A social navigation method, device and medium for human intention perception understanding

Abstract

The application discloses a social navigation method and device for human intention perception understanding and a medium, wherein the method comprises the following steps: automatically constructing a training set of human action navigation through a large language model and a text action generation model; embedding an intention predictor module in a basic navigation model to construct an intelligent agent navigation model with intention understanding capability; constructing a reinforcement learning reward function and an auxiliary loss function, and training the intelligent agent navigation model through reinforcement learning. The application proposes a human action automatic construction framework, generates diverse and real human actions by combining a large language model and a text action generation model, does not need to rely on manual labeling, effectively improves the generation efficiency of the cooperative navigation data set, and reduces the cost. Meanwhile, the application embeds a human action perception module in the basic navigation module, and gives the intelligent agent the ability to understand human action intentions. The application can be widely applied to the field of robot navigation.

Description

Technical Field

[0001] This invention relates to the field of robot navigation, and more particularly to a social navigation method, device, and medium oriented towards human intention perception and understanding. Background Technology

[0002] Embodied intelligence refers to intelligent agents achieving intelligent behavior through dynamic interaction with the physical environment, combining perception, movement, and cognition. Unlike traditional artificial intelligence, embodied intelligence emphasizes the presence and actions of intelligent agents in the physical world. It relies not only on information processing and decision-making but also on direct contact and feedback with the environment. Research on embodied intelligence can significantly improve the performance of robots in various application scenarios, enabling them to play a greater role in fields such as healthcare, disaster relief, industrial manufacturing, and home services. Among these, social navigation, which requires intelligent agents to collaborate with humans or other agents to move effectively and safely in shared physical environments and achieve goals, has gradually gained widespread attention in recent years and has become one of the research hotspots in embodied intelligence. Because intelligent agents need to understand and predict dynamic human behavior and intentions to perform real-time interaction and decision-making, social navigation is a more challenging task compared to object navigation and language navigation.

[0003] Current social navigation research focuses on minimizing human-robot collisions in point-to-point navigation tasks. Some methods construct human-centric navigation datasets in 2D space; however, since existing simulators lack 3D human animation, the agent cannot observe human actions in the scene to understand human movement intentions. Other methods enable humanoid agents to move in 3D simulation environments; however, the humanoid agent's movement is controlled by a simple collision avoidance dynamic model, resulting in aimless wandering. Therefore, the agent cannot be trained on these datasets to represent and model realistic human behavior and reactions. In contrast, while other works provide 3D humanoid agent actions, the high cost of animateting different human actions results in small datasets, and the lack of a natural connection between humanoid agent actions and their trajectories makes it difficult to generalize the agent's social navigation capabilities to unknown scenarios. Furthermore, aside from maintaining distance from humans, none of these methods consider how to train the agent to understand human movement intentions to complete social navigation tasks. Summary of the Invention

[0004] In order to at least partially solve one of the technical problems existing in the prior art, the present invention aims to provide a social navigation method, device and medium oriented towards human intention perception and understanding.

[0005] The first technical solution adopted in this invention is:

[0006] A social navigation method based on human intention perception and understanding includes the following steps:

[0007] Automatically construct a training set for human action navigation using large language models and text action generation models;

[0008] An intent predictor module is embedded in the basic navigation model to construct an intelligent agent navigation model with intent understanding capabilities;

[0009] Construct a reinforcement learning reward function and an auxiliary loss function, and train the agent navigation model through reinforcement learning.

[0010] Furthermore, the automatic construction of a training set for human action navigation through a large language model and a text action generation model includes:

[0011] The text description generated by the action is obtained through a pre-defined large language model;

[0012] Based on the text description, humanoid intelligent agent action fragments are automatically generated using a preset text action generation model;

[0013] The system obtains the walking trajectory of a humanoid intelligent agent, generates walking actions based on the trajectory, splices action fragments and walking actions to obtain complete human actions, and constructs a training set.

[0014] Furthermore, the step of obtaining the text description generated by the action through a preset large language model includes:

[0015] Obtain the object categories and coordinates that can interact with the humanoid intelligent agent in the 3D simulation scene, and construct an object set O. D Set O D These candidates are provided to the large language model to ensure that the generated activities can be executed in the scene;

[0016] The large language model generates templates based on preset actions to ensure that the humanoid agent describing the text performs meaningful actions in two different places, so as to avoid aimless wandering.

[0017] Furthermore, the text action generation model is an MDM model;

[0018] The text action generation model works as follows:

[0019] Human actions are represented as in This represents a single pose, where N is the length of the action sequence;

[0020] The MDM model is a denoising diffusion model that assumes the T-step noise steps modeled by a stochastic process are as follows:

[0021]

[0022] In the formula, t∈T represents the value t being randomly selected from T steps at a given time, and X t ~N(0, I) represents X t Sampling from a standard normal distribution, α t It is a cumulative scaling factor that represents the degree of change of the data from the initial state to the state at time t, where I is the identity matrix;

[0023] MDM model in a given noise action X t Predict the denoised action under the conditions of time step t and text description c encoded by CLIP. At each time step t, a denoised action Predicted, then noisy to X t-1 The process is repeated from t=T until X0 is obtained after denoising; therefore, the MDM model can generate human actions that match the description based on the noisy actions and the text description.

[0024] Furthermore, the process of acquiring the walking trajectory of the humanoid intelligent agent, generating walking actions based on the trajectory, and splicing action fragments and walking actions to obtain complete human actions includes:

[0025] Based on the coordinates of object A and object B and the navigable map of the scene, the collision-free trajectory of the humanoid intelligent agent walking in the scene is calculated using the A* algorithm.

[0026] Based on the obtained collision-free trajectory and the text description of walking, the walking action is generated using the PriorMDM model;

[0027] The generated motion fragments and walking motions are stitched together using linear interpolation to form complete human motion. This is then integrated into the 3D simulation environment based on coordinates, ensuring that humans can perform meaningful actions and walk purposefully in the 3D environment.

[0028] Furthermore, the basic navigation model includes a mapping network and a policy network;

[0029] The mapping network is used to map observations in the environment to a feature space; for image observations, a convolutional neural network is used for feature extraction, and for category observations, a multi-layer fully connected layer is used for feature extraction.

[0030] The policy network is implemented using an LSTM network, which is used to predict the action taken by the agent at the current time based on the current time mapping network output features, the action at the previous time, and the LSTM hidden layer features.

[0031] The input to the intent predictor module is human actions observed from the environment, including a third-person perspective image I of the human actions. t Objects that are being interacted with by human actions Human skeleton P based on human movement prediction t and the historical trajectory of human movement T t-5～t ;

[0032] To model the intrinsic relationship between human observation and human intent, a transformer architecture is constructed to transform intent prediction into a sequence-to-sequence transformation problem. Subsequently, three additional tags are embedded after the extracted features: CLS... obj CLS int and REG tra To enhance information and make predictions;

[0033] CLS after the transformer network obj CLS int and REG tra The intentional object is predicted using a linear network. Human activity category Y act And the future trajectory of humanity T t+1～t+K Finally, the output of the intent predictor module and the agent's observations are input into the mapping network.

[0034] Furthermore, the construction of the auxiliary loss function includes:

[0035] During training, real data corresponding to the predicted action intent is obtained from the environment, and auxiliary loss is calculated to supervise learning. The formula for auxiliary loss is as follows:

[0036]

[0037]

[0038]

[0039]

[0040] In the formula, P obj P tct , These are the predicted values ​​for the intended object, the activity category, and the future trajectory, respectively. Y act， T t～t+K These are the actual values ​​for the intended object, the activity category, and the future trajectory, respectively.

[0041] The supervision loss for the intentional object and activity categories is calculated using cross-entropy loss, while the supervision loss for the future trajectory is calculated using mean squared error loss. The final auxiliary loss is a weighted sum of the three types of losses and is directly added to the loss of the PPO algorithm for optimization along with reinforcement learning.

[0042] Furthermore, the construction of the reinforcement learning reward function includes:

[0043] During training, the agent is encouraged to reach its destination efficiently, and in accordance with human-centered navigation requirements, the agent should not collide with humans during navigation. Therefore, the reward function is designed as follows:

[0044]

[0045]

[0046] In the formula, R t The reward at time t. A 1 indicates task completion, and a 0 indicates task failure. The value of 1 indicates a collision between the agent and a human. The value is 1 for all cases and 0 for all others. The value is calculated by subtracting the current distance from the agent to the finish line from the previous distance. This indicates that the agent receives a positive reward for moving closer to the finish line, and a negative reward for moving closer. Δ(b) t h t ) represents the distance between the agent and the target point at the current moment.

[0047] The second technical solution adopted in this invention is:

[0048] An electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by the processor to implement a social navigation method oriented towards human intention perception and understanding as described above.

[0049] The third technical solution adopted in this invention is:

[0050] A computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement a social navigation method oriented towards human intention perception and understanding as described above.

[0051] The fourth technical solution adopted in this invention is:

[0052] A computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned social navigation method based on human intention perception and understanding.

[0053] The beneficial effects of this invention are as follows: This invention proposes an automatic human action construction framework, which generates diverse and realistic human actions by combining a large language model and a text action generation model, without relying on manual annotation, effectively improving the generation efficiency of collaborative navigation datasets and reducing costs. Simultaneously, this invention embeds a human action perception module into the basic navigation module, endowing the intelligent agent with the ability to understand human action intentions. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating the steps of a social navigation method for understanding human intentions in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of an intelligent agent navigation model with intent understanding capability in an embodiment of the present invention. Detailed Implementation

[0057] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0058] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0059] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0060] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0061] Terminology Explanation:

[0062] MDM model: Text Action Generation Diffusion Model, is a model used to generate 3D human actions based on text descriptions.

[0063] PriorMDM model: Action generation diffusion model, is an improved version of the MDM model, which can generate 3D human actions based on text descriptions or motion trajectories.

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment provides a social navigation method based on human intention perception and understanding, specifically including the following steps:

[0066] S1. Automatically construct a training set for human action navigation using a large language model and a text action generation model.

[0067] As an optional implementation, step S1 specifically includes the following steps:

[0068] S11. Obtain the text description generated by the action through a preset large language model.

[0069] Specifically, firstly, the object categories and coordinates that can interact with the humanoid intelligent agent in the 3D simulation scene are obtained to construct an object set O. D , will O D The candidate actions are provided to the large language model to ensure that the generated actions can be performed in the scene. The large language model ensures that the humanoid agent described in the text performs meaningful actions in two different places based on the action generation template "do [action A] at {object A}, then walk to {object B} and do [action B]", avoiding aimless wandering.

[0070] S12. Based on the text description, automatically generate humanoid intelligent agent action fragments using a preset text action generation model.

[0071] In this embodiment, the state-of-the-art text action generation model, MDM, is used, which can generate realistic and diverse human actions based on text descriptions. Human actions are represented as... in This represents a single pose. MDM is a denoising diffusion model that assumes a T-step noise step modeled by a stochastic process as follows:

[0072]

[0073] Where t∈T, X T ~N(0, I). MDM modeling and denoising process, which, given noise action X t Predict the denoised action under the conditions of time step t and text description c encoded by CLIP. At each time step t, a denoised action Predicted, then noisy to X t-1 The process repeats from t=T until X0 is obtained after denoising. Therefore, it can generate human-like actions that match the description based on the noisy actions and the text description.

[0074] S13. Obtain the walking trajectory of the humanoid intelligent agent, generate walking actions based on the trajectory, splice action fragments and walking actions to obtain complete human actions, and construct a training set.

[0075] Based on the coordinates of objects A and B and the navigable map of the scene, the collision-free trajectory of a humanoid agent walking in the scene can be calculated using the A* algorithm. A* is a common path planning algorithm and will not be explained in detail. Based on the generated trajectory and the text description of the walking motion, we use the PriorMDM model to generate the walking action. PriorMDM is an improved version of the MDM model, which allows control over the model to generate actions that conform to predetermined trajectory features. In this patent, the model is controlled to generate actions that conform to predetermined trajectory features. The principle of PriorMDM is that during the MDM denoising process, the predicted action is calculated for each step. The trajectory parameters are replaced with the trajectory calculated by A*. Therefore, the final output X0 will be an action that conforms to the specified trajectory characteristics. Finally, the action segments generated by S12 and S13 are stitched together using linear interpolation to form a complete human activity. This is then integrated into the 3D simulation environment based on coordinates, ensuring that humans can perform meaningful actions and purposeful walking in the 3D environment.

[0076] S2. Embed an intent predictor module into the basic navigation model to build an intelligent agent navigation model with intent understanding capabilities.

[0077] In some embodiments, see Figure 2The basic navigation model consists of two parts: a mapping network and a policy network. The mapping network maps observations in the environment to a feature space. For image observations, a convolutional neural network is used for feature extraction, while for category observations, multiple fully connected layers are used for feature extraction. The policy network uses an LSTM network to predict the agent's action at the current moment based on the output features of the mapping network, the action from the previous moment, and the features from the LSTM hidden layers. The main action categories are forward, backward, left turn, right turn, and stop. This basic module endows the agent with the ability to navigate and explore in an indoor environment.

[0078] To understand human intentions and enable navigation, this embodiment embeds an intention predictor module into the basic module. The input to the intention predictor module is human movement observed from the environment, primarily including a third-person perspective image (I) of human actions. t Objects that are being interacted with by human actions Human skeleton P based on human movement prediction t and the historical trajectory of human movement T t-5～t Extracting I through a ResNet convolutional network t For the first three categories of observations, features are extracted using a linear network. To model the intrinsic relationship between human observations and human intent, we further construct a transformer architecture that transforms intent prediction into a sequence-to-sequence transformation problem. Subsequently, three additional tags are embedded after the extracted features: CLS... obj CLS int and REG tra This enhances information and facilitates prediction. CLS after passing through the transformer network... obj CLS int and REG tra The intended object is predicted using a linear network. Human activity category Y act And the future trajectory of humanity T t+1～t+K Finally, the output of the intent predictor module, along with the agent's observations, is input into the mapping network. By embedding the intent predictor, the agent achieves human-like intent perception and understanding while possessing basic navigation capabilities, enabling it to complete tasks in complex human motion environments.

[0079] S3. Construct a reinforcement learning reward function and an auxiliary loss function, and train the agent navigation model through reinforcement learning.

[0080] As an optional implementation, step S3 specifically includes the following steps:

[0081] S31. Construct the auxiliary loss function.

[0082] During training, we obtain real-world data corresponding to the predicted action intent from the environment and directly calculate the auxiliary loss for supervised learning. The auxiliary loss formula is as follows:

[0083]

[0084]

[0085]

[0086]

[0087] Among them, P obj P act , These are the predicted values ​​for the intended object, the activity category, and the future trajectory, respectively. Y act T t～t+K These represent the ground truth values ​​for the intended object, activity category, and future trajectory, respectively. The supervision loss for the intended object and activity category is calculated using cross-entropy loss, while the supervision loss for the future trajectory is calculated using mean squared error loss. The final auxiliary loss is a weighted sum of the three losses and is directly added to the loss of the PPO algorithm for optimization during reinforcement learning.

[0088] S32. Reinforcement learning reward function design.

[0089] During training, the agent is encouraged to reach its destination efficiently, and in accordance with human-centered navigation requirements, the agent should not collide with humans during navigation. Therefore, we design the reward function as follows:

[0090]

[0091]

[0092] Successful navigation earns a reward of +10, while a collision results in a -1 reward. The value is calculated by subtracting the current distance from the agent to the finish line from the previous distance. This indicates that the agent receives a positive reward for moving closer to the finish line, and a negative reward for moving otherwise. Therefore, the agent is encouraged to move towards the destination. R t The final element is a time penalty to encourage the agent to reach its destination as quickly as possible. Through this reinforcement learning design, the agent is trained using PPO reinforcement learning on a constructed simulation dataset, thus learning intention-aware social navigation capabilities.

[0093] In summary, existing social navigation simulation datasets only construct human movement trajectories or generate a limited number of human actions manually. Therefore, current social navigation simulations are not close to reality, and dataset generation faces bottlenecks in terms of cost and efficiency. Furthermore, because existing social navigation datasets only contain human trajectories or a limited number of human actions, the designed methods lack an intent understanding module, resulting in the agent lacking intent understanding capabilities. This invention proposes an automatic human action construction framework that combines a large language model and a text action generation model to generate diverse and realistic human actions without relying on manual annotation, effectively improving the generation efficiency of collaborative navigation datasets and reducing costs. Simultaneously, this invention embeds a human action perception module into the basic navigation module, endowing the agent with the ability to understand human action intentions.

[0094] In summary, the method of the present invention has at least the following advantages and beneficial effects compared with the prior art:

[0095] (1) Traditional methods usually rely on manually creating human actions in social navigation datasets, resulting in small dataset size, high construction cost, and low generation efficiency. However, the method of this invention can automatically generate a large number of diverse and realistic human actions using only existing deep learning methods (large language models and text action generation models), thus having significant advantages in terms of dataset construction scale, construction cost, and construction efficiency.

[0096] (2) Existing methods either fail to incorporate human action intent due to the scarcity of human actions in the dataset, or suffer from weak generalization ability in social navigation due to the lack of datasets. The method in this paper integrates the human action intent module into the traditional navigation module and uses a large and diverse dataset of real human actions for training. Therefore, the agent has significant advantages and generalization ability in intention-understanding navigation.

[0097] Example 2

[0098] This invention also provides an electronic device, which includes a processor and a memory. The memory stores at least one instruction, at least one program, a code set, or an instruction set. The at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to achieve the following: Figure 1 This illustrates a social navigation method oriented towards understanding human intentions.

[0099] It is understood that the memory may include random access memory (RAM) or read-only memory. Optionally, the memory may include non-transitory computer-readable storage medium. The memory can be used to store instructions, programs, code, code sets, or instruction sets. The memory may include a stored program area and a stored data area, wherein the stored program area may store instructions for implementing an operating system, instructions for at least one function, instructions for implementing the various method embodiments described above, etc.; the stored data area may store data created according to the use of the server, etc.

[0100] A processor may include one or more processing cores. The processor connects to various parts of the server via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in memory, and accessing data stored in memory to perform various server functions and process data. Optionally, the processor may be implemented using at least one of the following hardware forms: Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor may integrate one or more of the following: Central Processing Unit (CPU) and Modem. The CPU primarily handles the operating system and applications; the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor.

[0101] Since this electronic device is an electronic device corresponding to a social navigation method for human intention perception and understanding in an embodiment of the present invention, and the principle of solving the problem by this electronic device is similar to that of this method, the implementation of this electronic device can refer to the implementation process of the above method embodiment, and repeated parts will not be described again.

[0102] Example 3

[0103] This invention also provides a computer-readable storage medium storing at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to achieve the following: Figure 1 This illustrates a social navigation method oriented towards understanding human intentions.

[0104] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0105] Since this storage medium is the storage medium corresponding to a social navigation method for human intention perception and understanding in an embodiment of the present invention, and the principle of the storage medium in solving the problem is similar to that of the method, the implementation of this storage medium can refer to the implementation process of the above method embodiment, and the repeated parts will not be described again.

[0106] Example 4

[0107] In some possible implementations, various aspects of the methods of the embodiments of the present invention can also be implemented as a program product comprising program code that, when run on a computer device, causes the computer device to perform the steps of a social navigation method oriented towards human intention perception and understanding according to various exemplary embodiments of this application as described above. The executable computer program code or "code" for performing the various embodiments can be written in high-level programming languages ​​such as C, C++, C#, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0108] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0109] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0110] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A social navigation method oriented towards human intention perception and understanding, characterized in that, Includes the following steps: Automatically construct a training set for human action navigation using large language models and text action generation models; An intent predictor module is embedded in the basic navigation model to construct an intelligent agent navigation model with intent understanding capabilities; Construct a reinforcement learning reward function and an auxiliary loss function, and train the agent navigation model through reinforcement learning; The method of automatically constructing a training set for human action navigation through a large language model and a text action generation model includes: The text description generated by the action is obtained through a pre-defined large language model; Based on the text description, humanoid intelligent agent action fragments are automatically generated using a preset text action generation model; The system acquires the walking trajectory of a humanoid intelligent agent, generates walking actions based on the trajectory, splices action fragments and walking actions to obtain complete human actions, and constructs a training set. The basic navigation model includes a mapping network and a policy network; The mapping network is used to map observations in the environment to a feature space; for image observations, a convolutional neural network is used for feature extraction, and for category observations, a multi-layer fully connected layer is used for feature extraction. The policy network is implemented using an LSTM network, which is used to predict the action taken by the agent at the current time based on the current time mapping network output features, the action at the previous time, and the LSTM hidden layer features. The input to the intent predictor module is human actions observed from the environment, including third-person perspective images of the human actions. Objects that are being interacted with by human actions Human skeleton predicted based on human movement and the trajectory of human movement history ; To model the intrinsic relationship between human observation and human intent, a transformer architecture is constructed to transform intent prediction into a sequence-to-sequence transformation problem; subsequently, three additional markers are embedded after the extracted features, namely... , and To enhance information and make predictions; After passing through the transformer network 、 and The intentional object is predicted using a linear network. Categories of human activities and the future trajectory of humanity Finally, the output of the intent predictor module and the agent's observations are input into the mapping network.

2. The social navigation method based on human intention perception and understanding according to claim 1, characterized in that, The text description generated by obtaining the action through a preset large language model includes: Obtain the object categories and coordinates in the 3D simulation scene that can interact with the humanoid intelligent agent, and construct an object set. , will set These candidates are provided to the large language model to ensure that the generated activities can be executed in the scene; The large language model generates templates based on preset actions to ensure that the humanoid agent describing the text performs meaningful actions in two different places.

3. The social navigation method based on human intention perception and understanding according to claim 1, characterized in that, The text action generation model is an MDM model; The text action generation model works as follows: Human actions are represented as ,in Indicates a single gesture. The length of the action sequence; The MDM model is a denoising diffusion model that assumes modeling by a stochastic process. The noise step is as follows: In the formula, It is a cumulative scaling factor that represents the data's scaling from its initial state to time. The degree of change in state, It is the identity matrix; MDM model in the presence of noise action Time step and text descriptions encoded via CLIP Predicting denoised actions under the given conditions ; at each time step A noise-reduced action Predicted, then noise is added. ,from Start repeating until the noise is removed. Therefore, the MDM model can generate human actions that match the description based on noisy actions and text descriptions.

4. A social navigation method based on human intention perception and understanding according to claim 1, characterized in that, The process of acquiring the walking trajectory of the humanoid intelligent agent, generating walking actions based on the trajectory, and splicing action fragments and walking actions to obtain complete human actions includes: Based on the coordinates of object A and object B and the navigable map of the scene, the collision-free trajectory of the humanoid intelligent agent walking in the scene is calculated using the A* algorithm. Based on the obtained collision-free trajectory and the text description of walking, the walking action is generated using the PriorMDM model; The generated motion fragments and walking motions are stitched together using linear interpolation to form complete human motion. This is then integrated into the 3D simulation environment based on coordinates, ensuring that humans can perform meaningful actions and walk purposefully in the 3D environment.

5. A social navigation method based on human intention perception and understanding according to claim 1, characterized in that, The construction of the auxiliary loss function includes: During training, real data corresponding to the predicted action intent is obtained from the environment, and auxiliary loss is calculated to supervise learning. The formula for auxiliary loss is as follows: In the formula, , , These are the predicted values ​​for the intended object, the activity category, and the future trajectory, respectively. , , These are the actual values ​​for the intended object, the activity category, and the future trajectory, respectively. The supervision loss for the intentional object and activity categories is calculated using cross-entropy loss, while the supervision loss for the future trajectory is calculated using mean squared error loss. The final auxiliary loss is a weighted sum of the three types of losses and is directly added to the loss of the PPO algorithm for optimization along with reinforcement learning.

6. A social navigation method based on human intention perception and understanding according to claim 1, characterized in that, The construction of the reinforcement learning reward function includes: During training, the agent is encouraged to reach its destination efficiently, and in accordance with human-centered navigation requirements, the agent should not collide with humans during navigation. Therefore, the reward function is designed as follows: In the formula, For time Momentary rewards A 1 indicates task completion, and a 0 indicates task failure. The value of 1 indicates a collision between the agent and a human. The value is 1 for all cases and 0 for all others. The value is the difference between the distance from the previous agent to the finish line and the current agent's distance to the finish line. This indicates that the agent receives a positive reward if it gets closer to the finish line, and a negative reward otherwise. This indicates the distance between the agent and the target point at the current moment.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the method as described in any one of claims 1 to 6.

Images