Automatic driving evolution method and system based on progressive expert hybrid network

Abstract

The present disclosure relates to an automatic driving evolution method and system based on a progressive expert hybrid network, which can solve the problems that the existing expert hybrid network is difficult to evolve with new driving conditions, has catastrophic forgetting and insufficient knowledge transfer. The method periodically trains the expert hybrid network: for new driving conditions, new experts are added to the existing expert network to form a new expert network, and the first layer of the new expert is connected to the first layer of the historical expert through a horizontal transfer interface. During training, the parameters of the historical experts are frozen, and the parameters of the new experts are optimized. The outputs of the new experts and the historical experts are input into the gate layer to obtain the control decision signal. Because the historical experts only participate in reasoning and do not update the weights, the existing knowledge can be stably retained to prevent catastrophic forgetting; and the new experts absorb existing knowledge in the initialization stage and learn differently for new conditions, achieving a balance between transfer and expansion, improving the accuracy and safety of automatic driving control.

Description

Technical Field

[0001] This disclosure relates to the fields of intelligent driving and artificial intelligence, and in particular to an autonomous driving evolution method and system based on a progressive expert hybrid network, which enables dynamic adaptation and iterative optimization for complex traffic environments and improves the accuracy of control decisions. Background Technology

[0002] Autonomous driving, as a cutting-edge field deeply integrating artificial intelligence and transportation, has made significant progress in decision-making and control, path planning, and other areas. With technological advancements, learning-based algorithms such as deep learning, reinforcement learning, and imitation learning are gradually becoming the mainstream methods for autonomous driving decision-making and control. These methods can automatically extract features from large-scale driving data and learn complex driving behaviors, offering greater adaptability and flexibility compared to earlier rule-based methods, thus laying a crucial foundation for the realization of high-level autonomous driving.

[0003] However, existing learning-based autonomous driving methods still suffer from serious shortcomings in complex edge scenarios. Real-world road environments exhibit a long-tail distribution, with abundant data on common scenarios but extremely scarce samples of rare and critical hazardous scenarios, leading to insufficient generalization ability of models in critical situations. Furthermore, existing methods lack continuous adaptability, often exhibiting catastrophic forgetting in dynamically changing task sequences—that is, overwriting knowledge gained from older tasks when learning new tasks, making it difficult to achieve stable balance across multiple tasks. This problem severely restricts the evolution and optimization capabilities of autonomous driving systems in dynamic environments.

[0004] To alleviate these problems, researchers have proposed a hybrid expert approach, which improves cross-task adaptability by dividing tasks among multiple expert subnetworks. However, the traditional hybrid expert network structure is static; the number and combination of experts are fixed when the model is built, making it difficult to evolve gradually as the task sequence expands. This still cannot fundamentally avoid catastrophic forgetting and insufficient knowledge transfer, thus affecting the accuracy of control decisions. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this disclosure proposes an autonomous driving evolution scheme based on a progressive expert hybrid network. This scheme enables the expert hybrid network to gradually expand with the introduction of new tasks, achieve cross-scenario knowledge fusion through dynamic weighting, and retain the memory of old knowledge during the evolution process. This endows the autonomous driving system with true continuous adaptation and evolution capabilities, improving the accuracy of driving decisions. The specific technical solution is as follows.

[0006] Firstly, this disclosure proposes an autonomous driving evolution method based on a progressive expert hybrid network. The method involves periodically training the expert hybrid network, including: constructing a comprehensive scene state from vehicle state, navigation state, and environmental perception data; determining the driving condition based on the comprehensive scene state; and for a new driving condition, adding a new expert to the existing expert network to form a new expert network. The new expert then uses a lateral transfer interface to transfer its first-order expert... The first layer and the history expert Layers connected, n is the number of network layers for the expert; during training, the comprehensive scene state is used as the input to the new expert network, the parameters of the historical experts are frozen, and the parameters of the new expert are optimized; the output of the new expert and the output of each historical expert are input into the gating layer to obtain executable control decision signals, which include control quantities for steering, acceleration and / or braking.

[0007] In one embodiment of the above technical solution, the training is performed in the cloud, and the parameters of the evolved expert hybrid network obtained after training are sent to the vehicle using a registry synchronization mechanism.

[0008] In one embodiment of the above technical solution, the new expert's first... The layer receives each historical expert's data through the horizontal migration interface at the 1st level. The features output by the layer are compressed and mapped to form a reusable aggregated representation. New experts in the The feature representation of the layer output is as follows , Let n be a non-linear activation function, and n be the number of historical experts. The first expert at time t Layer input, For the new experts Layer weight parameters, For the new experts Layer horizontal mapping matrix.

[0009] In one embodiment of the above technical solution, the new expert network updates the parameters of the new expert using the following objective function: , This represents the model's control output at the current moment. Indicates the actual reference control signal. These are the parameters for the current training expert. This serves as a reference parameter for historical experts when freezing data. To balance the constraint weights of fitting the current driving task with preserving historical knowledge.

[0010] In one embodiment of the above technical solution, the new driving condition is determined based on the comprehensive scene state. The steps include: calculating the similarity between the current comprehensive scene state and each historical comprehensive scene state; if the similarity is within a preset threshold, the driving condition to which it belongs is determined to be the same as the driving condition corresponding to the most similar historical comprehensive scene state; otherwise, the driving condition to which it belongs is determined to be a new driving condition.

[0011] In one embodiment of the above technical solution, the aggregated representation is obtained through the following steps: the feature representations extracted by n historical experts from the comprehensive scene state at time t are reduced in dimension and transformed nonlinearly using a linear transformation layer and a nonlinear activation layer.

[0012] Secondly, this disclosure proposes an autonomous driving evolution system based on a progressive expert hybrid network. The system includes an establishment module and a training module, which periodically train the expert network. The establishment module is configured to construct a comprehensive scene state from vehicle state, navigation state, and environmental perception data. Based on the comprehensive scene state, the driving condition is determined. For a new driving condition, a new expert is added to the existing expert network to form a new expert network. The new expert uses a lateral migration interface to transfer its first-order expert network to the next-order expert network. The first layer and the history expert Layers connected, , where n is the number of network layers for the expert; the training module is configured to take the comprehensive scene state as the input to the new expert network, freeze the parameters of the historical experts during training, and optimize the parameters of the new expert; the output of the new expert and the output of each historical expert are input into the gating layer to obtain executable control decision signals, which include control quantities for steering, acceleration and / or braking.

[0013] The beneficial technical effects of this disclosure are as follows: This solution improves upon existing hybrid expert methods, enabling the continuous expansion and gradual evolution of the expert hybrid network as new driving conditions emerge, thus avoiding problems such as catastrophic forgetting and insufficient knowledge transfer. This solution implements a continuously evolving autonomous driving decision-making structure, capable of long-term retention of historical knowledge and rapid learning of new knowledge in complex multi-source scenarios. This solution allows for progressively scalable training and learning in the cloud, and lightweight deployment on the vehicle side, improving the control performance of autonomous driving. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of a progressive expert hybrid network architecture in one implementation.

[0016] Figure 2 This is a schematic diagram of an autonomous driving evolution system architecture based on a progressive expert hybrid network in one implementation. Detailed Implementation

[0017] As the background technology indicates, real-world road environments exhibit a long-tail distribution characteristic, with abundant data for common scenarios but scarce samples for rare scenarios. Existing autonomous driving methods perform poorly in complex edge scenarios, particularly in rare but critical hazardous scenarios where data scarcity leads to weak generalization ability. In dynamic task sequences, existing models are prone to catastrophic forgetting, struggling to balance the learning and retention of knowledge across multiple tasks. While hybrid expert methods improve adaptability through the division of labor among multiple expert sub-networks, traditional static and fixed structures are difficult to evolve with task expansion, failing to address the core issues of catastrophic forgetting and insufficient knowledge transfer, thus hindering the dynamic evolution capability of autonomous driving systems.

[0018] Based on this, this invention proposes an autonomous driving evolution method based on a progressive expert hybrid network. This method achieves rapid adaptation to new tasks while preserving historical task knowledge through a mechanism of "progressive expansion + gating weighting + knowledge transfer." The method introduces a progressive expansion mechanism to dynamically expand the expert hybrid network, combines gating weighting to achieve expert collaborative reasoning, and utilizes a knowledge transfer mechanism to stably retain historical task knowledge. This invention supports the continuous evolution of autonomous driving systems in multi-source and multi-changing scenarios, significantly improving their long-term adaptability and decision robustness.

[0019] The following provides a clear and complete description of how the technical solution of this case is implemented. Obviously, the described implementation methods are only a part of the implementation methods of this case, and not all of them. Based on the implementation methods in this case, all other implementation methods obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0020] (a) Existing expert hybrid networks

[0021] In autonomous driving, expert hybrid network models primarily refer to systems that enhance decision-making capabilities through technologies such as multimodal fusion, risk prediction, and semantic reasoning. Each expert in an expert hybrid network has multiple network layers.

[0022] In one embodiment, the existing expert hybrid network has 3 experts , and A branch of experts Depend on , , ...composed of the content of the input expert hybrid network, passing through multiple layers of network. , , ...output after calculation Similarly, expert branches The input content to the expert hybrid network is processed through multiple layers of networks. , , … Output after calculation After multiple layers of network , , … Output after calculation Features extracted from each expert branch , and This constitutes a sequence input gating layer, which assigns weights to the matching degree of each expert. , and These weights are used to evaluate the extracted features. , and Weighted fusion is performed to obtain fused features. Based on the fused features, the control decision signals that the vehicle can execute can be obtained.

[0023] (ii) Progressive expert hybrid network

[0024] A sequence of driving tasks is generated based on road information, with each task corresponding to a driving condition. These driving conditions include intersections, roundabouts, ramps, and narrow passages. Existing autonomous driving decisions often suffer from catastrophic forgetting in dynamically changing task sequences, meaning that when learning new tasks, knowledge acquired in older tasks is overwritten, making it difficult to achieve a stable balance among multiple tasks. To address this issue, this application proposes a progressive expert hybrid network. Unlike existing expert hybrid networks (MoE), the construction of this application's expert hybrid network is dynamically scalable.

[0025] Specifically, in constructing the expert hybrid network of this application: for a new driving condition, a new expert is added to the existing expert network to form a new expert network. The new expert transfers its first-level expert to the next level through a lateral migration interface. The first layer and the history expert Layers connected, Here, n represents the number of network layers for each expert. By having each expert learn from a specific driving condition, the expert hybrid network is expanded when encountering a new driving condition. This allows the control decisions of the expert primarily learning from that driving condition to serve as the main reference when encountering the same driving condition, thus improving the safety of autonomous driving decisions. Furthermore, the new expert is connected to historical experts, allowing them to learn from their experience and improving the learning rate of the new expert.

[0026] See Figure 1 For example, a hybrid expert network has one expert. , its origin , , ...composed of the content of the input expert hybrid network, passing through multiple layers of network. , , ...output after calculation The gating layer is based on Output control decision signal. A new driving condition arises at this point, as indicated by the expert... Add a new expert to the existing one The new expert Initialize multilayer network , , ...output after calculation .in, Receive itself Layer output and experts middle The output of the layer, Receive itself Layer output and experts In The output of the layer, and so on. The gating layer is based on... and Weights are adjusted, and control decision signals are output. When a new driving condition arises, experts... and experts Add new experts on the basis .at this time, At the same time with experts and experts Connected. (From the diagram) Receive itself The layer output, while receiving , The output of the layer, Receive itself The output, while receiving and The output is then calculated and so on, until the final output is obtained. Gating layer , and After internal weight allocation, the control decision signal is output.

[0027] One method for determining the driving condition is based on a comprehensive scenario state. Specifically, for the current comprehensive scenario state, a similarity calculation is performed between it and each historical comprehensive scenario state. If the similarity falls within a preset threshold, the driving condition is determined to be the same as the driving condition corresponding to the most similar historical comprehensive scenario state; otherwise, the driving condition is determined to be a new driving condition.

[0028] In this application, the specific input to the expert hybrid network is a comprehensive scene state, which consists of vehicle state, navigation state, and environmental perception data. Providing a unified and complete scene representation from the comprehensive scene state improves the accuracy of decision-making and reasoning in the expert hybrid network. In one exemplary representation of the comprehensive scene state, the time step... The overall scene state is represented as , Indicates vehicle status. Indicates navigation status. This represents environmental perception data.

[0029] The vehicle status may include, but is not limited to, information such as the vehicle's observed position coordinates, the vehicle's observed longitudinal velocity, and the vehicle's observed longitudinal acceleration.

[0030] The navigation status is a dynamic state of vehicle driving and road calculation. The navigation status information includes one of the following: whether the vehicle position matches the road, whether road calculation or road recalculation has been performed, and whether deviation has occurred in the state of having no destination.

[0031] The environmental perception data includes road environment perception data collected by autonomous vehicles or roadside perception devices, including environmental data corresponding to motor vehicles, non-motor vehicles, pedestrians, buildings, plants, traffic signs, traffic lights, road markings, fences, etc., such as the position, size, status, color, texture, etc. of scene elements.

[0032] The overall scenario status of this application can be collected through a data acquisition interface. The data acquisition interface refers to the interface that connects to sensors such as cameras, radar, and GPS, as well as V2X devices, to realize data on the vehicle's surrounding environment, vehicle status, and driving behavior.

[0033] (III) Extended Training

[0034] When a new driving condition is detected, the newly added expert branch has an independent parameter space during the initialization phase, avoiding coupling and mutual interference with historical experts. To improve the efficiency of new experts in adapting to new driving conditions, this invention does not train from scratch, but rather expands the training by referencing the experience of historical experts. This avoids the high cost of retraining all experts, reduces the risk of catastrophic forgetting, and enables the expert database to progressively expand and evolve as the task flow grows.

[0035] Specifically, in the extended training method, the comprehensive scene state is used as the input of the new expert network. During training, the parameters of the historical experts are frozen, and the parameters of the new experts are optimized. The outputs of the new experts and the outputs of each historical expert are input into the gating layer to obtain executable control decision signals, which include control quantities for steering, acceleration, and / or braking.

[0036] In the aforementioned training process, only experts corresponding to new driving conditions are incrementally trained, while the parameters of historical experts remain frozen, participating only in inference without updating their weights. This ensures the stable retention of existing knowledge, effectively preventing catastrophic forgetting and improving the accuracy and safety of autonomous driving control. Through this incremental training approach, new experts can efficiently absorb existing knowledge during the initialization phase and further differentiate their learning for new driving conditions during subsequent training, effectively achieving a balance between transfer and extension. By combining incremental training with parameter freezing, the system can continuously expand expert capabilities without retraining the entire model, effectively improving long-term adaptability and evolutionary efficiency.

[0037] Because the new expert uses the lateral migration interface to transfer their first... The first layer and the history expert Layers connected, , where n is the network layer number of the expert, and therefore the first layer of the new expert... The layer receives each historical expert's data through the horizontal migration interface at the 1st level. Features of the layer output. The lateral transfer interface is a mapping subnetwork containing linear transformation layers and nonlinear activation layers, which transfers the new expert to the first layer. The first layer received by all historical experts The layer output, through continuous dimensionality reduction and nonlinear transformation, yields a reusable aggregated representation. New experts in the first The feature representation of the layer's output is as follows , Let n be a non-linear activation function, and n be the number of historical experts. For the new expert at time t Layer input, For the new experts Layer weight parameters, For the new experts Layer horizontal mapping matrix.

[0038] The lateral migration interface allows new experts to reference the knowledge and experience of historical experts when learning new driving conditions. Historical experts, in turn, participate in the optimization and updates of new experts in a controlled manner through the lateral migration interface. This ensures that the model absorbs the characteristics of new driving conditions without disrupting its existing learned handling of those conditions, resulting in good system performance and stability.

[0039] The aforementioned mapping subnetwork can also consist of a linear transformation layer and a nonlinear activation layer. The linear transformation layer performs a linear mapping on the input features through a fully connected layer, transforming the features from one dimension to another. The main function of this layer is to adjust the feature representation to better suit downstream tasks. The nonlinear activation layer is used to introduce a nonlinear transformation by applying a nonlinear activation function (such as ReLU) after the linear transformation, thereby enhancing the network's expressive power. The role of the mapping subnetwork is to process the features of historical experts through linear transformation and nonlinear activation, generating a new aggregated representation that effectively supports the learning of new experts.

[0040] In expert hybrid networks, the gate layer has a lightweight scoring function. This function is used to evaluate the matching degree of each expert, and the weights are normalized and allocated using a softmax structure. Its calculation form is as follows: This allows for the construction of the corresponding expert network weight table. This mechanism dynamically assigns weights to each registered expert branch. It adaptively adjusts the expert network composition based on environmental changes, ensuring both flexibility and stability in the decision-making process.

[0041] After weighting, the expert outputs are fused based on these weights to obtain the final control decision signal, which includes control variables for steering, acceleration, and / or braking. The mathematical expression of the fusion process is as follows: , The control output at the inference moment represents continuous control quantities such as steering, acceleration, and braking. Indicates the first The control response of an expert network in the current state.

[0042] The system achieves cross-scenario knowledge fusion through dynamic weighting, and integrates the outputs of various expert networks for reasoning. This enables the vehicle to achieve efficient collaboration among multiple expert networks during driving, has the advantage of quickly adapting to changes in the traffic environment, and can complete policy switching without relying on global model retraining. This makes the vehicle's decision response efficient and supports the beneficial effects of stable online operation in multiple scenarios.

[0043] During training, the extended expert hybrid network adopts a constrained optimization objective function, which minimizes the deviation from the distribution of control signals corresponding to old driving conditions while fitting the control signals corresponding to new driving conditions. Therefore, the optimization objective comprehensively considers the fitting error of the current learning and the preservation of historical knowledge.

[0044] In one implementation, the objective function of the extended expert hybrid network is a combination of least squares terms and regularization constraint terms. The objective function can be defined as follows: , Mean square error, This represents the model's control output at the current moment. Indicates the actual reference control signal. These are the parameters for currently training the expert network. This serves as a reference parameter for historical experts when the network is frozen. To balance the constraint weights of fitting the current driving task with preserving historical knowledge. This is a set value; an example value is 0.1.

[0045] (iv) Application of Reasoning

[0046] This invention adopts an architecture that separates the inference and training ends, consisting of a vehicle-side terminal and a cloud-based terminal. See [link to relevant documentation]. Figure 2 As shown.

[0047] Vehicle-side: It uses its own sensing devices to perceive the environment, collect data, and upload the collected data to the cloud through communication base stations.

[0048] In the cloud: Data collected from the vehicle is received via communication base stations, and periodic training is performed. The steps include: first, constructing a comprehensive scene state from vehicle status, navigation status, and environmental perception data; then, determining the driving condition based on this comprehensive scene state; and finally, for a new driving condition, adding a new expert to the existing expert network to form a new expert network. The new expert then uses a lateral migration interface to transfer its first-level... The first layer and the history expert Layers connected, Here, n represents the number of network layers for the expert network. During training, the comprehensive scene state is used as the input to the new expert network, the parameters of historical experts are frozen, and the parameters of the new expert are optimized. The outputs of the new expert and the outputs of each historical expert are input into the gating layer to obtain executable control decision signals. The trained new expert hybrid network is then deployed to the vehicle. During deployment, since the parameters of historical experts are frozen and not updated, the parameters of the new expert and the gating layer only need to be distributed through the registry synchronization mechanism to achieve progressive expansion and updating of the vehicle-side expert hybrid network without requiring a complete redeployment.

[0049] In summary, the present invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the present invention.

[0050] For example, an autonomous driving evolution method based on a progressive expert hybrid network periodically expands and trains the expert hybrid network using new driving conditions. This includes: constructing a comprehensive scene state from vehicle state, navigation state, and environmental perception data; determining the driving condition based on the comprehensive scene state; and for a new driving condition, adding a new expert to the existing expert network to form a new expert network. The new expert then transfers its first-order expert to the next expert network via a lateral transfer interface. The first layer and the history expert Layers connected, n is the number of network layers of the expert; the comprehensive scene state is used as the input of the new expert network, the parameters of the historical experts are frozen during training, and the parameters of the new expert are optimized; the output of the new expert and the output of each historical expert are input into the gating layer to obtain the executable control decision signal, which includes the control quantity of steering, acceleration and / or braking.

[0051] Accordingly, an autonomous driving evolution system based on a progressive expert hybrid network can be obtained. The system includes a setup module and a training module. The setup and training modules periodically extend and train the expert network using new driving conditions. The setup module is configured to construct a comprehensive scene state from vehicle state, navigation state, and environmental perception data. Based on this comprehensive scene state, the system determines the driving condition. For a new driving condition, a new expert is added to the existing expert network to form a new expert network. The new expert uses a lateral transfer interface to transfer its first... The first layer and the history expert Layers connected, , where n is the number of network layers for the expert; the training module is configured to take the comprehensive scene state as the input to the new expert network, freeze the parameters of the historical experts during training, and optimize the parameters of the new expert; the output of the new expert and the output of each historical expert are input into the gating layer to obtain executable control decision signals, which include control quantities for steering, acceleration and / or braking.

[0052] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0053] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0054] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0055] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0056] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0057] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0058] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be well known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0059] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. An evolutionary method for autonomous driving based on a progressive expert hybrid network, characterized in that, The expert hybrid network is trained periodically, including: The vehicle status, navigation status, and environmental perception data are combined to form a comprehensive scene status. The driving condition is determined based on the comprehensive scene status. The steps include: calculating the similarity between the current comprehensive scene status and each historical comprehensive scene status. If the similarity is within a preset threshold, the driving condition is determined to be the same as the driving condition corresponding to the most similar historical comprehensive scene status; otherwise, the driving condition is determined to be a new driving condition. For a new driving condition, a new expert network is formed by adding a new expert to the existing expert network. The new expert then uses a lateral migration interface to transfer its first-level expert to the next level. The first layer and the history expert Layers connected, , where n is the network layer number of the expert; the first layer of the new expert. The layer receives each historical expert's data through the horizontal migration interface at the 1st level. The features output by the layer are compressed and mapped to form a reusable aggregated representation. New experts in the The feature representation of the layer's output is as follows , Let n be a non-linear activation function, and n be the number of historical experts. For the new expert at time t Layer input, For the new experts Layer weight parameters, For the new experts Layer horizontal mapping matrix; During training, the comprehensive scene state is used as the input to the new expert network, the parameters of historical experts are frozen, and the parameters of the new experts are optimized. The training is performed periodically in the cloud using data collected from the vehicle. At the end of the training, an evolved hybrid expert network is obtained. During deployment, the parameters of the new experts and the parameters of the gating layer are sent to the vehicle through a registry synchronization mechanism. The new expert network uses the following objective function to update the parameters of the new experts: , This represents the model's control output at the current moment. Indicates the actual reference control signal. These are the parameters for the current training expert. This serves as a reference parameter for historical experts when freezing data. To balance the constraints of fitting the current driving task with preserving historical knowledge; The outputs of the new experts and the outputs of each historical expert are input into the gating layer to obtain executable control decision signals, which include control quantities for steering, acceleration and / or braking.

2. An autonomous driving evolutionary system based on a progressive expert hybrid network, characterized in that: The system includes a setup module and a training module, through which the expert network is periodically trained; The establishment module is configured to construct a comprehensive scene state from vehicle status, navigation status, and environmental perception data, and determine the driving condition based on the comprehensive scene state. The steps include: calculating the similarity between the current comprehensive scene state and each historical comprehensive scene state; if the similarity is within a preset threshold, the driving condition is determined to be the same as the driving condition corresponding to the most similar historical comprehensive scene state; otherwise, the driving condition is determined to be a new driving condition; for a new driving condition, a new expert is added to the existing expert network to form a new expert network, and the new expert uses a lateral migration interface to transfer its first... The first layer and the history expert Layers connected, , where n is the network layer number of the expert; the first layer of the new expert. The layer receives each historical expert's data through the horizontal migration interface at the 1st level. The features output by the layer are compressed and mapped to form a reusable aggregated representation. New experts in the The feature representation of the layer's output is as follows , Let n be a non-linear activation function, and n be the number of historical experts. For the new expert at time t Layer input, For the new experts Layer weight parameters, For the new experts Layer horizontal mapping matrix; The training module is configured to use the comprehensive scene state as input to the new expert network. During training, the parameters of historical experts are frozen, and the parameters of the new experts are optimized. The training is performed periodically in the cloud using data collected from the vehicle. At the end of the training, an evolved hybrid expert network is obtained. During deployment, the parameters of the new experts and the parameters of the gating layer are sent to the vehicle via a registry synchronization mechanism. The new expert network uses the following objective function to update the parameters of the new experts: , This represents the model's control output at the current moment. Indicates the actual reference control signal. These are the parameters for the current training expert. This serves as a reference parameter for historical experts when freezing data. To balance the constraint weights of fitting the current driving task and preserving historical knowledge, the outputs of the new experts and the outputs of each historical expert are input into the gating layer to obtain executable control decision signals, which include control quantities for steering, acceleration, and / or braking.

Images