Vehicle-Cloud Collaborative Autonomous Driving Model Iterative Training Method and System

Abstract

This application relates to the field of autonomous driving technology, specifically a vehicle-cloud collaborative method and system for iterative training of autonomous driving models. Based on multi-dimensional indicators, it accurately identifies the key scenarios most valuable for model improvement, significantly enhancing the quality and diversity of training data. A trajectory segmenter is used to compress trajectory fragments into compact token sequences, combined with contextual feature dimensionality reduction, greatly reducing the amount of uploaded data. This effectively alleviates bandwidth pressure, reduces cloud storage overhead, and shortens data transmission and processing cycles, accelerating model iteration. The uploaded content consists only of token sequences and dimensionality-reduced features, rather than raw sensor data, thus reducing the risk of leakage. The cloud can reconstruct high-fidelity scenes based on the compact data, and combined with targeted training strategies, achieves targeted learning for challenging scenarios. Coupled with the feedback loop of vehicle-cloud collaboration, it continuously optimizes the model's performance in challenging scenarios, improving the overall system's iterative efficiency and adaptability.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, specifically to a vehicle-cloud collaborative autonomous driving model iterative training method and system. Background Technology

[0002] The performance of end-to-end autonomous driving models is highly dependent on the quality and diversity of training data. In real-world operations, vehicles constantly encounter new scenarios and challenges; these "difficult scenarios" are key data sources for improving model capabilities. However, efficiently collecting, transmitting, and utilizing this difficult scenario data remains a significant challenge for current autonomous driving technology.

[0003] Traditional data collection methods typically employ either a "full upload" or "random sampling" strategy. The full upload strategy uploads all sensor data to the cloud, which preserves complete information but incurs enormous communication bandwidth and storage costs. While the random sampling strategy reduces the amount of data, it may lead to the omission of key and challenging scenarios, affecting model training performance.

[0004] The amount of data generated by autonomous vehicles is enormous. A Level 4 autonomous vehicle equipped with multiple sensors (LiDAR, cameras, millimeter-wave radar, etc.) can generate several terabytes of raw data per day. Uploading all this data to the cloud for training presents the following challenges:

[0005] (1) Bandwidth limitation: The actual upload bandwidth of 4G / 5G networks is limited, making it difficult to support large-scale real-time data uploads;

[0006] (2) High cost: The large amount of data transmission and storage generates high operating costs;

[0007] (3) Poor real-time performance: The data upload and processing cycle is long, which affects the model iteration speed;

[0008] (4) Privacy risks: Raw sensor data may contain sensitive information and pose a risk of privacy leakage. Summary of the Invention

[0009] In view of this, the purpose of this application is to provide a method and system for iterative training of autonomous driving models in a vehicle-cloud collaborative manner, so as to solve the problems in the background art.

[0010] To achieve the above objectives, this application adopts the following technical solution:

[0011] The iterative training method for the vehicle-cloud collaborative autonomous driving model in this application includes the following steps:

[0012] The system acquires trajectory segment data of the vehicle in multiple historical scenarios, probability distribution of model-predicted actions, and information on manual takeover, wherein the model-predicted actions are output by an end-to-end model deployed on the vehicle.

[0013] Difficult scenarios were selected based on the trajectory segment data, the probability distribution of the model's predicted actions, and information on human intervention.

[0014] The trajectory fragment data of the difficult scenario is input into a pre-built trajectory segmenter to obtain the target token sequence; and the context information and scene metadata of the difficult scenario are extracted. The context information of the difficult scenario is dimensionality reduced to obtain dimensionality reduction feature information. The context information includes environmental perception features, map features, vehicle state features and model intermediate layer features. The scene metadata includes time, location and vehicle ID.

[0015] A scene compressed data packet is constructed based on the target token sequence, the dimensionality reduction feature information, and the scene metadata, and the scene compressed data packet is sent to the cloud;

[0016] The scene data is reconstructed in the cloud based on the scene compression data packet to obtain trajectory fragment data, context information and scene metadata of the difficult scene; and the autonomous driving model is iteratively trained based on the trajectory fragment data and context information of the difficult scene.

[0017] In one embodiment of this application, difficult scenarios are identified based on the trajectory segment data, the probability distribution of model-predicted actions, and information on human intervention, including:

[0018] The trajectory segment data is encoded using a pre-built trajectory segmenter to obtain the token encoding result of the trajectory segment data; the target codeword most similar to the token encoding result is extracted from the codebook of the trajectory segmenter, and the Euclidean distance between the token encoding result and the target codeword is calculated. ;

[0019] Calculate the entropy of the probability distribution of the predicted actions by the model. And extract the highest probability ;

[0020] Constructing indicator functions for manual takeover information ;

[0021] For the Euclidean distance respectively and the entropy value Normalization is performed to obtain the normalized distance. and normalized entropy value ; and based on the normalized distance The normalized entropy value The maximum probability and the indicated function Calculate the overall difficulty score The comprehensive difficulty score The mathematical expression is:

[0022]

[0023] In the formula, As the first weight, As the second weight, As the third weight, It is the fourth weight;

[0024] Based on the Euclidean distance The entropy value The maximum probability The indicated function Or overall difficulty score Filter out difficult scenarios.

[0025] In one embodiment of this application, based on the Euclidean distance The entropy value The maximum probability The indicated function Or overall difficulty score Filtering difficult scenarios includes:

[0026] A scenario that satisfies any one of the target conditions is considered a suffering scenario. The target conditions include:

[0027] The Euclidean distance The distance is greater than the preset distance threshold;

[0028] The entropy value Greater than the preset entropy threshold;

[0029] The maximum probability Less than the preset probability threshold;

[0030] The value of the indicator function indicates that manual intervention is present;

[0031] The overall difficulty score is greater than the preset difficulty threshold.

[0032] In one embodiment of this application, the trajectory segmenter includes an encoder for mapping trajectory segments to a latent space vector sequence, a codebook for converting the latent space vector sequence into a token sequence, and a decoder for mapping the token sequence to trajectory decoding segments. The method for constructing the trajectory segmenter includes:

[0033] Obtain driving trajectory segment samples;

[0034] The trajectory segment sample is input into the encoder to obtain the predicted token sequence output by the codebook and the trajectory decoded segment output by the decoder;

[0035] The loss is calculated based on a pre-constructed loss function, the driving trajectory segment samples, the trajectory decoding segments, and the predicted token sequence, wherein the mathematical expression of the loss function is:

[0036]

[0037] In the formula, Indicates loss, This represents a sample of a driving trajectory segment. This represents a trajectory decoding segment. This represents the sequence of latent space vectors output by the encoder. This indicates that the gradient operation is stopped. For the codeword sequence in the codebook, The index is the token sequence. For loss balancing parameters;

[0038] The trajectory segmenter is trained based on the loss.

[0039] In one embodiment of this application, scene data is reconstructed in the cloud based on the scene compressed data packet to obtain trajectory fragment data, context information, and scene metadata of the difficult scene, including:

[0040] The decoder based on the trajectory segmenter decodes the target token sequence in the scene compressed data packet to obtain trajectory fragment data; and it upscales the dimensionality reduction feature information in the scene compressed data packet to obtain context information, and directly extracts the scene metadata in the scene compressed data packet.

[0041] In one embodiment of this application, iterative training of an autonomous driving model based on trajectory segment data and contextual information of a difficult scenario includes:

[0042] The context information is used as a data sample, and the trajectory segment data is used as a label to construct training data samples;

[0043] The training data samples are augmented to obtain augmented data samples and an augmented data sample set, wherein the data augmentation includes geometric transformation, temporal perturbation and parameter perturbation;

[0044] The end-to-end autonomous driving model in the cloud is trained based on the enhanced data sample set and a pre-configured training strategy, wherein the training strategy includes a course learning strategy, a hard sample mining strategy, an adversarial training strategy, and an incremental training strategy.

[0045] The end-to-end autonomous driving model on the vehicle is differentially updated based on the trained cloud-based end-to-end autonomous driving model.

[0046] In one embodiment of this application, training an end-to-end autonomous driving model in the cloud based on the enhanced data sample set and a pre-configured training strategy includes:

[0047] Based on the comprehensive difficulty score, the samples in the augmented data sample set are arranged in ascending order;

[0048] In the early stages of training, the simplest multiple samples combined with standard gradient descent are used to incrementally train the end-to-end autonomous driving model in the cloud.

[0049] During the middle of training, the parameters of the end-to-end autonomous driving model in the cloud are frozen, and the loss of each augmented data sample is calculated. Augmented data samples with losses greater than a preset loss threshold are designated as hard samples, and a hard sample set is constructed based on the hard samples. Acquisition weights are assigned to each augmented data sample based on the loss, where the acquisition weights of non-hard samples are multiplied by a decay factor. Augmented data samples are acquired based on the acquisition weights of all samples and combined with standard gradient descent to incrementally train the end-to-end autonomous driving model in the cloud.

[0050] In the later stages of training, the sampling weight mechanism is retained, and adversarial examples that maximize the loss of each augmented data sample are generated. The end-to-end autonomous driving model in the cloud is incrementally trained based on the augmented data samples and the adversarial examples of the augmented data samples, wherein the adversarial examples are generated by adding perturbations to the augmented data samples.

[0051] This application also provides an iterative training system for autonomous driving models based on vehicle-cloud collaboration, including:

[0052] The acquisition module is used to acquire trajectory segment data of the vehicle in multiple historical scenarios, probability distribution of model predicted actions, and manual takeover information, wherein the model predicted actions are output by an end-to-end model deployed on the vehicle.

[0053] The scene filtering module is used to filter out difficult scenes based on the trajectory segment data, the probability distribution of the model's predicted actions, and the information on human intervention.

[0054] The data dimensionality reduction module is used to input the trajectory fragment data of the difficult scenario into a pre-built trajectory segmenter to obtain the target token sequence; and to extract the context information and scene metadata of the difficult scenario, and to reduce the dimensionality of the context information of the difficult scenario to obtain dimensionality reduction feature information. The context information includes environmental perception features, map features, vehicle state features and model intermediate layer features, and the scene metadata includes time, location and vehicle ID.

[0055] The upload module is used to construct a scene compressed data packet based on the target token sequence, the dimensionality reduction feature information and the scene metadata, and send the scene compressed data packet to the cloud;

[0056] The iterative training module is used to reconstruct scene data in the cloud based on the scene compressed data package to obtain trajectory fragment data, context information and scene metadata of the difficult scene; and to iteratively train the autonomous driving model based on the trajectory fragment data, context information and scene metadata of the difficult scene.

[0057] This application also provides an electronic device, including: a processor and a memory;

[0058] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to cause the electronic device to perform the methods described above.

[0059] This application also provides a computer-readable storage medium having a computer program stored thereon: when the computer program is executed by a processor, it implements the method described above.

[0060] The beneficial effects of this application are as follows: The vehicle-cloud collaborative autonomous driving model iterative training method and system of this application, based on multi-dimensional indicators such as trajectory token anomalies, prediction confidence, and manual intervention, can accurately identify the key scenarios most valuable for model improvement, avoiding the omission of key data caused by traditional random sampling, and significantly improving the quality and diversity of training data. A trajectory segmenter is used to compress trajectory fragments into compact token sequences, and combined with contextual feature dimensionality reduction, greatly reducing the amount of uploaded data, effectively alleviating bandwidth pressure, reducing cloud storage overhead, and shortening the data transmission and processing cycle, thus accelerating model iteration. The uploaded content is only the token sequence and dimensionality-reduced features, rather than the original sensor data, thereby reducing the risk of leakage of sensitive information such as road environment and pedestrians. The cloud can reconstruct high-fidelity scenes based on the compact data, and combined with targeted training strategies, achieve targeted learning for difficult scenarios; in conjunction with the feedback loop of vehicle-cloud collaboration, the model's performance in difficult scenarios is continuously optimized, improving the overall system's iterative efficiency and adaptability. Attached Figure Description

[0061] The present application will be further described below with reference to the accompanying drawings and embodiments:

[0062] Figure 1 This is a structural diagram of an iterative training system for a vehicle-cloud collaborative autonomous driving model according to an embodiment of this application.

[0063] Figure 2 This is a flowchart illustrating an iterative training method for a vehicle-cloud collaborative autonomous driving model in one embodiment of this application;

[0064] Figure 3 This is a logical structure diagram of an autonomous driving model iterative training system with vehicle-cloud collaboration shown in one embodiment of this application. Detailed Implementation

[0065] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0066] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the layers related to this application and are not drawn according to the actual number, shape and size of the layers in the actual implementation. In the actual implementation, the form, number and proportion of each layer can be arbitrarily changed, and the layer layout may also be more complex.

[0067] Numerous details are explored in the following description to provide a more thorough explanation of embodiments of this application; however, it will be apparent to those skilled in the art that embodiments of this application may be practiced without these specific details.

[0068] This application aims to provide an iterative training method for autonomous driving models in a vehicle-cloud collaborative manner, with the following technical objectives:

[0069] (1) Achieve efficient data compression: Use a trajectory segmenter to compress data from difficult scenarios into a compact token sequence, significantly reducing data transmission and storage requirements;

[0070] (2) Accurately identify difficult scenarios: Based on multi-dimensional indicators such as token anomalies and prediction confidence, accurately identify difficult scenarios that are valuable for model training;

[0071] (3) Supports cloud-based scene reconstruction: The cloud-based system efficiently reconstructs difficult scenes based on token sequences and contextual features for targeted training;

[0072] (4) Achieve rapid model iteration: Through the vehicle-cloud collaboration mechanism, shorten the cycle from data collection to model update and achieve rapid iterative optimization of the model.

[0073] Figure 1 This is an overall structural diagram of the vehicle-cloud collaborative autonomous driving model iterative training system in one embodiment of this application, as shown below. Figure 1 As shown, the iterative training system includes the following core modules:

[0074] (1) Vehicle-side data filtering module: Deployed on autonomous vehicles, responsible for real-time identification of difficult scenarios and data compression encoding;

[0075] (2) Vehicle-side Token Encoder: Based on the trajectory segmenter, the trajectory data is encoded into a token sequence;

[0076] (3) Cloud data receiving module: Receives uploaded data from multiple vehicles, and performs aggregation and preprocessing;

[0077] (4) Cloud-based scene reconstruction module: Reconstructing the training scene based on token sequence and context features;

[0078] (5) Cloud-based model training module: Targeted training of end-to-end models based on reconstructed scenarios;

[0079] (6) Model distribution module: Distribute the trained model to the vehicle for deployment and updates.

[0080] Figure 2 This is a flowchart of an iterative training method for a vehicle-cloud collaborative autonomous driving model according to an embodiment of this application, as shown below. Figure 2 As shown, the iterative training method for the vehicle-cloud collaborative autonomous driving model in this application includes:

[0081] S210, acquire trajectory segment data of the vehicle in multiple historical scenarios, probability distribution of model predicted actions, and manual takeover information, wherein the model predicted actions are output by an end-to-end model deployed on the vehicle.

[0082] The three methods for obtaining data are as follows:

[0083] (1) Trajectory segment data: During vehicle operation, the vehicle's pose information (position, heading angle, velocity, acceleration, etc.) is recorded at a fixed frequency (e.g., 10Hz) to form a continuous time series. The trajectory segments are obtained by slicing the data into fixed time windows (e.g., the past 2 seconds, a total of 20 frames). ,in, The above data is stored in a circular buffer on the vehicle side for the difficult scene detection module to read.

[0084] (2) Probability distribution of predicted actions: The end-to-end model outputs the probability distribution of future actions (such as steering wheel angle and acceleration) at each time step. If the action is discretized (e.g., the steering angle is divided into 51 intervals), the softmax probability vector is output. If the action is continuous (such as a Gaussian policy), the output is the mean and variance. This distribution is cached along with the current timestamp and used to calculate the prediction entropy or maximum probability.

[0085] (3) Manual Takeover Information: A takeover signal is triggered when the safety officer presses the takeover button or when the steering wheel / brake is manually engaged. The system records the takeover moment. It automatically saves all sensor data, model output, and vehicle status for 10 seconds before and after that moment. (Takeover sign) As an additional label for all samples within that time window.

[0086] S220, based on the trajectory segment data, the probability distribution of the model-predicted actions, and the information on manual intervention, difficult scenarios are selected, specifically including:

[0087] S221, the trajectory segment data is encoded based on a pre-built trajectory segmenter to obtain the token encoding result of the trajectory segment data; the target codeword most similar to the token encoding result is extracted from the codebook of the trajectory segmenter, and the Euclidean distance between the token encoding result and the target codeword is calculated. ;

[0088] Trajectory segmenters (such as VQ-VAE) map continuous trajectory segments to discrete token indices, and their codebook stores typical trajectory primitives. The Euclidean distance between the input trajectory and the nearest codeword in the codebook is considered. A large value indicates that the current trajectory pattern lacks sufficiently similar representations in the training codebook, meaning the trajectory belongs to a rare or insufficiently learned pattern. Therefore, It directly measures the "novelty" of the trajectory; the greater the distance, the more likely the scenario is to be a difficult one.

[0089] S222, Calculate the entropy of the probability distribution of the predicted actions by the model. And extract the highest probability ;

[0090] The action probability distribution output by the end-to-end model reflects the model's determinism regarding the current scene. Entropy The larger the value, the flatter the distribution, indicating that the model is hesitant about which action to take; maximum probability The smaller the value, the less confident the model is in its dominant actions. Both indicate that the model's current ability is insufficient to confidently handle the scenario, thus serving as an important indicator of challenging scenarios.

[0091] S223, Construct the indicator function for manual takeover information ;

[0092] Human intervention directly indicates that the autonomous driving system is unable to handle the current scenario safely and requires human intervention. This is a strong supervisory signal; regardless of the model output, a takeover event signifies that the scenario exceeds the system's capabilities and must be collected and used as a challenging scenario for training.

[0093] S224, respectively, for the Euclidean distance and the entropy value Normalization is performed to obtain the normalized distance. and normalized entropy value ; and based on the normalized distance The normalized entropy value The maximum probability and the indicated function Calculate the overall difficulty score The comprehensive difficulty score The mathematical expression is:

[0094]

[0095] In the formula, As the first weight, As the second weight, As the third weight, It is the fourth weight;

[0096] The comprehensive score can balance the contributions of four dimensions: trajectory novelty, model uncertainty, takeover signal, and output confidence, which facilitates subsequent threshold selection.

[0097] S225, a scenario that satisfies any one of the target conditions is designated as a suffering scenario, wherein the target conditions include:

[0098] (1) The Euclidean distance The distance is greater than the preset distance threshold;

[0099] (2) The entropy value Greater than the preset entropy threshold;

[0100] (3) The maximum probability Less than the preset probability threshold;

[0101] (4) The value of the indicator function indicates that manual intervention exists;

[0102] (5) The overall difficulty score is greater than the preset difficulty threshold.

[0103] A single metric might miss certain types of difficult scenarios (e.g., common trajectories but extremely uncertain model predictions). Using an OR logic where "meeting any one of the target conditions qualifies as difficult" can cover different types of difficulty patterns.

[0104] Condition (1) The trajectory pattern itself is rare;

[0105] Conditions (2) and (3) focus on the uncertainty of the model's current understanding;

[0106] Condition (4) depends on real external intervention;

[0107] Condition (5) is a weighted score that combines all dimensions. This multi-threshold parallel mechanism ensures high recall in difficult scenarios and avoids missing key samples.

[0108] S230, input the trajectory segment data of the difficult scene into a pre-built trajectory segmenter to obtain the target token sequence; and extract the context information and scene metadata of the difficult scene, perform dimensionality reduction on the context information of the difficult scene to obtain dimensionality reduction feature information, wherein the context information includes environmental perception features, map features, vehicle state features and model intermediate layer features, and the scene metadata includes time, location and vehicle ID;

[0109] This application pre-constructs a trajectory segmenter based on a VQ-VAE network to encode trajectory fragments into token sequences. The training process of the trajectory segmenter includes:

[0110] (1) Obtain driving trajectory segment samples;

[0111] The trajectory segmenter needs to learn to discretize continuous driving behavior into a finite number of typical patterns (code words). Therefore, the training data must cover diverse driving scenarios and contain complete dynamic processes. Each trajectory segment sample consists of continuous trajectory points within a fixed duration (e.g., 2 seconds of historical data + 5 seconds of future data). Each point contains motion states such as position, velocity, acceleration, and heading angle.

[0112] (2) Input the trajectory segment sample into the encoder to obtain the predicted token sequence output by the codebook and the trajectory decoding segment output by the decoder;

[0113] Specifically, the three sub-modules work together as follows:

[0114] Encoder: Employs multi-layer Transformer or convolutional networks to process the input trajectory segments. Mapped to a sequence of vectors in the latent space The encoder's role is to extract high-order features of the trajectory and compress them into a low-dimensional manifold.

[0115] Quantization module: For each latent vector In the codebook The most similar codeword is found through nearest neighbor search:

[0116]

[0117] in This is the token at that time step. The tokens from all time steps form the predicted token sequence. The quantization process forces continuous and diverse latent vectors to be mapped onto a finite number of discrete prototypes, compelling the model to learn typical behavioral patterns.

[0118] Decoder: Receives the quantized vector sequence (That is, the codeword corresponding to the token), and reconstructs the trajectory segment with the same dimension as the input through another set of neural networks (such as Transformer or MLP). The goal of the decoder is to make the reconstructed trajectory as close as possible to the original input, thereby ensuring that the token sequence retains the core information of the original trajectory.

[0119] (3) Calculate the loss based on the pre-constructed loss function, the driving trajectory segment samples, the trajectory decoding segments, and the predicted token sequence, wherein the mathematical expression of the loss function is:

[0120]

[0121] In the formula, Indicates loss, This represents a sample of a driving trajectory segment. This represents a trajectory decoding segment. This represents the sequence of latent space vectors output by the encoder. This indicates that the gradient operation is stopped. For the codeword sequence in the codebook, The index is the token sequence. For loss balancing parameters;

[0122] The reconstruction loss is defined as the loss that forces the decoder's output trajectory to be geometrically and kinematically consistent with the input trajectory. By minimizing the mean squared error, the token sequence (i.e., the index of the selected codewords) must contain enough information to recover the original trajectory. This is fundamental to ensuring that the discrete representation does not lose key behavioral features.

[0123] This term represents the codebook loss, which reduces the codeword vectors in the codebook. The latent vector output to the encoder Close to. Due to Indicates to The gradient is stopped; this loss only updates the codebook parameters and does not affect the encoder. By continuously adjusting the codewords, each codeword corresponding to a token can represent the latent space center of its region, thus forming a semantically clear dictionary of atomic behaviors.

[0124] This represents the commitment loss, which imposes a constraint on the encoder, causing its output latent vector to... Don't stray too far from your chosen code. This indicates that gradients are stopped for codewords, therefore this loss only updates the encoder. Parameters (Typically set to 0.25) This balances the encoder's degrees of freedom and quantization error. Without this, the encoder might arbitrarily map latent vectors to regions far from the codeword, resulting in severe information loss after quantization.

[0125] (4) The trajectory segmenter is trained based on the loss.

[0126] Specifically, in the first stage, the codebook is frozen, and the encoder and the decoder are trained based on the loss;

[0127] At the end of the first stage, the second stage begins and the codebook is decoded. The encoder, the codebook, and the decoder are trained based on the loss until training is complete.

[0128] In the first stage, at the start of training, the codebook is randomly initialized and completely mismatched with the true latent space distribution. Immediately allowing the codebook to learn from the encoder output could cause the codewords to get stuck in local optima, or even result in a large number of "dead codewords" (codewords that were never selected). Without codebook intervention, the encoder and decoder learn how to compress trajectories into the latent space and recover them through a pure reconstruction task (similar to a regular autoencoder). At this point, the latent vectors... It can move freely, unconstrained by quantization, thus quickly forming an effective continuous representation of the input trajectory. This provides a good initial latent space for subsequent quantization learning, preventing the codebook from degrading in chaotic early updates.

[0129] In the second stage, after the first stage ends (usually marked by the reconstruction error decreasing to a stable state or reaching a preset number of rounds), the codebook is set to a trainable state. At this point, all three modules (encoder, codebook, and decoder) participate in optimization simultaneously. After sufficient joint training, the three losses reach a balance: small reconstruction error (good trajectory fidelity), small quantization error (high precision of discrete representation), and high codebook utilization (no dead codewords). The resulting trajectory segmenter compresses the trajectory information into a discrete token sequence while simultaneously restoring the original trajectory with high quality.

[0130] After training, the trajectory segment data is compressed into a token sequence using an encoder, and the context information is reduced in dimension and compressed based on PCA (Principal Component Analysis) or an autoencoder. The composition of the context information is shown in the table below:

[0131] Table 1. Composition of contextual information in difficult scenarios

[0132]

[0133] S240, construct a scene compressed data packet based on the target token sequence, the dimensionality reduction feature information, and the scene metadata, and send the scene compressed data packet to the cloud. The specific process includes:

[0134] (1) Assemble the target token sequence (integer array), dimensionality reduction feature information (floating-point vector), and scene metadata (timestamp, location code, vehicle ID, etc.) into a binary or JSON format data block in a pre-defined order. To reduce redundancy, variable-length encoding (such as Varint) can be used to store the token sequence, and half-precision floating-point numbers (FP16) can be used to store the feature vector.

[0135] (2) Apply a lightweight general-purpose compression algorithm (such as LZ4 or Snappy) to the serialized data blocks to further reduce the transmission volume.

[0136] (3) Encapsulate the compressed data block into a data packet with a header. The header includes a scene unique identifier, data version number, checksum (used for integrity verification) and total data packet length.

[0137] (4) The vehicle maintains a priority queue, based on the difficulty of the scenario. Data packets are sorted. When a batch quantity is reached (e.g., 50 packets) or a timed trigger is triggered (e.g., every 10 seconds), multiple data packets are merged into a batch upload unit, and interrupted transmission and burst transmission mechanisms are used to send them to the designated receiving endpoint in the cloud via the 5G network.

[0138] (5) After receiving the data from the cloud, the vehicle will return an ACK. If the vehicle receives the acknowledgment, the data packet will be deleted from the cache. If no acknowledgment is received within the time limit, a retransmission will be triggered to ensure that the data arrives reliably.

[0139] S250, the scene data is reconstructed in the cloud based on the scene compression data packet to obtain trajectory fragment data, context information and scene metadata of the difficult scene; and the autonomous driving model is iteratively trained based on the trajectory fragment data and context information of the difficult scene.

[0140] In one embodiment of this application, scene data is reconstructed in the cloud based on the scene compressed data packet to obtain trajectory fragment data, context information, and scene metadata of the difficult scene, including:

[0141] S251, the target token sequence in the scene compressed data packet is decoded based on the decoder of the trajectory segmenter to obtain trajectory fragment data; and the dimensionality reduction feature information in the scene compressed data packet is upgraded to obtain context information, and the scene metadata in the scene compressed data packet is directly extracted.

[0142] After reconstructing the difficult scenario, the training process includes:

[0143] S252, the context information is used as a data sample, and the trajectory segment data is used as a label to construct training data samples;

[0144] S253, perform data augmentation on the training data samples to obtain augmented data samples and an augmented data sample set, wherein the data augmentation includes geometric transformation, temporal perturbation and parameter perturbation;

[0145] S254, The end-to-end autonomous driving model in the cloud is trained based on the enhanced data sample set and a pre-configured training strategy, wherein the training strategy includes a course learning strategy, a hard sample mining strategy, an adversarial training strategy, and an incremental training strategy.

[0146] The training process based on training strategies includes:

[0147] S254-1, Based on the comprehensive difficulty score, the samples in the enhanced data sample set are arranged in ascending order;

[0148] Specifically, the training set Each sample Attached is its difficulty rating. Then according to Sort by size from smallest to largest.

[0149] S254-2, in the early stages of training, uses the simplest multiple samples combined with standard gradient descent to incrementally train the end-to-end autonomous driving model in the cloud.

[0150] This allows the model to quickly grasp simple scenarios, establish a reliable foundation, and avoid being misled by difficult examples or noise from the outset.

[0151] S254-3, during the middle of training, freezes the parameters of the end-to-end autonomous driving model in the cloud and calculates the loss for each augmented data sample. Augmented data samples with losses exceeding a preset loss threshold are designated as hard samples, and a hard sample set is constructed based on these hard samples. Assign acquisition weights to each augmented data sample based on the loss. The collection weights of non-difficult samples are multiplied by a decay factor (e.g., 0.1) to highlight difficult samples; and augmented data samples are collected based on the collection weights of all samples, combined with standard gradient descent, to incrementally train the end-to-end autonomous driving model in the cloud.

[0152] Training according to weights Sampling with replacement is performed, with each batch primarily consisting of difficult samples. The loss function remains the behavior cloning loss, without an adversarial term. A smaller regularization coefficient can be added at this point. (e.g., 0.01) to prevent overfitting to difficult examples, thus allowing the model to focus on learning the currently poorly performing samples and specifically improve the accuracy of difficult scenarios.

[0153] S254-4, In the later stage of training, the sampling weight mechanism is retained, and adversarial examples that maximize the loss of each augmented data sample are generated. The end-to-end autonomous driving model in the cloud is incrementally trained based on the augmented data samples and the adversarial examples of the augmented data samples. The adversarial examples are generated by adding perturbations to the augmented data samples.

[0154] In the later stages, the entire training set is used (no longer filtered by course), but weighted sampling is still retained (optional).

[0155] For each sample in each batch :

[0156] (1) Generate adversarial examples ,in, For disturbance, (Approximated by PGD or FGSM).

[0157] (2) During training, augmented data samples will be used. Adversarial examples and augmented data samples The data is input into the model sequentially, and the total loss of both is calculated:

[0158]

[0159] In the formula, This is the loss balancing parameter, typically set to 0.5.

[0160] The above process improves the model's robustness to small input perturbations through adversarial training, preventing dangerous behavior in real-world environments due to noise or slight perceptual errors.

[0161] S255 performs differential updates to the vehicle's end-to-end autonomous driving model based on the trained cloud-based end-to-end autonomous driving model.

[0162] In this application, when updating the vehicle-side model, only the parameter changes are transmitted to avoid catastrophic omissions.

[0163] In addition, the new model will be validated on a selection of vehicles before being rolled out nationwide; it also supports hot updates of the model during vehicle operation.

[0164] The vehicle records the performance of the new model and feeds it back to the cloud; the cloud continuously optimizes the model training strategy based on the feedback data.

[0165] The specific embodiments of this application are as follows:

[0166] The implementation environment of this invention includes:

[0167] (1) Vehicle-side environment:

[0168] - Computing platform: NVIDIA DriveOrin;

[0169] - Communication module: 5G communication module;

[0170] - Deployment scale: 100 L4 level unmanned logistics vehicles.

[0171] (2) Cloud environment:

[0172] - Computing resources: 1000+ GPU computing nodes;

[0173] - Storage resources: Distributed object storage system;

[0174] - Training framework: PyTorch distributed training framework.

[0175] Data compression effect

[0176] In this embodiment, the data compression effect is as follows:

[0177] (1) Raw data volume: Approximately 500MB for a single difficult scenario (including 2 seconds of sensor data);

[0178] (2) Token sequence size: approximately 2KB (containing 200 tokens);

[0179] (3) Context feature size: approximately 50KB (feature vector after dimensionality reduction);

[0180] (4) Total compressed size: approximately 52KB;

[0181] (5) Compression ratio: approximately 10000:1.

[0182] Communication efficiency

[0183] In this embodiment, the efficiency metrics of vehicle-to-cloud communication are as follows:

[0184] (1) Average daily data upload volume: approximately 100MB per vehicle (approximately 2000 challenging scenarios);

[0185] (2) Upload bandwidth requirement: approximately 10Kbps average bandwidth per vehicle;

[0186] (3) Model distribution size: Differential update is approximately 50MB;

[0187] (4) Update frequency: The model is updated once a week.

[0188] Training effect

[0189] In this embodiment, the model training results are as follows:

[0190] (1) Ability to handle difficult scenarios: improved by more than 40%;

[0191] (2) Model iteration cycle: shortened from 3 months to 1 month;

[0192] (3) Data utilization efficiency: improved by more than 100 times;

[0193] (4) Operating costs: reduced by more than 90%.

[0194] Vehicle-side system deployment

[0195] (1) Track segmenter integration: Deploy the pre-trained track segmenter to the vehicle-side computing platform;

[0196] (2) Development of the difficult scene detection module: Implement a multi-dimensional difficult scene recognition algorithm;

[0197] (3) Data compression module development: Implement token encoding and feature dimensionality reduction functions;

[0198] (4) Communication module development: Implement data upload and model download functions.

[0199] Cloud system deployment

[0200] (1) Data receiving service: Deploy a high-concurrency data receiving service;

[0201] (2) Scene reconstruction service: Deploy a token-based scene reconstruction service;

[0202] (3) Model training platform: Build a distributed model training platform;

[0203] (4) Model distribution service: Deploy model version management and distribution services.

[0204] System integration and optimization

[0205] (1) End-to-end testing: Verify the integrity of vehicle-side data upload and cloud-based scene reconstruction;

[0206] (2) Performance optimization: Optimize token encoding and decoding performance to reduce vehicle-side computing overhead;

[0207] (3) Compression optimization: Adjust the feature dimensionality reduction parameters to balance the compression ratio and reconstruction quality;

[0208] (4) Collaborative optimization: Optimize vehicle-cloud collaboration strategy to improve overall efficiency.

[0209] The main features of this application include:

[0210] (1) Extreme compression ratio: A data compression ratio of 10000:1 is achieved through tokenization, which greatly reduces communication and storage costs;

[0211] (2) Precise screening: Multi-dimensional difficult scene identification indicators to accurately locate samples that are valuable for model training;

[0212] (3) Efficient reconstruction: The training scenario is efficiently reconstructed in the cloud based on a compact token sequence;

[0213] (4) Rapid iteration: The vehicle-cloud collaboration mechanism shortens the model iteration cycle from monthly to weekly;

[0214] (5) Closed-loop optimization: A complete closed loop from data collection to model update, supporting continuous optimization.

[0215] The beneficial effects of this application include:

[0216] (1) Significantly reduce operating costs: Data compression and precise filtering reduce communication and storage costs by more than 90%;

[0217] (2) Significantly improves model capabilities: Targeted training improves the ability to handle difficult scenarios by more than 40%;

[0218] (3) Accelerate model iteration speed: The iteration cycle is shortened to 1 / 3 of the original, and the market demand is responded to quickly;

[0219] (4) Improve data utilization efficiency: accurately extract valuable samples from massive amounts of data, improving data utilization efficiency by 100 times;

[0220] (5) Supports large-scale deployment: Low-cost data collection solutions support large-scale fleet deployment and operation.

[0221] like Figure 3 As shown, this application also provides a vehicle-cloud collaborative autonomous driving model iterative training system, including:

[0222] The acquisition module is used to acquire trajectory segment data of the vehicle in multiple historical scenarios, probability distribution of model predicted actions, and manual takeover information, wherein the model predicted actions are output by an end-to-end model deployed on the vehicle.

[0223] The scene filtering module is used to filter out difficult scenes based on the trajectory segment data, the probability distribution of the model's predicted actions, and the information on human intervention.

[0224] The data dimensionality reduction module is used to input the trajectory fragment data of the difficult scenario into a pre-built trajectory segmenter to obtain the target token sequence; and to extract the context information and scene metadata of the difficult scenario, and to reduce the dimensionality of the context information of the difficult scenario to obtain dimensionality reduction feature information. The context information includes environmental perception features, map features, vehicle state features and model intermediate layer features, and the scene metadata includes time, location and vehicle ID.

[0225] The upload module is used to construct a scene compressed data packet based on the target token sequence, the dimensionality reduction feature information and the scene metadata, and send the scene compressed data packet to the cloud;

[0226] The iterative training module is used to reconstruct scene data in the cloud based on the scene compressed data package to obtain trajectory fragment data, context information and scene metadata of the difficult scene; and to iteratively train the autonomous driving model based on the trajectory fragment data, context information and scene metadata of the difficult scene.

[0227] This application presents a vehicle-cloud collaborative autonomous driving model iterative training method and system. Based on multi-dimensional indicators such as trajectory token anomalies, prediction confidence, and human intervention, the system can accurately identify the key scenarios most valuable for model improvement, avoiding the omission of key data caused by traditional random sampling, and significantly improving the quality and diversity of training data. A trajectory segmenter is used to compress trajectory fragments into compact token sequences, and contextual feature dimensionality reduction is combined to greatly reduce the amount of uploaded data, effectively alleviating bandwidth pressure, reducing cloud storage overhead, and shortening the data transmission and processing cycle, thus accelerating model iteration. The uploaded content is only the token sequence and dimensionality-reduced features, rather than the original sensor data, thereby reducing the risk of leakage of sensitive information such as road environment and pedestrians. The cloud can reconstruct high-fidelity scenarios based on the compact data, and combined with targeted training strategies, achieve targeted learning for difficult scenarios; coupled with the feedback loop of vehicle-cloud collaboration, the model's performance in difficult scenarios is continuously optimized, improving the overall system's iterative efficiency and adaptability.

[0228] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0229] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory so that the terminal performs any of the methods in this embodiment.

[0230] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0231] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0232] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0233] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0234] In the above embodiments, although the present application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. The embodiments of the present application are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0235] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. An iterative training method for an autonomous driving model based on vehicle-cloud collaboration, characterized in that, Including the following steps: The system acquires trajectory segment data of the vehicle in multiple historical scenarios, probability distribution of model-predicted actions, and information on manual takeover, wherein the model-predicted actions are output by an end-to-end model deployed on the vehicle. Difficult scenarios were selected based on the trajectory segment data, the probability distribution of the model's predicted actions, and information on human intervention. The trajectory fragment data of the difficult scenario is input into a pre-built trajectory segmenter to obtain the target token sequence; and the context information and scene metadata of the difficult scenario are extracted. The context information of the difficult scenario is dimensionality reduced to obtain dimensionality reduction feature information. The context information includes environmental perception features, map features, vehicle state features and model intermediate layer features. The scene metadata includes time, location and vehicle ID. A scene compressed data packet is constructed based on the target token sequence, the dimensionality reduction feature information, and the scene metadata, and the scene compressed data packet is sent to the cloud; The scene data is reconstructed in the cloud based on the scene compression data packet to obtain trajectory fragment data, context information and scene metadata of the difficult scene; and the autonomous driving model is iteratively trained based on the trajectory fragment data and context information of the difficult scene.

2. The iterative training method for vehicle-cloud collaborative autonomous driving model according to claim 1, characterized in that, Based on the trajectory segment data, the probability distribution of model-predicted actions, and information on human intervention, difficult scenarios were identified, including: The trajectory segment data is encoded using a pre-built trajectory segmenter to obtain the token encoding result of the trajectory segment data; the target codeword most similar to the token encoding result is extracted from the codebook of the trajectory segmenter, and the Euclidean distance between the token encoding result and the target codeword is calculated. ; Calculate the entropy of the probability distribution of the predicted actions by the model. And extract the highest probability ; Constructing indicator functions for manual takeover information ; For the Euclidean distance respectively and the entropy value Normalization is performed to obtain the normalized distance. and normalized entropy value ; and based on the normalized distance The normalized entropy value The maximum probability and the indicated function Calculate the overall difficulty score The comprehensive difficulty score, among which... The mathematical expression is: In the formula, As the first weight, As the second weight, As the third weight, It is the fourth weight; Based on the Euclidean distance The entropy value The maximum probability The indicated function Or overall difficulty score Filter out difficult scenarios.

3. The iterative training method for vehicle-cloud collaborative autonomous driving model according to claim 2, characterized in that, Based on the Euclidean distance The entropy value The maximum probability The indicated function Or overall difficulty score Filtering difficult scenarios includes: A scenario that satisfies any one of the target conditions is considered a suffering scenario. The target conditions include: The Euclidean distance The distance is greater than the preset distance threshold; The entropy value Greater than the preset entropy threshold; The maximum probability Less than the preset probability threshold; The value of the indicator function indicates that manual intervention is present; The overall difficulty score is greater than the preset difficulty threshold.

4. The iterative training method for vehicle-cloud collaborative autonomous driving model according to claim 1, characterized in that, The trajectory segmenter includes an encoder for mapping trajectory segments to a latent space vector sequence, a codebook for converting the latent space vector sequence into a token sequence, and a decoder for mapping the token sequence to trajectory decoding segments. The method for constructing the trajectory segmenter includes: Obtain driving trajectory segment samples; The trajectory segment sample is input into the encoder to obtain the predicted token sequence output by the codebook and the trajectory decoded segment output by the decoder; The loss is calculated based on a pre-constructed loss function, the driving trajectory segment samples, the trajectory decoding segments, and the predicted token sequence, wherein the mathematical expression of the loss function is: In the formula, Indicates loss, This represents a sample of a driving trajectory segment. This represents a trajectory decoding segment. This represents the sequence of latent space vectors output by the encoder. This indicates that the gradient operation is stopped. For the codeword sequence in the codebook, The index is the token sequence. For loss balancing parameters; The trajectory segmenter is trained based on the loss.

5. The iterative training method for vehicle-cloud collaborative autonomous driving model according to claim 1, characterized in that, The scene data is reconstructed in the cloud based on the scene compression data packet to obtain trajectory fragment data, context information, and scene metadata of the difficult scene, including: The decoder based on the trajectory segmenter decodes the target token sequence in the scene compressed data packet to obtain trajectory fragment data; and it upscales the dimensionality reduction feature information in the scene compressed data packet to obtain context information, and directly extracts the scene metadata in the scene compressed data packet.

6. The iterative training method for vehicle-cloud collaborative autonomous driving model according to claim 2, characterized in that, Iterative training of autonomous driving models based on trajectory fragment data and contextual information from challenging scenarios includes: The context information is used as a data sample, and the trajectory segment data is used as a label to construct training data samples; The training data samples are augmented to obtain augmented data samples and an augmented data sample set, wherein the data augmentation includes geometric transformation, temporal perturbation and parameter perturbation; The end-to-end autonomous driving model in the cloud is trained based on the enhanced data sample set and a pre-configured training strategy, wherein the training strategy includes a course learning strategy, a hard sample mining strategy, an adversarial training strategy, and an incremental training strategy. The end-to-end autonomous driving model on the vehicle is differentially updated based on the trained cloud-based end-to-end autonomous driving model.

7. The iterative training method for vehicle-cloud collaborative autonomous driving model according to claim 6, characterized in that, The end-to-end autonomous driving model in the cloud is trained based on the enhanced data sample set and a pre-configured training strategy, including: Based on the comprehensive difficulty score, the samples in the augmented data sample set are arranged in ascending order; In the early stages of training, the simplest multiple samples combined with standard gradient descent are used to incrementally train the end-to-end autonomous driving model in the cloud. During the middle of training, the parameters of the end-to-end autonomous driving model in the cloud are frozen, and the loss of each augmented data sample is calculated. Augmented data samples with losses greater than a preset loss threshold are designated as hard samples, and a hard sample set is constructed based on the hard samples. Acquisition weights are assigned to each augmented data sample based on the loss, where the acquisition weights of non-hard samples are multiplied by a decay factor. Augmented data samples are acquired based on the acquisition weights of all samples and combined with standard gradient descent to incrementally train the end-to-end autonomous driving model in the cloud. In the later stages of training, the sampling weight mechanism is retained, and adversarial examples that maximize the loss of each augmented data sample are generated. The end-to-end autonomous driving model in the cloud is incrementally trained based on the augmented data samples and the adversarial examples of the augmented data samples, wherein the adversarial examples are generated by adding perturbations to the augmented data samples.

8. A vehicle-cloud collaborative autonomous driving model iterative training system, characterized in that, include: The acquisition module is used to acquire trajectory segment data of the vehicle in multiple historical scenarios, probability distribution of model predicted actions, and manual takeover information, wherein the model predicted actions are output by an end-to-end model deployed on the vehicle. The scene filtering module is used to filter out difficult scenes based on the trajectory segment data, the probability distribution of the model's predicted actions, and the information on human intervention. The data dimensionality reduction module is used to input the trajectory fragment data of the difficult scenario into a pre-built trajectory segmenter to obtain the target token sequence; and to extract the context information and scene metadata of the difficult scenario, and to reduce the dimensionality of the context information of the difficult scenario to obtain dimensionality reduction feature information. The context information includes environmental perception features, map features, vehicle state features and model intermediate layer features, and the scene metadata includes time, location and vehicle ID. The upload module is used to construct a scene compressed data packet based on the target token sequence, the dimensionality reduction feature information and the scene metadata, and send the scene compressed data packet to the cloud; The iterative training module is used to reconstruct scene data in the cloud based on the scene compressed data package to obtain trajectory fragment data, context information and scene metadata of the difficult scene; and to iteratively train the autonomous driving model based on the trajectory fragment data, context information and scene metadata of the difficult scene.

9. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images