Robotic complex scene social self-adaptive navigation knowledge transfer method and system

Abstract

This invention relates to the field of human-computer interaction navigation technology, and discloses a method and system for knowledge transfer in complex social adaptive navigation of robots. The method includes: constructing a multi-level progressive human-computer collaborative navigation scenario library; generating expert demonstration path datasets for each level; constructing a recursive generative model including a GRU social feature encoder based on a gated recurrent unit; determining model parameters through ablation experiments; embedding the GRU social feature encoder into the node cost evaluation module of the RRT* path planning algorithm to generate a GRU-RRT* path planner; training the GRU-RRT* path planner using a simple social scenario library; and performing knowledge transfer using a complex social scenario library. This invention can effectively learn long-term dependencies in social interactions, improving the anthropomorphism and social compliance of navigation paths.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction navigation technology, and in particular to a method and system for knowledge transfer in social adaptive navigation of robots in complex scenarios. Background Technology

[0002] With the rapid development of artificial intelligence technology, service mobile robots are undergoing profound changes in the field of human-computer interaction and navigation. Their navigation behavior not only exhibits increasingly human-like characteristics, but also demonstrates superior knowledge transfer and adaptive capabilities when facing new environments, thus enabling them to learn a more comprehensive and diverse social navigation knowledge system.

[0003] Social adaptive navigation, as an interdisciplinary field, integrates theories and methods from robotics, machine learning, psychology, and sociology. It aims to enable robots to achieve safe, efficient, and socially compliant autonomous movement in dynamic scenarios involving human activity. In recent years, this field has made significant progress in knowledge acquisition and transfer. In navigation decision modeling, recurrent neural networks (RNNs), due to their inherent ability to process temporal data, have been widely introduced into social navigation research to model continuous state sequences and capture the temporal dependencies in robot-dynamic interactions. Simultaneously, in human-robot collaborative environments, embedding recurrent neural networks as inverse reinforcement learning modules within the RRT* planner allows the path planner to acquire social adaptive navigation behavior knowledge while learning expert-demonstrated paths. When facing new scenarios, iterative optimization of network parameters facilitates the transfer of navigation knowledge. This enables the path planner to utilize learned knowledge to guide the robot in planning more human-like paths that conform to social habits in the interactive environment.

[0004] However, while existing RNN-RRT* path planners incorporate recurrent neural networks with memory capabilities and can utilize historical state information for sequence modeling to some extent, they still have several limitations, mainly: 1. Traditional recurrent neural networks (RNNs) are prone to gradient vanishing or exploding problems when modeling long sequence interactions, resulting in insufficient ability to learn long-term dependencies in social interactions in long sequences, which affects the prediction effect.

[0005] 2. Recurrent Neural Networks (RNNs) have a relatively simple structure and gating mechanism, making it difficult to adaptively select and forget historical navigation information. Therefore, when dealing with dense, dynamic human-machine collaborative scenarios with high social rule constraints, they often exhibit problems such as low anthropomorphism in path generation and limited social acceptance. They also struggle to continuously model and predict pedestrian intentions and behavioral patterns, lacking sufficient social adaptive navigation capabilities. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for knowledge transfer in social adaptive navigation of robots in complex scenarios, which can effectively learn the long-term dependencies of social interactions and improve the anthropomorphism and social compliance of navigation paths.

[0007] To address the aforementioned technical problems, this invention provides a method for knowledge transfer in social adaptive navigation of robots in complex scenarios, comprising: A multi-level, progressive human-machine collaborative navigation scenario library is constructed, wherein the levels of the human-machine collaborative navigation scenario library include a simple social scenario library and a complex social scenario library; Based on the different levels in the human-machine collaborative navigation scenario library, an expert demonstration path dataset is generated for each level. A recursive generation model based on gated loop units is constructed. The recursive generation model based on gated loop units includes a GRU social feature encoder constructed based on gated loop units. A hierarchical dataset consisting of expert demonstration path datasets corresponding to simple social scenarios is input into the recursive generation model based on gated loop units. The recursive generative model based on gated loop units is trained, and the model parameters of the recursive generative model based on gated loop units are determined through ablation experiments. The GRU social feature encoder is embedded into the node cost evaluation module of the RRT* path planning algorithm to generate a GRU-RRT* path planner. All simple social scenarios in the simple social scenario library are set as a training loop and the GRU-RRT* path planner is trained. The GRU-RRT* path planner trained in simple social scenarios is used as the social navigation behavior meta-model. Prior knowledge is introduced, and the same training method used in simple social scenarios is used to train the social navigation behavior meta-model in complex social scenarios, resulting in a trained recursive generative model based on gated recurrent units.

[0008] Further, training the GRU-RRT* path planner includes: Calculate the social adaptive navigation feature vector of each node on each expert demonstration path to generate a social cost graph representing the social rules of the current simple social scenario; Based on the social cost graph of social rules in simple social scenarios, search and generate planned paths; The recursive generation module completes the learning of social adaptive navigation feature knowledge in simple social scenarios. It determines whether the training cycle has ended. If it has ended, the GRU-RRT* path planner training is complete. If it has not ended, it updates the next simple social scenario and returns to the step of calculating the social adaptive navigation feature vector of each node on each expert demonstration path.

[0009] Furthermore, when determining the model parameters of the recursive generation model based on the gated recurrent unit through ablation experiments, the model parameters are selected by constructing a comprehensive efficiency score, which is: , In the formula, To score overall efficiency, To score performance, This is the performance rating coefficient. Scoring is given for convergence speed. This is the convergence speed scoring coefficient. Score the model complexity. This is the model complexity scoring coefficient. Rate time efficiency. This is a time efficiency rating coefficient. This is the Dropout penalty factor.

[0010] Furthermore, the search and generation of planned paths based on the social cost graph of simple social scenarios and social rules includes: Within the feasible state space of a simple social scenario, a new potential growth point is selected through random sampling. This new potential growth point is denoted as... ; right Perform collision detection. If the detection fails, return to the step of selecting a new potential growth point through random sampling; if the detection succeeds, then... As an effective growth point, let the effective growth point be denoted as... ; calculate The social adaptive navigation feature vector is input into the GRU social feature encoder. The GRU social feature encoder fuses the current input with the hidden state that encodes the historical information of past path nodes through a gating mechanism to obtain the hidden state at the current moment. A comprehensive social cost function is constructed, and the comprehensive social cost value is calculated based on the hidden state at the current moment. Based on the calculation The instantaneous comprehensive social value is calculated from the starting point through potential parent nodes. The cumulative comprehensive social cost value; from the neighborhood node set of the random tree, select the node that minimizes the cumulative comprehensive social cost value as... Find the optimal parent node and establish a connection; judge If the distance to the target point exceeds a threshold, return to the step of selecting a new potential growth point by random sampling; if it does not exceed the threshold, it is considered that the destination has been reached and the path planning ends.

[0011] Furthermore, the comprehensive social cost function is: , In the formula, This represents the comprehensive social value at time t. This represents the hidden state at time t. It is a nonlinear fusion function. For efficiency cost components, characterization Within the implied temporal context, For stability cost, characterization in The degree of abruptness in the robot's future navigation behavior relative to its recent historical behavior patterns. For compliance cost components, characterization in The severity of violating social norms by choosing this node in the scenario it reflects.

[0012] Furthermore, the aforementioned The calculation method is as follows: , In the formula, The reference time increment is related to the geometric distance between nodes. It is a nonlinear mapping; The The calculation method is as follows: , In the formula, The diagonal weight matrix is ​​learnable. It is a moving average or smoothed representation of the recent historical hidden state; The The calculation method is as follows: , In the formula, Let T be the learnable weight vector, T be the total training time, and exp be the exponential function. Let d be the indicator function, and d be the robot's expected intrusion distance. For the comfort threshold, when hour ,when hour .

[0013] Furthermore, the nonlinear fusion function is specifically: , In the formula, , , To integrate the weighting coefficients, , To fuse the order parameters, >1.

[0014] Furthermore, the iterative update process of the social navigation behavior meta-model during training is specifically as follows: The social navigation behavior metamodel extracts the social adaptive navigation feature vectors of each node on the currently generated planned path to form an input sequence, denoted as . ,Will The input sequence serves as the input sequence for the social navigation behavior meta-model; the comprehensive social cost value of the corresponding nodes on the expert demonstration path is extracted to form the target value sequence, denoted as . ,Will As a meta-model of social navigation behavior, the target value sequence that needs to be fitted needs to be aligned. and ; Aligned Input the social navigation behavior meta-model sequentially, and output a sequence of predicted social costs for each node of the planned path; An adaptive weighting mechanism for predicting uncertainty is constructed, and a loss function is built by combining the predicted social cost value sequence to train the meta-model of social navigation behavior. The gradient of the loss function with respect to all trainable parameters of the social navigation behavior meta-model is calculated using automatic differentiation techniques. The Adam optimizer is used to iteratively update the parameters of the social navigation behavior meta-model using the gradient.

[0015] Furthermore, the loss function for training the social navigation behavior meta-model is: , In the formula, The loss function used when training the social navigation behavior meta-model, where n is the number of nodes. The dynamic adaptive weight of the i-th node. To demonstrate the comprehensive social value of the i-th node on the expert demonstration path. Let be the predicted social cost of the i-th node on the planned path; The method for calculating the dynamic adaptive weight of the i-th node is as follows: , In the formula, It is an exponential function. To focus the intensity coefficient, This represents the average of the predicted social costs for the current batch. The standard deviation of the predicted social costs for the current batch. As a preset constant, These are complexity weighting coefficients. Let i be the social adaptive navigation feature vector of the i-th node on the planning path. for The variance.

[0016] This invention also provides a robot social adaptive navigation knowledge transfer system for complex scenarios, comprising: The scenario building module is used to build a multi-level progressive human-machine collaborative navigation scenario library, the levels of which include a simple social scenario library and a complex social scenario library; The dataset generation module is used to generate an expert demonstration path dataset for each level based on different levels in the human-machine collaborative navigation scenario library; The recursive generation model module is used to construct a recursive generation model based on gated loop units. The recursive generation model based on gated loop units includes a GRU social feature encoder constructed based on gated loop units. A hierarchical dataset consisting of expert demonstration path datasets corresponding to simple social scenarios is input into the recursive generation model based on gated loop units. The ablation training module is used to train the recursive generative model based on the gated loop unit and to determine the model parameters of the recursive generative model based on the gated loop unit through ablation experiments. The simple social scene training module is used to embed the GRU social feature encoder into the node cost evaluation module of the RRT* path planning algorithm to generate a GRU-RRT* path planner, and to set all simple social scenes in the simple social scene library as a training cycle and train the GRU-RRT* path planner. The complex social scenario knowledge transfer module is used to take the GRU-RRT* path planner trained in simple social scenarios as a social navigation behavior meta-model, introduce prior knowledge, and use the same training method as in simple social scenarios to train the social navigation behavior meta-model in complex social scenarios, resulting in a trained recursive generative model based on gated recurrent units.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention employs gated recurrent units (RNNs) to construct a GRU-RRT* path planner, replacing traditional RNNs. By leveraging the collaborative control mechanism of the update and reset gates within the gated recurrent units, it achieves more refined dynamic filtering and memory management of scene temporal features. This significantly enhances the model's long-term prediction and understanding capabilities for complex dynamic interaction patterns such as pedestrian social intentions, group movement trends, and scene semantic rules. Furthermore, it continuously optimizes social adaptive navigation strategies through adaptive weighted memory and forgetting of historical navigation states. Simultaneously, this invention introduces a hierarchical transfer learning paradigm into the GRU-RRT* path planner framework, focusing on social adaptive navigation scenarios. This enables the GRU-RRT* path planner to not only recursively learn social adaptive navigation knowledge in simple scenarios but also achieve more stable, efficient, and generalizable cross-scenario knowledge transfer in complex human-machine collaborative environments with rich social intentions and challenging scene navigation. Ultimately, this improves the anthropomorphism, social compliance, and acceptability of navigation paths in practical applications. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating the steps of a method in a preferred embodiment of the present invention.

[0020] Figure 3 This is an example diagram of a multi-level progressive human-machine collaborative navigation scene library in a preferred embodiment of the present invention.

[0021] Figure 4 This is an example diagram of an expert demonstration path in a preferred embodiment of the present invention.

[0022] Figure 5 This is a structural diagram of the GRU social feature encoder in a preferred embodiment of the present invention.

[0023] Figure 6 This is an example diagram of the paths planned by the GRU-RRT* path planner in a simple social scenario before and after training, according to a preferred embodiment of the present invention.

[0024] Figure 7 This is an example diagram of the paths planned by the GRU-RRT* path planner before and after migration in a complex social scene, according to a preferred embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0026] This invention discloses a knowledge transfer method for social adaptive navigation in complex robot scenarios, such as... Figure 1 and Figure 2 As shown, it includes the following steps: S1: Construct a multi-level progressive human-machine collaborative navigation scenario library. The human-machine collaborative navigation scenario library includes a simple social scenario library and a complex social scenario library. The human-machine collaborative navigation scenario library is different from the homogenization or randomization of scenarios in the existing technology. It is a scenario library with a clear complexity gradient.

[0027] In this embodiment, the human-machine collaborative navigation scenario library includes two levels, such as... Figure 3 As shown, this includes both simple and complex social scenarios. This design aims to provide a structured data foundation for subsequent incremental knowledge transfer. Figure 3 (a) is a simple social scene, including static obstacles, the robot's starting and target points, and a small number of pedestrians (no more than 3). Figure 3 In the case of complex social scenarios, more (no less than 4) dynamic pedestrians with interactive intentions are introduced on the basis of static environment to form complex social avoidance challenges for the learning and transfer verification of adaptive navigation knowledge in complex society.

[0028] S2: Based on the different levels in the human-machine collaborative navigation scenario library, generate an expert demonstration path dataset for each level. The expert demonstration path dataset includes a simple expert demonstration path dataset and a complex expert demonstration path dataset. Specifically, the simple expert demonstration path dataset consists of M demonstration paths generated by an operator remotely controlling a robot in M ​​(e.g., 60) simple social scenarios, following basic social rules (such as minimum safe distance and avoiding talking pedestrians). Specifically, the complex expert demonstration path dataset consists of N smooth paths generated by an operator remotely controlling a robot in N (e.g., 18) complex social scenarios, reflecting more complex social etiquette and cooperative avoidance, serving as expert demonstration paths. Examples of expert demonstration paths are shown below. Figure 4 As shown.

[0029] S2 clarifies the hierarchical nature of data sources, supporting the transfer learning strategy of "first learning basic rules in simple scenarios, then adapting to complex situations." The generation process of the expert demonstration path dataset at each level includes the following steps: S2-1: Set the starting point and target point of the navigation task. For each preset human-machine collaborative navigation scenario, set an initial starting point coordinate and a final target point coordinate to define the complete spatial constraints of a single navigation task.

[0030] S2-2: Expert Path Generation Based on Teleoperation. Using the Visual Interaction Tool (RVIZ) integrated into the Robot Operating System (ROS) and a handheld remote control, the operator remotely drives the robot, guiding it along a smooth path that conforms to human social habits from the starting point to the target point. This path serves as the expert demonstration path for this scenario.

[0031] S2-3: Synchronous Acquisition and Encapsulation of Multimodal Scene Data. During the expert demonstration path generation process, the robot's 360-degree LiDAR sensor and positioning system collect and record multimodal scene dynamic information in real time. The collected structured training data mainly includes: 1. The robot's real-time pose sequence in the global coordinate system; 2. The real-time position and heading angle of quasi-dynamic pedestrians in the scene; 3. The relative distance and orientation information of the nearest obstacle. Finally, using the data recording tool (rosbag) provided by the ROS framework, all the above time-synchronized data streams, including the coordinates of the starting point and the target point, as well as the states of all real-time collected objects, are uniformly encapsulated and stored as a standard bag data package, forming the formatted dataset required for subsequent model training.

[0032] In this embodiment, a hierarchical dataset containing 100 scenarios was constructed, including 75 simple social scenarios (60 for training and 15 for testing) and 25 complex social scenarios (18 for transfer training and 7 for transfer testing).

[0033] S3: Construct a recursive generation model based on gated recurrent units. The recursive generation model based on gated recurrent units includes a GRU social feature encoder constructed based on gated recurrent units (GRU). The hierarchical dataset consisting of expert demonstration path datasets corresponding to simple social scenarios (i.e., the bag of M simple expert demonstration path datasets collected in S2) is input into the recursive generation model based on gated recurrent units. The learning of basic social rules and other navigation behavior knowledge in simple social scenarios is completed in the iteration of network parameters.

[0034] The recursive generative model based on gated recurrent units uses a GRU social feature encoder to replace traditional fully connected neural networks or transitive recursive neural networks. The GRU social feature encoder is used to extract and generate social navigation knowledge from expert demonstrations. The structure of the GRU social feature encoder is as follows: Figure 5As shown, the GRU social feature encoder takes a six-dimensional social navigation feature vector (including target distance, obstacle distance, pedestrian social field, and interaction prediction probability, etc.) as input. Through its update gate and reset gate gating mechanism, it adaptively fuses current features with historical hidden states, and finally outputs accurate node social cost values. This design enables the model to effectively capture long-term temporal dependencies, avoid gradient vanishing, and thus have stronger scene memory and continuous decision-making capabilities.

[0035] A state-value implicit mapping model is constructed using a gated recurrent unit. This model, under the supervision of expert demonstration data, learns an intrinsic evaluation criterion for the social adaptability of navigation nodes. Specifically, hierarchical data is input into a recursive generative model based on the gated recurrent unit. This module does not directly fit an explicit reward function formula, but rather uses its hidden states (…). Constructing a sequence of historical states to the social value of nodes ( The end-to-end mapping of ), where, It is a historical state sequence. These are the model parameters for a recursive generation model based on gated loop units.

[0036] The training follows a hierarchical paradigm of "simple first, complex later": first, the model is trained using a basic set of scenarios to master general rules; then, the learned knowledge is transferred to a set of complex scenarios for fine-tuning and adaptation.

[0037] S4: Train the recursive generative model based on gated recurrent units (GRUs), and determine the model parameters of the GRU-based recursive generative model through ablation experiments. The ablation experiments are designed primarily to determine the rationality of the model parameters of the GRU-based recursive generative model. In this embodiment, nine control groups were designed for the number of hidden layer units, the number of hidden layers, the dropout size, and the stride size of the GRU-based recursive generative model, and S hierarchical expert demonstration path datasets were randomly selected to complete the training and testing under different model parameters.

[0038] When determining the model parameters of the recursive generative model based on gated recurrent units through ablation experiments, an efficiency scoring function was designed based on the concept of multi-objective optimization to achieve a balance between model performance, training efficiency, and computational cost. Multiple conflicting optimization objectives were fused into a single scoring index using a weighted linear combination method, facilitating architecture comparison and selection. Model parameters were selected by constructing a comprehensive efficiency score, which is: , In the formula, To score overall efficiency, To score performance, For performance rating coefficients, in this embodiment =0.45, Scoring is given for convergence speed. The convergence speed scoring coefficient is used in this embodiment. =0.25, Score the model complexity. The model complexity scoring coefficient, in this embodiment =0.15, Rate time efficiency. The time efficiency rating coefficient, in this embodiment =0.1, This is the Dropout penalty factor.

[0039] The performance score is calculated as follows: , In the formula, L is the final verification loss value. The intention is to make the score higher as the validation loss is smaller, thereby reflecting the direct and explicit expression of the validation loss on network performance.

[0040] The convergence rate score is calculated as follows: , In the formula, For training rounds, As a penalty factor. In this embodiment, a penalty mechanism is imposed on the convergence speed: (1) premature convergence ( <5 rounds): Penalty factor (1) Reflects the risk of underfitting; (2) Moderate convergence (5≤ ≤30 rounds): Penalty factor , optimal interval; (3) slow convergence (>30 rounds): penalty factor This reflects low training efficiency.

[0041] The method for calculating the model complexity score is as follows: , In the formula, P represents the total number of model parameters. Considering the impact of different GRU network architecture parameters on the number of parameters in the recursive generative network model, following Occam's razor, the fewer the parameters, the higher the score.

[0042] The method for calculating time efficiency scores is as follows: , In the formula, T is the total training time (seconds).

[0043] For the Dropout value, when d > 0.3, a high Dropout penalty is set: based on the information loss theory, an excessively high dropout rate may cause the network to forget important features, especially in RNNs / GRUs, potentially disrupting temporal dependencies; when d < 0.1, a low Dropout penalty is set: based on the overfitting risk, too low a dropout rate may not provide sufficient regularization, potentially leading to a decrease in the model's generalization ability. The optimal Dropout value range is 0.1 ≤ d ≤ 0.3: empirical studies show that this range usually balances the risks of underfitting and overfitting. The specific settings for the Dropout penalty factor are as follows: .

[0044] The specific design and scoring comparison of the model parameters of the recursive generation model based on gated recurrent units in this invention and the model parameters of the nine control groups are shown in Table 1. The final model parameters of the recursive generation model based on gated recurrent units in this invention are determined as follows: 6 input layer units (depending on the social adaptive feature vector of the node), 64 hidden layer units, 2 hidden layers, 1 output layer unit, stride of 10, and dropout of 0.2.

[0045] Table 1. Comparison of specific design and scoring for different model parameters.

[0046] S5: The GRU social feature encoder is embedded into the node cost evaluation module of the RRT* path planning algorithm to replace the cost function of the original linear or simple neural network or traditional recurrent neural network, thereby generating the GRU-RRT* path planner; during the path search process, the GRU-RRT* path planner can call the GRU social feature encoder to perform real-time social comprehensive cost evaluation and prediction that integrates historical information.

[0047] The core function of the GRU social feature encoder lies in fusing multi-dimensional social navigation features and providing accurate social cost estimates for node selection in the path planner. Specifically, for each node to be evaluated during the planning process, a six-dimensional social adaptive navigation feature vector is defined (denoted as ). , This is used to quantify the suitability of the node in a social environment. Among these, The goal-oriented cost represents the Euclidean distance between the current node and the target point. Static obstacle avoidance cost represents the distance from the node to the nearest obstacle; The cost of the dynamic social field is defined by Gaussian functions for the front, back and right sides of the pedestrian, and is used to model the comfort cost of the pedestrian's close area. The cost of pedestrian interaction prediction is used to quantify the probability that the current node is in an area where pedestrians may intersect in the future.

[0048] With nodes For example, the corresponding six feature values ​​are: , , , , , ; In the formula, , , , , , They are respectively The six eigenvalues, This represents the distance from the current node to the target node. This represents the Euclidean distance between two points. Represents the regularization parameter. , This represents the adjustment parameters of the function. express To the The distance between pedestrians express With the From the perspective of a pedestrian express Distance to the convergence point express Distance to the nearest obstacle This indicates the cost or danger ahead for pedestrians. Indicates the cost behind the pedestrian. This indicates the cost to pedestrians on the right side. express The cost of reaching the convergence point Indicates the convergence point. Represents a set of nodes.

[0049] In a dynamic navigation environment where humans and machines coexist, there are complex temporal coupling relationships and dynamic interactive influences among the various dimensions of the social adaptive navigation feature vector (e.g., the direction of pedestrian movement). It will change the social field around it in real time. Meanwhile, navigation decisions heavily rely on the understanding of continuous historical states (such as the evolution trend of pedestrian trajectories). Simple fully connected neural networks, due to their inherent memorylessness, struggle to effectively capture such dependencies and dynamic relationships between features across time steps, resulting in limitations in their knowledge learning for complex, time-varying social scenarios.

[0050] To this end, this invention adopts a recurrent neural network architecture with temporal modeling capabilities to replace the traditional fully connected network. The aim is to enable the model to deeply explore the continuous impact of historical navigation state sequences on current and future planning decisions by explicitly introducing a historical information memory mechanism, thereby achieving more comprehensive and forward-looking social navigation knowledge learning.

[0051] Compared to basic recurrent neural networks (RNNs), this invention further selects gated recurrent units (GRUs) as the core temporal modeling component, possessing the following key advantages: First, fine-grained encoding of dynamic interactions: The update gate and reset gate of the GRU work together to adaptively filter and fuse feature information from different time steps. This enables more accurate modeling, such as... Pedestrian interaction prediction and First, it recognizes the dynamic causal relationships between real-time social fields, rather than simple memory. Second, it ensures robust maintenance of long-term dependencies: In complex scenarios, key social cues (such as pedestrian merging intentions) may appear at earlier times. GRU effectively mitigates the vanishing gradient problem through its gating mechanism, maintaining and propagating such long-range dependencies more stably than standard RNNs, ensuring that historical key events have a continuous and appropriate impact on current planning. Third, it enhances robustness to noisy time-series data: Navigation sensor data streams (such as obstacle distances used for calculation) There is transient noise. GRU's gating mechanism can learn to filter out irrelevant fluctuations in time series data and focus on feature changes with continuous trends, thereby generating more stable and reliable latent state representations in dynamic environments.

[0052] S6: Set all the simple social scenarios in the simple social scenario library as a training loop and train the GRU-RRT* path planner.

[0053] S6-1: Calculate the social adaptive navigation feature vector of each node on each expert demonstration path. The GRU social feature encoder encodes the social adaptive navigation feature vector to obtain the social cost weight corresponding to each node. The cost weights of all nodes are combined to generate a probabilistic social cost map that represents the social rules of the current simple social scenario, which serves as the prior knowledge input for the planner.

[0054] S6-2: The GRU-RRT* path planner searches and generates r planned paths based on the social cost graph of social rules in a simple social scenario; the paths planned by the GRU-RRT* path planner before and after training in a simple social scenario are as follows: Figure 6 As shown, Figure 6 (a) shows the path planned before training. The white path is the planned path, and the black path is the expert demonstration path. Figure 6 (b) shows the path planned after training, starting from... Figure 6 It can be seen that after training, the consistency between the planned path and the expert demonstration path has been significantly improved.

[0055] S6-2-1: Random State Sampling. Within the feasible state space of a simple social scenario, a new potential growth point is selected through random sampling. This new potential growth point is denoted as... .

[0056] S6-2-2: Collision Feasibility Test. (Regarding...) Perform collision detection. If the detection fails, return to step S6-2-1 to execute the step of selecting a new potential growth point through random sampling; if the detection succeeds, then... As an effective growth point, let the effective growth point be denoted as... Execute S6-2-3.

[0057] S6-2-3: Time-series social cost assessment based on GRU. Calculation. Social adaptive navigation feature vector (i.e. ),Will The social adaptive navigation feature vector is used as the input at the current time (denoted as...). ) and input it into the GRU social feature encoder, which uses its internal gating mechanism to process the current input ( ) and hidden states that encode historical information of past path nodes ( The hidden state at the current moment is obtained by merging the two states. Construct a comprehensive social cost function and calculate the comprehensive social cost value based on the hidden state at the current moment.

[0058] In simple social scenarios, the comprehensive social cost for path node optimization should be a composite function that comprehensively measures the potential loss of navigation efficiency, the risk of behavioral stability, and the degree of conformity to social norms. The calculation of this function is based on GRU-encoded temporal social states. In this embodiment, the comprehensive social cost function is defined as: , In the formula, This represents the comprehensive social value at time t. This represents the hidden state at time t. This represents the comprehensive social cost function. It is a nonlinear fusion function. For efficiency cost components, characterization Within the implied temporal context, the potential increase in travel time that might result from selecting this node. For stability cost, characterization in The degree of abruptness in the robot's future navigation behavior relative to its recent historical behavior patterns. For compliance cost components, characterization in The severity of violating social norms (such as personal space) by choosing this node in the scenario it reflects.

[0059] The The calculation method is as follows: , In the formula, This is a baseline time increment related to the geometric distance between nodes. It is a nonlinear mapping; ( ) is a by The efficiency decay factor derived through a nonlinear mapping (such as a small neural network submodule). ( )≥1, ( The larger the value of ), the greater the negative impact of the current social state on traffic efficiency.

[0060] It can be defined as a weighted norm based on the change of the state vector, the The calculation method is as follows: , In the formula, It is a learnable diagonal weight matrix used to amplify fluctuations in key dimensions of social states. It is the recent historical hidden state (i.e.) , The moving average or smoothed representation of (, ...). Encourage behavioral smoothing;

[0061] The The calculation method is as follows: , In the formula, A learnable weight vector, used to learn from The "scene sensitivity" is decoded in the middle, where T is the total training time, exp is the exponential function, ensuring that the cost is positive and can amplify violations in highly sensitive scenes; Let d be the indicator function, and d be the robot's expected intrusion distance. For the comfort threshold, when hour ,when hour .

[0062] The nonlinear fusion function is specifically: , In the formula, , , To integrate the weighting coefficients, , To fuse the order parameters, >1, in this embodiment = 2. The core of the nonlinear fusion function is that it is not a simple linear weighting, but a nonlinear amplification of the high cost (i.e., performance bottleneck) in any single dimension, thereby forcing the planner to seek the optimal path that balances efficiency, stability and compliance.

[0063] In practical implementation, the function Φ( ) and its contained subfunctions ( ), , , , , , All of these can be achieved through a single [method / mechanism]. For input, with This is used to unify the implementation and end-to-end learning of the output backend neural network module (e.g., a multilayer perceptron MLP). During the training phase in simple scenarios, this module is optimized together with the pre-built GRU social feature encoder. The changes in its internal weights implicitly learn the evaluation rules and fusion logic of the costs of the aforementioned components. Therefore, the final evaluation formula... The technical concept and performance objectives of this invention are clearly expressed in mathematical form, while its parameterized implementation method (MLP) ensures the feasibility of the solution.

[0064] The design of the integrated social cost function enables the GRU-RRT* path planner to select nodes in simple social scenarios not only based on the instantaneous social characteristics of the node, but also adaptively considering the historical state evolution of the scenario reflected by a series of past nodes through the gated memory of GRU. This understanding of temporal context is the key to the model's ability to efficiently learn general and robust social rules in simple scenarios. These general rules, which are fixed in the GRU parameters (W,U,b) in simple social scenarios, will serve as high-quality initial prior knowledge, directly serving rapid adaptation and fine-tuning in complex social scenarios, demonstrating the core advantage of the hierarchical transfer paradigm of this invention.

[0065] S6-2-4: Parent Node Optimization and Path Rerouting Based on Temporal Social Cost. Based on the calculated... The GRU-RRT* path planner calculates the instantaneous comprehensive social value of the path from the starting point through potential parent nodes. The cumulative comprehensive social cost value; from the neighborhood node set of the random tree, select the node that minimizes the cumulative comprehensive social cost value as... Find the optimal parent node and establish a connection.

[0066] Subsequently, perform socially cost-driven rewiring optimization: check After joining, can the cumulative comprehensive social value of existing nodes in its neighborhood be increased by changing the parent node? This further reduces costs. This process ensures that the random trees generated in simple social scenarios not only approach geometric optimality but also achieve continuous optimization at the level of social behavioral costs, making the social characteristics of the learned paths more consistent and efficient.

[0067] S6-2-5: Judgment If the distance to the target point exceeds a threshold, return to step S6-2-1 to select a new potential growth point through random sampling and continue expanding; if it does not exceed the threshold, it is considered to have reached the destination and the path planning ends.

[0068] S6-3: The recursive generation module learns social adaptive navigation feature knowledge in simple social scenarios, and S6-3 is the core of the model learning process. The GRU-RRT* path planner compares its generated candidate paths with expert demonstration paths in the same scenario, calculating the difference in social feature distribution (e.g., loss function). This difference is used to update the internal weight parameters of the GRU social feature encoder through backpropagation. Based on the recursive generation module, the GRU-RRT* path planner will be forced to learn how to extract and generate high-level social navigation feature representations consistent with expert behavior from the original features through iterative training on a large number of basic scenarios, thereby mastering the general rules of social navigation. This general knowledge learned in simple social scenarios will serve as a strong prior, laying the foundation for subsequent knowledge transfer to complex social scenarios.

[0069] S6-4: Determine whether the training cycle has ended. If it has ended, the GRU-RRT* path planner training is complete, and proceed to S7. If it has not ended, update the next simple social scenario and return to S6-1 to execute the step of calculating the social adaptive navigation feature vector of each node on each expert demonstration path.

[0070] S6 is the first stage of training, which completes the initial training of the model on the basic scene set (i.e., simple social scenes) to achieve general knowledge learning.

[0071] S7: Social Adaptive Navigation Knowledge Transfer. The GRU-RRT* path planner trained in simple social scenarios will be used as the social navigation behavior meta-model. Prior knowledge will be introduced, and the same training method used in simple social scenarios will be used to train the social navigation behavior meta-model in complex social scenarios. The specific training process is S7-1 to S7-5.

[0072] After training in simple social scenarios, the resulting GRU-RRT* path planner with updated weights is embedded into a social navigation behavior meta-model. This meta-model encodes generalizable prior knowledge of social rules extracted from a large number of simple social scenarios. When faced with a library of complex social scenarios, firstly, the GRU parameters of this meta-model are loaded as the network initialization weights, rather than being randomly initialized. Subsequently, based on strong priors, the steps of feature extraction, path generation and optimization, and feature comparison learning are repeatedly performed only for the current complex social scenario. The GRU's gating mechanism enables it to quickly adapt to the unique dynamics of new scenarios while firmly retaining the basic social rules, achieving efficient and stable knowledge transfer and scenario-specific fine-tuning.

[0073] S7-1: Set all complex social scenarios in the complex social scenario library as a training loop and start training the social navigation behavior meta-model.

[0074] S7-2: Calculate the social adaptive navigation feature vector of each node on each expert demonstration path. The GRU social feature encoder encodes the social adaptive navigation feature vector to obtain the social cost weight corresponding to each node. The cost weights of all nodes are combined to generate a probabilistic social cost map that represents the social rules of the current complex social scenario.

[0075] S7-3: The GRU-RRT* path planner searches and generates r planned paths based on the social cost graph of complex social scenarios.

[0076] S7-4: The transfer learning of social adaptive navigation feature knowledge in complex social scenarios is completed through a recursive generation module.

[0077] The methods of S7-2 to S7-4 are the same as those of S6-1 to S6-3, the difference being the further increase in scene complexity and social interactivity in hierarchical data.

[0078] S7-5: Iteratively update the recursive generative model based on the gated loop unit. If the loss value converges or the number of training rounds is completed during model training, execute S8; otherwise, return to step S7-2.

[0079] The iterative update process of the social navigation behavior metamodel is as follows: S7-5-1: Training Data Pairing and Sequence Alignment. The social navigation behavior meta-model extracts the social adaptive navigation feature vectors of each node on the currently generated planned path to form the input sequence, denoted as [the input sequence is missing from the original text]. ,Will The input sequence serves as the meta-model of social navigation behavior; simultaneously, the comprehensive social cost value of corresponding nodes on the expert demonstration path is extracted to form the target value sequence, denoted as [the target value sequence is not specified in the original text]. ,Will As a meta-model of social navigation behavior, the target value sequence that needs to be fitted needs to be aligned. and Since the number of nodes on the planned path and the expert demonstration path are usually different, linear interpolation is used to resample the target value sequence to ensure that it is accurately aligned with the input sequence in the time dimension, forming an effective supervised learning sample pair, denoted as ( , ).

[0080] S7-5-2: Time-series cost prediction based on GRU. (The text then abruptly shifts to a seemingly unrelated topic: "aligned...") The social navigation behavior meta-model is input sequentially. The model processes the entire sequence according to a gating mechanism, outputting a sequence of predicted social costs for each node of the planned path, denoted as... The gating structure of GRU ensures the effective use of social context information in long path sequences, making predictions not only based on single-point features, but also incorporating the historical trends of the path.

[0081] S7-5-3: Constructing an adaptive weighting mechanism for predicting uncertainty and a loss function for training a meta-model of social navigation behavior by combining the predicted social cost value sequence.

[0082] To force the social cost evaluation criteria learned by the GRU-RRT* planner in simple scenarios to be better transferred to unseen scenarios with higher dynamics and complexity, this invention abandons simple mean squared error calculation and proposes a multi-stage loss function weighted by dynamic scenario sensitivity. By introducing an adaptive weighting mechanism based on prediction uncertainty, the training samples are differentiated, focusing on difficult samples that are more decisive in complex social scenarios (i.e., where expert and model judgments are prone to divergence), thereby improving the model's generalization ability and robustness in marginal cases. The loss function for training the social navigation behavior meta-model is: , In the formula, The loss function used when training the social navigation behavior meta-model, where n is the number of nodes. The dynamic adaptive weight of the i-th node. The comprehensive social value of the i-th node on the expert demonstration path (i.e. (the i-th target value in the process). The predicted social cost of the i-th node on the planned path (i.e. The predicted social cost in the i-th value. The basic cost prediction error is the core error term, used to ensure overall alignment between the planned path and the expert demonstration path. Through The GRU social feature encoder is forced to learn how to map raw social features into values ​​consistent with expert judgment.

[0083] The method for calculating the dynamic adaptive weight of the i-th node is as follows: , In the formula, It is an exponential function. The focus intensity coefficient is α > 0. This represents the average of the predicted social costs for the current batch. The standard deviation of the predicted social costs for the current batch. These are preset constants, small constants used to prevent division by zero. To predict the divergence focus weights, this term assigns the model a prediction cost that deviates from its current batch mean. Samples that are further away are given lower weights. In the early stages of training, model predictions often fall within an inaccurate range. At this point, samples where the model itself is "hesitant" (predictions deviate far from the batch mean, i.e., high uncertainty) are likely complex social interaction patterns that the model has not yet understood. Temporarily reducing the weight of these "difficult samples" in the early stages of training can prevent the model from prematurely fitting noise or getting trapped in local optima, and promote the model's stable learning of mainstream and clearly defined social rules. These are complexity weighting coefficients. The social adaptive navigation feature vector of the i-th node on the planned path (i.e. (the i-th social adaptive navigation feature vector in the data). for The variance; As a weight for scene complexity, this item is related to The variance is proportional to the square root. High variance in the feature vector usually indicates a highly dynamic or complex social scene at that node (e.g., multiple pedestrians approaching at different speeds and directions). This ensures the loss function pays more attention to samples in complex scenes, directly optimizing the model's evaluation performance in difficult but crucial scenarios, which aligns with the fundamental goal of "adaptive social navigation in complex scenes." α and λ are hyperparameters that control the intensity of "focused learning" and "complexity sensitivity," respectively. They can be fixed during training or dynamically adjusted according to the training stage (e.g., λ gradually increases as training progresses to enhance the fitting of complex scenes later).

[0084] S7-5-4: Gradient backpropagation. Using automatic differentiation, the gradient of the loss function with respect to all trainable parameters of the social navigation behavior meta-model (including weight matrices W and U, and bias vector b) is calculated. The GRU social feature encoding architecture ensures a more stable and efficient flow of gradients between time steps, effectively alleviating the gradient vanishing problem under long-term dependence. This is the key to its superiority over basic RNNs and its ability to perform deep temporal learning.

[0085] S7-5-5: Adaptive Parameter Optimization. The Adam optimizer is employed, using the gradients calculated in S7-5-4 to iteratively update the parameters of the social navigation behavior meta-model. The adaptive learning rate adjustment characteristic of the Adam optimizer, combined with the stable gradients of the GRU, enables the model to achieve fast and smooth convergence, whether in large-scale training in basic scenarios or in transfer fine-tuning in complex scenarios.

[0086] S7-1 to S7-5 constitute the second stage of training, achieving knowledge transfer and adaptation on a set of complex scenarios (i.e., complex social scenarios). In each complex social scenario, the feature extraction, path generation, and comparative learning processes are repeated through S7-2 to S7-4. At this point, based on strong priors, the model quickly adapts to the dense interaction patterns unique to complex scenarios, achieving targeted knowledge transfer and refinement.

[0087] S8: After training is completed, the parameters of the recursive generative model based on gated recurrent units are output to obtain the trained recursive generative model based on gated recurrent units, thus realizing transfer learning in complex social scenarios.

[0088] After training terminates, the final optimized GRU social feature encoder parameters are output. The resulting GRU-RRT* path planner has now completed hierarchical learning and transfer, enabling it to generate safe, efficient, and highly human-like navigation paths in complex, unseen human-machine collaborative scenarios by integrating real-time features and historical interaction context.

[0089] The paths planned by the GRU-RRT* path planner before and after migration in complex social scenarios are as follows: Figure 7 As shown, Figure 7 Image 1(a) shows the path planned before migrating to a complex social scenario. The white path is the planned path, and the black path is the expert demonstration path. Figure 7 1(b) is in Figure 7 The path planned after migrating from the complex social scenario in section 1(a); Figure 7 2(a) shows the path planned before migrating to another complex social scenario. Figure 7 2(b) is in Figure 7 The path planned after training in the complex social scenario transfer in section 2(a). From... Figure 7 It can be seen that the consistency between the planned path and the expert demonstration path has been significantly improved.

[0090] The final output parameters of the GRU social feature encoder, after the above hierarchical training and transfer optimization, give the GRU-RRT* path planner the following advanced features: 1. Possesses deep scene understanding and memory capabilities. GRU's gating memory mechanism enables the GRU-RRT* path planner to understand and remember long-term social interaction patterns during navigation (such as pedestrian following and avoidance loops), thereby generating temporally coherent and reasonable paths, exhibiting decision-making memory similar to that of humans.

[0091] 2. Possesses hierarchical generalization ability: It can not only reproduce expert behavior in trained scenarios, but also combine basic social rules learned in simple scenarios with advanced interaction strategies learned in complex scenarios, thereby generating safe, efficient and socially compliant anthropomorphic navigation paths in unseen and dynamically changing new human-computer interaction scenarios.

[0092] 3. Dynamic adaptability: Thanks to GRU's sensitivity to temporal dynamics, the GRU-RRT* path planner can respond quickly and smoothly to changes in pedestrian intent in the environment, adjust the path plan in real time, and avoid abrupt changes in behavior.

[0093] This invention also discloses a robot social adaptive navigation knowledge transfer system for complex scenarios, comprising: The scenario building module is used to build a multi-level progressive human-machine collaborative navigation scenario library, the levels of which include a simple social scenario library and a complex social scenario library; The dataset generation module is used to generate an expert demonstration path dataset for each level based on different levels in the human-machine collaborative navigation scenario library; The recursive generation model module is used to construct a recursive generation model based on gated loop units. The recursive generation model based on gated loop units includes a GRU social feature encoder constructed based on gated loop units. A hierarchical dataset consisting of expert demonstration path datasets corresponding to simple social scenarios is input into the recursive generation model based on gated loop units. The ablation training module is used to train the recursive generative model based on the gated loop unit and to determine the model parameters of the recursive generative model based on the gated loop unit through ablation experiments. The simple social scene training module is used to embed the GRU social feature encoder into the node cost evaluation module of the RRT* path planning algorithm to generate a GRU-RRT* path planner, and to set all simple social scenes in the simple social scene library as a training cycle and train the GRU-RRT* path planner. The complex social scenario knowledge transfer module is used to take the GRU-RRT* path planner trained in simple social scenarios as a social navigation behavior meta-model, introduce prior knowledge, and use the same training method as in simple social scenarios to train the social navigation behavior meta-model in complex social scenarios, resulting in a trained recursive generative model based on gated recurrent units.

[0094] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for knowledge transfer in social adaptive navigation of robots in complex scenarios.

[0095] The present invention also discloses a device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a knowledge transfer method for social adaptive navigation in complex scenarios for robots.

[0096] The core of this invention lies in using a temporal feature encoder constructed with gated recurrent units (GRUs) to replace traditional fully connected neural networks and transitive recurrent neural networks, and designing a hierarchical transfer learning paradigm from simple to complex scenarios. The constructed GRU-RRT* path planner, through the GRU's gated memory mechanism, can more effectively learn navigation behavior knowledge in human-machine collaborative scenarios. It can not only handle the multi-dimensional social adaptive features of the current node (including pedestrian cost, path distance cost, obstacle distance cost, and pedestrian interaction convergence probability prediction), but also robustly integrate the state information of historical path nodes, thereby endowing the planner with the ability to continuously remember and understand the dynamic evolution of the scene.

[0097] When faced with new scenarios, this invention achieves efficient adaptation through a hierarchical knowledge transfer process. First, the model is trained in numerous simple, rule-based scenarios to master general navigation rules, and the fully trained GRU-RRT* planner is then solidified as a navigation behavior meta-model. Subsequently, when facing complex, highly dynamic, and challenging scenarios, the GRU parameters of this meta-model are directly loaded as strong priors, and rapid fine-tuning is performed on this basis, thereby completing the targeted transfer and enhancement of social adaptive navigation knowledge from simple to complex domains. Ultimately, this enables the planner to generate human-like navigation paths with high safety, high efficiency, and high social acceptance in complex human-computer interaction scenarios, comprehensively considering real-time features and historical information, significantly improving the generalization ability and practicality of the planning system in unknown dynamic environments.

[0098] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will 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 program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for knowledge transfer in social adaptive navigation of robots in complex scenarios, characterized in that, include: A multi-level, progressive human-machine collaborative navigation scenario library is constructed, wherein the levels of the human-machine collaborative navigation scenario library include a simple social scenario library and a complex social scenario library; Based on the different levels in the human-machine collaborative navigation scenario library, an expert demonstration path dataset is generated for each level. A recursive generation model based on gated loop units is constructed. The recursive generation model based on gated loop units includes a GRU social feature encoder constructed based on gated loop units. A hierarchical dataset consisting of expert demonstration path datasets corresponding to simple social scenarios is input into the recursive generation model based on gated loop units. The recursive generative model based on gated loop units is trained, and the model parameters of the recursive generative model based on gated loop units are determined through ablation experiments. The GRU social feature encoder is embedded into the node cost evaluation module of the RRT* path planning algorithm to generate a GRU-RRT* path planner. All simple social scenarios in the simple social scenario library are set as a training loop and the GRU-RRT* path planner is trained. The GRU-RRT* path planner trained in simple social scenarios is used as the social navigation behavior meta-model. Prior knowledge is introduced, and the same training method used in simple social scenarios is used to train the social navigation behavior meta-model in complex social scenarios, resulting in a trained recursive generative model based on gated recurrent units.

2. The robot complex scene social adaptive navigation knowledge transfer method according to claim 1, characterized in that: Training the GRU-RRT* path planner includes: Calculate the social adaptive navigation feature vector of each node on each expert demonstration path to generate a social cost graph representing the social rules of the current simple social scenario; Based on the social cost graph of social rules in simple social scenarios, search and generate planned paths; The recursive generation module completes the learning of social adaptive navigation feature knowledge in simple social scenarios. It determines whether the training cycle has ended. If it has ended, the GRU-RRT* path planner training is complete. If it has not ended, it updates the next simple social scenario and returns to the step of calculating the social adaptive navigation feature vector of each node on each expert demonstration path.

3. The robot complex scene social adaptive navigation knowledge transfer method according to claim 1, characterized in that: When determining the model parameters of the recursive generation model based on gated recurrent units through ablation experiments, the model parameters are selected by constructing a comprehensive efficiency score, which is: ， In the formula, To score overall efficiency, To score performance, This is the performance rating coefficient. Scoring is given for convergence speed. This is the convergence speed scoring coefficient. Score the model complexity. This is the model complexity scoring coefficient. Rate time efficiency. This is a time efficiency rating coefficient. This is the Dropout penalty factor.

4. The robot complex scene social adaptive navigation knowledge transfer method according to claim 2, characterized in that: The social cost graph based on social rules in simple social scenarios is searched and a planned path is generated, including: Within the feasible state space of a simple social scenario, a new potential growth point is selected through random sampling. This new potential growth point is denoted as... ; right Perform collision detection. If the detection fails, return to the step of selecting a new potential growth point through random sampling; if the detection succeeds, then... As an effective growth point, let the effective growth point be denoted as... ; calculate The social adaptive navigation feature vector is input into the GRU social feature encoder. The GRU social feature encoder fuses the current input with the hidden state that encodes the historical information of past path nodes through a gating mechanism to obtain the hidden state at the current moment. A comprehensive social cost function is constructed, and the comprehensive social cost value is calculated based on the hidden state at the current moment. Based on the calculation The instantaneous comprehensive social value is calculated from the starting point through potential parent nodes. The cumulative comprehensive social cost value; from the neighborhood node set of the random tree, select the node that minimizes the cumulative comprehensive social cost value as... Find the optimal parent node and establish a connection; judge If the distance to the target point exceeds a threshold, return to the step of selecting a new potential growth point by random sampling; if it does not exceed the threshold, it is considered that the destination has been reached and the path planning ends.

5. The robot complex scene social adaptive navigation knowledge transfer method according to claim 4, characterized in that: The comprehensive social cost function is: ， In the formula, This represents the comprehensive social value at time t. This represents the hidden state at time t. It is a nonlinear fusion function. For efficiency cost components, characterization Within the implied temporal context, For stability cost, characterization in The degree of abruptness in the robot's future navigation behavior relative to its recent historical behavior patterns. For compliance cost components, characterization in The severity of violating social norms by choosing this node in the scenario it reflects.

6. The robot complex scene social adaptive navigation knowledge transfer method according to claim 5, characterized in that: The The calculation method is as follows: ， In the formula, The reference time increment is related to the geometric distance between nodes. It is a nonlinear mapping; The The calculation method is as follows: ， In the formula, The diagonal weight matrix is ​​learnable. It is a moving average or smoothed representation of the recent historical hidden state; The The calculation method is as follows: ， In the formula, Let T be the learnable weight vector, T be the total training time, and exp be the exponential function. Let d be the indicator function, and d be the robot's expected intrusion distance. For the comfort threshold, when hour ,when hour .

7. The robot complex scene social adaptive navigation knowledge transfer method according to claim 5, characterized in that: The nonlinear fusion function is specifically: ， In the formula, , , To integrate the weighting coefficients, , To fuse the order parameters, >

1.

8. The robot complex scene social adaptive navigation knowledge transfer method according to claim 4, characterized in that: The iterative update process of the social navigation behavior meta-model during training is as follows: The social navigation behavior metamodel extracts the social adaptive navigation feature vectors of each node on the currently generated planned path to form an input sequence, denoted as . ,Will The input sequence serves as the input sequence for the social navigation behavior meta-model; the comprehensive social cost value of the corresponding nodes on the expert demonstration path is extracted to form the target value sequence, denoted as . ,Will As a meta-model of social navigation behavior, the target value sequence that needs to be fitted needs to be aligned. and ; Aligned Input the social navigation behavior meta-model sequentially, and output a sequence of predicted social costs for each node of the planned path; An adaptive weighting mechanism for predicting uncertainty is constructed, and a loss function is built by combining the predicted social cost value sequence to train the meta-model of social navigation behavior. The gradient of the loss function with respect to all trainable parameters of the social navigation behavior meta-model is calculated using automatic differentiation techniques. The Adam optimizer is used to iteratively update the parameters of the social navigation behavior meta-model using the gradient.

9. The robot complex scene social adaptive navigation knowledge transfer method according to claim 8, characterized in that: The loss function for training the social navigation behavior meta-model is: ， In the formula, The loss function used when training the social navigation behavior meta-model, where n is the number of nodes. The dynamic adaptive weight of the i-th node. To demonstrate the comprehensive social value of the i-th node on the expert demonstration path. Let be the predicted social cost of the i-th node on the planned path; The method for calculating the dynamic adaptive weight of the i-th node is as follows: ， In the formula, It is an exponential function. To focus the intensity coefficient, This represents the average of the predicted social costs for the current batch. The standard deviation of the predicted social costs for the current batch. As a preset constant, These are complexity weighting coefficients. Let i be the social adaptive navigation feature vector of the i-th node on the planning path. for The variance.

10. A robot complex scene social adaptive navigation knowledge transfer system, characterized in that, include: The scenario building module is used to build a multi-level progressive human-machine collaborative navigation scenario library, the levels of which include a simple social scenario library and a complex social scenario library; The dataset generation module is used to generate an expert demonstration path dataset for each level based on different levels in the human-machine collaborative navigation scenario library; The recursive generation model module is used to construct a recursive generation model based on gated loop units. The recursive generation model based on gated loop units includes a GRU social feature encoder constructed based on gated loop units. A hierarchical dataset consisting of expert demonstration path datasets corresponding to simple social scenarios is input into the recursive generation model based on gated loop units. The ablation training module is used to train the recursive generative model based on the gated loop unit and to determine the model parameters of the recursive generative model based on the gated loop unit through ablation experiments. The simple social scene training module is used to embed the GRU social feature encoder into the node cost evaluation module of the RRT* path planning algorithm to generate a GRU-RRT* path planner, and to set all simple social scenes in the simple social scene library as a training cycle and train the GRU-RRT* path planner. The complex social scenario knowledge transfer module is used to take the GRU-RRT* path planner trained in simple social scenarios as a social navigation behavior meta-model, introduce prior knowledge, and use the same training method as in simple social scenarios to train the social navigation behavior meta-model in complex social scenarios, resulting in a trained recursive generative model based on gated recurrent units.

Images