training and Privacy Protection: The PHMS-Fed framework is adopted to update the perception sub-models such as cameras, LiDAR, and RADAR locally and in parallel at each edge node. At the same time, it combines chassis feedback data to form multimodal gradients. The gradients are aggregated into the global world model through a differential privacy mechanism to prevent the leakage of sensitive information including driving trajectory and chassis condition.

[0025] S6. Closed-loop performance evaluation and multi-dimensional parameter iteration: In the digital twin platform, a full-process simulation test of "perception → multi-step prediction → slice communication → decision-making → chassis simulation" is performed. Key indicators such as end-to-end latency, trajectory prediction error and chassis motion deviation are collected. The single hybrid gradient update mode is abandoned and a hierarchical decoupled optimization strategy is adopted: gradient-based accuracy update is performed for the world model and resource correction based on feedback control is performed for the network slices until all indicators meet the preset thresholds, so as to realize the adaptive convergence of the system in communication-constrained and dynamic environments.

[0026] Through the above steps, this invention organically combines multi-step world model prediction, digital twin network slicing closed loop, and federated multimodal privacy fusion in a modular manner. It not only meets the real-time requirements of end-to-end low latency and future multi-step high-precision environmental prediction, but also significantly improves the robustness and data security of the system in complex traffic scenarios through secure and controllable resource scheduling and privacy protection measures.

[0027] Furthermore, in step S1, in order to generate a unified fusion hidden state, the present invention first processes the camera image... LiDAR point cloud RADAR echo and high-definition map features Process in the following order: Based on a unified GPS timestamp, the timestamps of each sensor are recorded. Collected data Mapped to the same reference coordinate system, and any missing frames are filled in by linear interpolation to ensure the synchronization of input data; Image First normalized to It will then be passed through a convolutional encoder. Generate image vectors ; Point cloud Through voxelization and encoder Obtain point cloud vectors ; RADAR data Through multilayer perceptron Output echo vector ; Map features Through graph convolutional mesh Embedded as map vector .

[0028] In HERMES multimodal fusion, the four modal features are aggregated along the feature dimension through a concatenation operation to generate a high-dimensional fused feature vector. :

[0029] The high-dimensional concatenated vector is then input into a fully connected layer (MLP). Nonlinear mapping is performed to extract deep correlations between modes and reduce the dimensionality to a unified latent space, ultimately generating a fused latent state vector. :

[0030] Furthermore, to meet the constraints of real-time inference in both vehicle and edge environments, the fusion encoder simultaneously satisfies the following lightweight constraints and achieves them in an integrated manner during the design and training process.

[0031] Using a knowledge distillation constraint strategy, let the output of the teacher encoder be... The student encoder output is Introducing the distillation term: , Will Joint optimization with the main task loss ensures that the student encoder reduces the number of parameters and computational complexity while maintaining semantic consistency.

[0032] Measurable pruning strategies are applied to the encoder's channels, attention heads, and other structures, and weights and activations are mapped to low-bit-width representations using Quantization-Aware Training (QAT), thereby reducing memory and computation load on edge devices.

[0033] Lightweight encoder output Compared with the original definition Maintain consistency in dimensions and semantic interfaces.

[0034] This step aims to generate the time. Fusion hidden state and The lightweight hidden state The observations used as the latest frame, together with the previous historical hidden states, form a time series, which serves as the input for step S2 to meet the low-latency, real-time prediction requirements of the vehicle-mounted terminal, while maintaining high accuracy. This step is used for cloud-based training or high-fidelity simulation verification. Through this step, a high-dimensional, continuous environmental representation that can be computed in real time by edge devices is generated for subsequent multi-step prediction and control, forming a solid foundation for the entire end-to-end low-latency closed loop.

[0035] Furthermore, in step S2, the "Prompting with the Future MPC" approach is applied to multi-step prediction, employing a dual closed-loop mechanism of "online planning + background self-supervision" to achieve coupled solution of future environment simulation and vehicle control strategy: Construct a prediction-decision coupled optimization problem, treating the world model as a differentiable environment simulator, and build a joint objective function that includes state prediction error and control cost: , , in The hidden state predicted by the model. Guided to the desired navigation path or safe state. For the pre-trained world model dynamic equations, predict the step size. State weight matrix : Among them It is an identity matrix with dimension 1. That is, consistent with the hidden state dimension; control weight matrix : Used to penalize abrupt changes in control input to ensure control smoothness; a tradeoff coefficient between prediction accuracy and control input overhead. , Features for future scenarios are generated by a cue network based on historical hidden states and a map. Specifically, the hidden states predicted by the model at time t are... For step 1 generated , It is obtained through the control command at time t-1.

[0036] Specifically, the parameters for the hidden states used in predicting future states are set at time t using subsequent methods: At each time step First, the latest hidden state observed in step 1 will be collected. The historical hidden state sequence, consisting of the historical hidden states and their preceding time steps, is used to capture the velocity, acceleration, and motion trends of all elements in the scene through a self-attention mechanism. Specifically, the internal Transformer encoder infers the initial context state, which contains complete dynamic information of the preceding historical hidden state sequence, based on the historical hidden state sequence. As the starting point for future multi-step predictions, the context state is... As For the initial The control command issued to the chassis at time t-1 is a historical action that has already been executed.

[0037] Subsequently, the system solves the above optimization equations using gradient descent or stochastic optimization algorithms to find a set of optimal future control sequences that minimize the cost function. In this solution process, the world model It will be repeatedly called to simulate future evolution under different control sequences, such as the simulation in the first step. .

[0038] After optimization, the system outputs two sets of sequences at once: the future hidden state sequence. and optimal control command sequence .

[0039] To enhance robustness to stochastic dynamic environments, the actual states observed after actual operation are utilized. Calculate the self-supervised loss and update the model parameters asynchronously. : , in Balancing multi-step prediction with autoregressive consistency, this process is performed in the background or cloud, enabling the model to continuously learn long-term dependency characteristics and ensuring... It can continuously adapt to new environmental evolution patterns.

[0040] Furthermore, in step S3, the present invention constructs an end-to-end digital twin platform: Digital twin of traffic environment - using simulation tools such as CARLA to recreate multiple scenarios such as urban roads, highways, and overpasses, and generate static and dynamic elements consistent with real roads; Intelligent chassis digital twin – Based on the vehicle dynamics model, a simulation subsystem for torque distribution, suspension response and tire friction is established to ensure that the chassis execution results and control commands maintain physical consistency.

[0041] At the same time, the "sensory data stream" is mapped to the sensing slice. It is responsible for ensuring the transmission of multimodal raw data from various sensors to multimodal encoders; the "state data stream" is mapped to state slices. It is responsible for ensuring the real-time hidden state after fusion from the multimodal encoder to the world model prediction and decision module. Transmission; "Control Data Stream" mapped to control slice It is responsible for ensuring the optimal control commands from the world model prediction and decision-making module to the chassis actuator. )transmission.

[0042] Each slice has pre-allocated bandwidth during deployment. Delay budget Computing resources This creates a complete closed-loop simulation environment. The output is a closed-loop simulation environment that can simultaneously support perception, prediction, decision-making, and chassis execution.

[0043] Furthermore, in step S4, the present invention performs dynamic scheduling and secure reconfiguration of four types of slice resources: Conduct AI-driven online self-optimization, that is, monitor each slice in real time. "end-to-end latency" With bandwidth utilization core resource share Updates are primarily latency-driven and follow these rules: , in, This represents the adaptive scheduling step size coefficient, with a preset upper limit on delay. As specified in the system design, when the real-time latency exceeds the target value, the resource share is increased in a positive direction to reduce the latency; otherwise, redundant resources are released.

[0044] Based on this, bandwidth utilization As an auxiliary judgment condition: if the time delay already meets the requirements ( )but If the value remains too low, the system will appropriately reduce it. To free up redundant resources.

[0045] Furthermore, when the system detects a sudden change in the network environment—based on a sharp deviation of key performance indicators from their historical moving averages within a short time window (100ms)—Oracle's security migration learning mechanism will be immediately triggered. Specific triggering conditions include, but are not limited to, bandwidth utilization. A sudden drop occurs, such as the current value being less than 30% of its average value over the past second; end-to-end latency. A sharp spike, such as when the current value is more than three times its average value over the past second and exceeds an absolute safety threshold (80ms).

[0046] When Oracle's security migration learning mechanism is triggered for emergency reconfiguration, the following strategy is used to quickly generate new slice configuration vectors. : , in, Indicates the adaptive scheduling configuration coefficient. Configure vectors (source domain policies) for the expert / Oracle slices retrieved from the provisioned library that best match the current operating conditions. This generates a configuration vector (target domain policy) for the current online learning algorithm. Through this weighted interpolation, expert policies guide the current policy, quickly generating new slice configuration vectors. This ensures that the Service Level Agreement (SLA) for sliced ​​slices is not affected.

[0047] In step S5, multimodal privacy and security fusion is achieved: Each edge node Using locally collected data, based on the self-supervised loss function defined in step S2 Calculate local loss And solve for its gradient with respect to the local model parameters. Each edge node Parallel update of local model parameters using the PHMS-Fed framework .

[0048] When updating global parameters, add Gaussian noise to satisfy the requirements. -Differential Privacy: , accomplish - Differential privacy protection effectively prevents the leakage of sensitive information such as driving trajectory and chassis condition. in and They are nodes Compared with the total number of samples.

[0049] This process enables each node to collaboratively learn the world model while protecting vehicle privacy and chassis operation data security.

[0050] Furthermore, in step S6, closed-loop verification and adaptive optimization are performed on the overall system: The following three core metrics are monitored in real time within the complete closed loop: end-to-end latency, multi-step prediction error, and chassis motion deviation.

[0051] Among them, end-to-end delay This reflects the delay from the moment sensors acquire data until the chassis executes control commands to complete the entire process. The acquisition process is as follows: each time a complete "perception-encoding-prediction-decision-control-chassis feedback" closed loop is executed, the system records the start timestamp. Perception data frame generation, recording end timestamp The chassis actuator receives and responds to control commands, then the end-to-end delay is... After multiple closed-loop cycles, the system statistically calculates the average and maximum values ​​to measure performance.

[0052] Multi-step prediction error: H represents the number of future time steps in a single prediction. These are the hidden state vectors obtained from actual observations and model predictions, respectively; Chassis motion deviation : Measures the deviation between the vehicle's actual trajectory (feedback from the chassis) and the predicted control path, defining the vehicle's... The actual location at any given time is The predicted trajectory points are ;sampling After the frame, define the motion bias: , Represents Euclidean distance; This directly reflects the accuracy of the "control-chassis" closed loop. Large slice delays or high prediction biases will lead to deviations. Significant increase.

[0053] Construct a comprehensive loss index to monitor the overall convergence status of the system: , All are weighting coefficients. During parameter updates, the following physically feasible decoupling iteration strategy is executed: Utilizing multi-step prediction error The world model parameters in step S2 are processed using the time backpropagation algorithm. Gradient updates are performed to improve the accuracy of the model's environmental extrapolation in specific traffic scenarios:

[0054] Due to end-to-end delay Since the network parameters are not differentiable, a feedback control mechanism is used to correct the slice resource allocation strategy in step S4. If it is detected that... If the latency exceeds the limit, then the adaptive scheduling step size coefficient defined in step S4 will be used. Or directly add specific slices resource share :

[0055] That is, by increasing the adjustment sensitivity or allocating more bandwidth resources, the communication latency is reduced at the physical level.

[0056] Repeat the above process until... , and At the same time, if the values ​​are below the preset safety threshold, it indicates that the parameter adaptation of the entire "sensing-transmission-prediction-control" link is complete, forming a stable adaptive closed loop.

[0057] The above technical solution is described in conjunction with the relevant accompanying drawings: like Figure 1 As shown, the end-to-end autonomous driving intelligent chassis control system and method based on world model and network slicing closed loop described in this invention mainly includes a multimodal perception and communication assurance module, a prediction and decision-making integrated generation module, and a closed-loop iteration and federated learning module.

[0058] The multimodal perception and communication assurance module is responsible for the real-time, low-latency acquisition and processing of environmental information by the system. Specifically, in step S3, within the digital twin environment, the system acquires time-series data streams in real time through sensors (cameras, radar, lidar, and high-definition maps). In step S1, the multimodal encoder performs spatiotemporal alignment, fusion, and encoding of the heterogeneous data streams into a unified lightweight latent state. The entire data transmission process is ensured by the S4 network slice management module, which monitors the communication links of each data stream (sensing, status, and control) in real time and dynamically adjusts slice resources through an AI-driven online self-optimization strategy to ensure low-latency and high-reliability transmission of critical information.

[0059] The integrated prediction and decision generation module is the core of the system, responsible for the integrated solution of environmental prediction and vehicle control. This module uses... As input, the PWTF-MPC framework is used in S2 to input the pre-trained world model. As a differentiable environmental dynamics simulator, this module can generate two key outputs at once by solving an optimization problem that includes state tracking error and control costs: one is a multi-step predicted hidden state sequence for assessing future risks. Secondly, optimal control commands are directly issued to the chassis actuators. This enables end-to-end coupling from perception to decision-making.

[0060] The closed-loop iteration and federated learning module is responsible for the system's continuous learning and adaptive optimization. This module contains two parallel closed loops: the first is the closed-loop evaluation and decoupling iteration in step S6, which collects end-to-end latency data in real time. Prediction error and chassis motion deviation Heterogeneous performance metrics, through a decoupled feedback mechanism, are used to fine-tune the network slice scheduling strategy in step S4 and the world model parameters in step S2 online, respectively; the second step is federated training, which collects data from a large number of edge nodes (vehicles) and performs distributed asynchronous training and global model updates on the world model in step S2 under the premise of protecting user data security through differential privacy, thereby continuously improving the model's generalization ability in various scenarios.

[0061] The end-to-end autonomous driving intelligent chassis control method described in this invention can be divided into six collaboratively executed steps, such as... Figure 1 As shown, the processing flow is as follows: S1. Multimodal Data Preprocessing and Latent State Encoding—Acquiring and fusing multi-source sensor data, and generating a unified latent state through a lightweight encoder. .

[0062] S2. Integration of Spatiotemporal Prediction and Decision Planning Based on World Model – Within the MPC framework, the optimal control command is solved in one step using the world model. and future environmental predictions .

[0063] S3. Digital Twin Scenario and Slice Deployment - Run the entire system in a high-fidelity simulation environment and pre-define dedicated network slice resources for each data stream.

[0064] S4. Online self-optimization of slices—real-time monitoring of communication link quality, dynamic reconfiguration of network resources through AI scheduler and Oracle migration learning mechanism to ensure communication service level.

[0065] S5. Distributed Federated Training and Privacy Protection – Through a federated learning framework, multi-vehicle data is utilized and differential privacy is combined to enable secure distributed training and updating of the world model.

[0066] S6. Closed-loop performance evaluation and multi-dimensional parameter iteration—Collect key performance indicators of the system and optimize the world model and slice scheduling strategy online through a decoupled feedback mechanism.

[0067] Figure 2 The diagram shown illustrates the integrated prediction and decision-making mechanism driven by the world model, detailing the internal workings of step S2. This mechanism consists of two parallel processes: Online planning (real-time) process: This process is executed in real time while the vehicle is running. It receives three types of inputs: historical state from S1, Prompt network, and navigation system. The core PWTF-MPC solver then processes the world model... As an internal simulator, it aims to minimize the cost function (Cost) and solves for the optimal control command through an optimization algorithm. (Used to control the vehicle) and the corresponding predicted state (Used for evaluation and background training).

[0068] Background training (asynchronous) process: This process is used for continuous learning of the model. It receives the predicted state from the online planning output. and the real state observed from the digital twin environment In the loss calculation unit, the loss function between the two is calculated, and the gradient used to update the model parameters is generated accordingly. Ultimately, the optimizer optimizes the world model parameters. Update.

[0069] Figure 3 The diagram illustrating the online self-optimization mechanism for network slicing details the internal workflow of step S4. The monitoring module continuously collects network performance data, which first enters the link mutation detection module. If no mutation is detected, the system proceeds to the "No" path, sending the current status information to the AI ​​scheduler to perform routine, feedback-controlled resource fine-tuning. If a mutation is detected, the system proceeds to the "Yes" path, triggering a migration trigger (Oracle) to generate a high-priority coverage configuration using pre-defined expert strategies. The configuration schemes generated by both paths are ultimately updated uniformly in the slice configuration database, which then distributes the latest slice parameters to the actual communication links.

[0070] Figure 4 The flowchart illustrating the distributed federated training and privacy protection process details the working mechanism of step S5. During the local training phase, each edge node uses local data to calculate the model update amount, as shown in the diagram as "Calculating Local Gradients". Before uploading, Gaussian noise is added through the privacy protection module to meet differential privacy requirements. In the global aggregation phase, the central server collects the noisy updates from all nodes, performs weighted merging of sample weights, and calculates the new global model parameters. Finally, the updated global parameters are broadcast back to each edge node via the distribution channel to begin the next round of local training.

[0071] Figure 5 The schematic diagram of the digital twin closed-loop evaluation system interface shown illustrates a concrete and implementable monitoring and debugging platform of this invention. The interface consists of four core panels: The real-time key performance indicator panel provides real-time visualization of key performance indicators such as end-to-end latency, RMSE, and chassis deviation angle monitored by S6.

[0072] The slice status panel shows that step S4 is for different data streams (sensory data streams). Status data stream Control data flow The allocated bandwidth, latency, and other resources and their priorities.

[0073] The simulation scene view panel presents a realistic traffic scene generated by the digital twin environment in step S3.

[0074] The Status and Debug panel provides human-computer interaction functions. Operators can test the emergency response mechanism of step S4 by using the "Trigger Transfer Learning" button, or manually trigger a round of federated learning aggregation of step S5 by using the "Re-aggregate" button, thereby conducting comprehensive testing and verification of the system.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An autonomous driving chassis control method based on world model and network slicing closed loop, characterized in that, include: The system senses and acquires multimodal data in a traffic scenario, encodes the multimodal data, and obtains the hidden state. The world model is used to predict and generate future states and control commands based on the hidden states, thereby obtaining future states and control commands; the world model is then optimized based on the future states. A digital twin scenario is constructed, in which the autonomous driving chassis is controlled according to the control commands, and multimodal data is perceived and acquired. Specifically, the data streams corresponding to the perception, prediction of future states, and control command transmissions are mapped to network slice resources, and the network slice resources are dynamically adjusted and configured. The world model is optimized through federated training using a distributed federated training method; the world model and network slice resources are optimized online based on a feedback mechanism.

2. The method according to claim 1, characterized in that, The process of encoding the multimodal data includes: The multimodal data is encoded separately to obtain corresponding modal feature vectors, wherein the multimodal data includes image data, point cloud data, radar echo data, and map data; The hidden states are obtained by performing HERMES multimodal fusion on the modal feature vectors.

3. The method according to claim 1, characterized in that, The process of predicting and controlling the future state of the hidden state includes: The predicted hidden state at the current time step is obtained by using the world model to predict the hidden state at the previous time step based on the hidden state at the previous time step, the control instructions at the previous time step, and the auxiliary features. A joint objective function is constructed, which includes the hidden state prediction error and the chassis control cost. With the goal of minimizing the joint objective function, the future state is predicted and the optimal control command is generated by combining the world model, thus obtaining the future state and the control command.

4. The method according to claim 1, characterized in that, The process of optimizing the world model based on future states includes: The real-time observation state is obtained, and the error loss at the same time step and the corresponding next time step is applied based on the real state and the future state to update the model parameters of the world model.

5. The method according to claim 1, characterized in that, The digital twin scenario includes a traffic environment digital twin and an intelligent chassis digital twin. The traffic environment digital twin maps the real road environment and obtains multimodal data of the real road environment in real time. The intelligent chassis digital twin simulates the behavior of the chassis based on the vehicle dynamics model of the chassis and according to control commands.

6. The method according to claim 1, characterized in that, The process of dynamically adjusting and configuring the network slice resources includes: Determine if there is a link mutation. If there is no link mutation, obtain the end-to-end latency and bandwidth utilization of the network slice resources, and update the resource share of different network slice resources according to the end-to-end latency and bandwidth utilization. When a link malfunctions, network slice resources are configured using Oracle's secure migration learning method.

7. The method according to claim 1, characterized in that, The process of online optimization of the world model and network slice resources includes: Performance metrics are obtained through a feedback mechanism, including end-to-end latency throughout the entire process of sensing control execution, multi-step prediction error between future state and actual observation, and deviation between the actual motion trajectory of chassis control and the predicted control path. Monitor the overall convergence status based on the aforementioned performance indicators; The world model parameters are updated using a gradient method based on the multi-step prediction error and the time backpropagation method; the resource share of the network slice resources is adjusted based on the end-to-end latency throughout the process; until the performance index is within the safety threshold.

8. An autonomous driving chassis control system based on a world model and network slicing closed loop, characterized in that, Used to perform the method described in any one of claims 1-7.

Images