Social navigation method and equipment oriented to human intention perception and understanding and medium

Abstract

The invention discloses a social navigation method and device oriented to human intention perception and understanding and a medium, and 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 into the basic navigation model to construct an agent navigation model with intention understanding ability; and constructing a reinforcement learning reward function and an auxiliary loss function, and training the intelligent agent navigation model through reinforcement learning. According to the method, a human action automatic construction framework is provided, various real human actions are generated by combining a large language model and a text action generation model, manual labeling is not needed, the generation efficiency of a collaborative navigation data set is effectively improved, and the cost is reduced. Meanwhile, the human action sensing module is embedded in the basic navigation module, so that the intelligent agent is endowed with the ability of understanding the human action intention. The method can be widely applied to the field of robot navigation.

Description

Technical Field

[0001] The present invention relates to the field of robot navigation, and in particular to a social navigation method, device and medium for perceiving and understanding human intentions. Background Art

[0002] Embodied intelligence refers to the intelligent behavior of intelligent agents through dynamic interaction with the physical environment, combining perception, movement and cognitive abilities. Unlike traditional artificial intelligence, embodied intelligence emphasizes the existence 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. The research on embodied intelligence can greatly improve the performance of robots in various application scenarios, enabling them to play a greater role in medical care, disaster relief, industrial manufacturing, home services and other fields. Among them, social navigation requires intelligent agents to collaborate with humans or other agents to effectively and safely move and complete goals in a shared physical environment. In recent years, it has gradually received widespread attention and has become one of the research hotspots of embodied intelligence. Since intelligent agents need to understand and predict dynamic human behaviors and intentions in order to interact and make decisions in real time, social navigation is a more challenging task compared with object navigation and language navigation.

[0003] Existing social navigation research methods focus on minimizing collisions between humans and robots in point target navigation tasks. Some methods construct human-centric navigation datasets in 2D space. Since existing simulators do not provide 3D human animations, the agent cannot observe the human's actions in the scene to understand the human's movement intentions. Other methods enable humanoid agents to move in 3D simulation environments. However, the movement of humanoid agents is controlled by a simple collision avoidance dynamic model, so the humanoid agents just wander aimlessly. Therefore, the agent cannot represent and model real human behaviors and reactions by training on these datasets. In contrast, although some other works provide 3D humanoid agent actions, the animation of different human actions is very costly, resulting in a small dataset size, and there is no natural connection between humanoid agent actions and their movement trajectories, which makes it difficult for the agent's social navigation capabilities to generalize to unknown scenarios. In addition, in addition to keeping a distance from humans, all these methods do not consider how to train agents to complete social navigation tasks by understanding the intention of human actions. Summary of the invention

[0004] In order to solve at least one of the technical problems existing in the prior art to a certain extent, the purpose of the present invention is to provide a social navigation method, device and medium for perceiving and understanding human intentions.

[0005] The first technical solution adopted by the present invention is:

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

[0007] Automatically build a training set for human action navigation through a large language model and text-action generation model;

[0008] Embed the intention predictor module in the basic navigation model to build an intelligent agent navigation model with intent understanding capability;

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

[0010] Furthermore, the method of automatically constructing a training set for human action navigation by using a large language model and a text action generation model includes:

[0011] Obtain text descriptions of action generation through a preset large language model;

[0012] According to the text description, the humanoid agent action clips are automatically generated through the preset text action generation model;

[0013] Obtain the walking trajectory of the humanoid agent, generate walking actions based on the trajectory, splice action clips and walking actions, obtain complete human actions, and construct a training set.

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

[0015] Obtain the object categories and coordinates that can interact with the humanoid agent in the 3D simulation scene, and construct the object set O D , the set O D Provided as candidates to the large language model to ensure that the generated activities can be executed in the scenario;

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

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

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

[0019] Human actions are characterized as in represents a single posture, and N is the length of the action sequence;

[0020] The MDM model is a denoising diffusion model that assumes T noise steps modeled by a random process:

[0021]

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

[0023] The MDM model is given a noisy action X t , time step t and the text description c encoded by CLIP to predict the denoised action At each time step t, a denoised action is predicted and then noised to X t-1 , repeating from t=T until denoising obtains X0; therefore, the MDM model can generate human actions that meet the description based on the noisy actions and text descriptions.

[0024] Furthermore, the step of obtaining the walking trajectory of the humanoid intelligent body, generating walking movements according to the trajectory, and splicing the movement clips and the walking movements to obtain complete human movements includes:

[0025] Based on the coordinates of object A and object B and the navigable map of the scene, the A* algorithm is used to calculate the collision-free trajectory of the humanoid agent walking in the scene;

[0026] Based on the obtained collision-free trajectory and text description of walking, the PriorMDM model is used to generate walking actions;

[0027] The generated action clips and walking movements are stitched together through linear interpolation to form complete human movements, which are integrated into the 3D simulation environment according to the coordinates, ensuring that humans can perform meaningful movements and purposeful walking in the 3D environment.

[0028] Further, 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 feature space; a convolutional neural network is used to extract features for image observations, and a multi-layer fully connected layer is used to extract features for category observations;

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

[0031] The input of the intention predictor module is the human action observed in the environment, including the third-person perspective image I of the human action t , objects that human actions are interacting with Human skeleton P predicted from human motion t And the historical trajectory of human movement T t-5～t ;

[0032] In order to model the intrinsic connection between human observation and human intention, a transformer architecture is constructed to convert intention prediction into a sequence-to-sequence conversion problem; then three additional tags are embedded after the extracted features, namely, CLS obj 、CLS int and REG tra , to reinforce information and make predictions;

[0033] CLS after the transformer network obj 、CLS int and REG tra , respectively predict the intended objects through the linear network Human activity category Y act And the future trajectory of mankind t+1～t+K ; Finally, the output of the intention predictor module and the agent observation are input into the mapping network together.

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

[0035] During the training process, real data corresponding to the action intention prediction results are obtained from the environment, and the auxiliary loss is calculated for supervised learning. The auxiliary loss formula is as follows:

[0036]

[0037]

[0038]

[0039]

[0040] Where P obj , P tct , are the predicted values ​​of intended objects, activity categories, and future trajectories, respectively. Y act， T t～t+K are the true values ​​of the intended object, activity category, and future trajectory, respectively;

[0041] The supervision loss of the intended object and activity category is calculated using the cross entropy loss, and the supervision loss of the future trajectory is calculated using the mean square error loss; the final auxiliary loss is the weighted sum of the three types of losses, and is directly added to the loss of the PPO algorithm and optimized along with reinforcement learning.

[0042] Furthermore, constructing a reinforcement learning reward function includes:

[0043] During the training process, the agent is encouraged to reach the destination efficiently, and according to the human-oriented navigation requirements, the agent should not collide with humans during navigation, so the reward function is designed as follows:

[0044]

[0045]

[0046] In the formula, R t is the reward at time t, 1 indicates task completion, 0 indicates task failure, and when the agent collides with a human The value is 1, otherwise it is 0; The distance from the previous agent to the end point minus the distance from the current agent to the end point indicates that the agent can get a positive reward if it is close to the end point, otherwise the opposite; 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 by the present invention is:

[0048] An electronic device comprises 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 a social navigation method for human intention perception and understanding as described above.

[0049] The third technical solution adopted by the present invention is:

[0050] A computer-readable storage medium stores 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 implement a social navigation method for human intent perception and understanding as described above.

[0051] The fourth technical solution adopted by the present invention is:

[0052] A computer program product or a 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 the processor executes the computer instructions, so that the computer device executes the above-mentioned social navigation method for human intention perception and understanding.

[0053] The beneficial effects of the present invention are as follows: the present invention proposes a human action automatic construction framework, which generates various real 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 data sets and reducing costs. At the same time, the present invention embeds a human action perception module in the basic navigation module, giving the intelligent agent the ability to understand the intention of human actions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the embodiments of the present invention or the drawings of related technical solutions in the prior art are introduced below. It should be understood that the drawings introduced below are only for the convenience of clearly describing 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 work.

[0055] Figure 1 is a flowchart of steps of a social navigation method for human intention perception and understanding in an embodiment of the present invention;

[0056] Figure 2 Schematic diagram of an intelligent agent navigation model with intent understanding capability in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limitations of the present invention. For the step numbers in the following embodiments, they are only provided for the convenience of explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0058] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0059] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0060] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0061] Terminology explanation:

[0062] MDM model: Text-action generation diffusion model is a model used to generate three-dimensional human actions based on text descriptions.

[0063] PriorMDM model: Action generation diffusion model, which is an improved version of the MDM model and can generate three-dimensional 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 for human intention perception and understanding, which specifically includes the following steps:

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

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

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

[0069] Specifically, first obtain the object categories and coordinates that can interact with the humanoid agent in the 3D simulation scene to construct an object set O D , O D The large language model is provided as a candidate to ensure that the generated activities can be performed in the scene. The large language model generates a template based on the action "do [action A] at {object A}, then walk to {object B} and do [action B]" to ensure that the humanoid agent described in the text performs meaningful actions in two different places and avoids wandering aimlessly.

[0070] S12. According to the text description, a humanoid agent action segment is automatically generated through a preset text action generation model.

[0071] In this embodiment, the best current text action generation model MDM is used to generate realistic and diverse human actions based on text descriptions. Human actions are characterized as in represents a single pose. MDM is a denoising diffusion model that assumes T noise steps modeled by a random process:

[0072]

[0073] Where t∈T, X T ~ N(0, I). MDM models the denoising process, which, given a noise action X t , time step t and the text description c encoded by CLIP to predict the denoised action At each time step t, a denoised action is predicted and then noised to X t-1 , repeating from t=T until denoising obtains X0. Therefore, it can generate human actions that meet the description based on the noise action and text description.

[0074] S13. Obtain the walking trajectory of the humanoid agent, generate walking movements according to the trajectory, splice the action clips and the walking movements, obtain complete human movements, and construct a training set.

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

[0076] S2. Embed the intention predictor module in the basic navigation model to build an intelligent navigation model with intent understanding capability.

[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 feature space. Convolutional neural networks are used to extract features for image observations, and multi-layer fully connected layers are used to extract features for category observations. The policy network uses an LSTM network to predict the action that the agent should take at the current moment based on the output features of the mapping network at the current moment, the action at the previous moment, and the features of the LSTM hidden layer. The executable action categories mainly include forward, backward, left turn, right turn, and stop. The basic module gives the agent the ability to navigate and explore in indoor environments.

[0078] In order to understand human intention and achieve navigation, this embodiment embeds an intention predictor module in the basic module. The input of the intention predictor module is the human movement observed from the environment, mainly including the third-person perspective image I of the human movement t , objects that human actions are interacting with Human skeleton P predicted based on human motion t And the historical trajectory of human movement T t-5～t . Extract I through ResNet convolutional network t Features are extracted from the other three types of observations through a linear network. In order to model the intrinsic connection between human observations and human intentions, we further constructed a transformer architecture to transform intent prediction into a sequence-to-sequence conversion problem. Three additional tags are then embedded after the extracted features: CLS obj 、CLS int and REG tra , to strengthen information and make predictions. CLS after the transformer network obj 、CLS int and REG tra Predict the intended objects separately through the linear network Human activity category Y act And the future trajectory of mankind t+1～t+K Finally, the output of the intention predictor module and the agent observation are input into the mapping network. By embedding the intention predictor, the agent can achieve human intention perception and understanding with basic navigation capabilities and complete tasks in complex human motion environments.

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

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

[0081] S31. Construct an auxiliary loss function.

[0082] During the training process, we obtain real data corresponding to the action intention prediction results 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 , are the predicted values ​​of intended objects, activity categories, and future trajectories, respectively. Y act , T t～t+K They are the true values ​​of the intended object, activity category, and future trajectory. The supervision loss of the intended object and activity category is calculated using the cross entropy loss, and the supervision loss of the future trajectory is calculated using the mean square error loss. The final auxiliary loss is the weighted sum of the three types of losses, and is directly added to the loss of the PPO algorithm and optimized along with reinforcement learning.

[0088] S32. Reinforcement learning reward function design.

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

[0090]

[0091]

[0092] If the navigation is successful, you will get a reward of +10, and if a collision occurs, you will get a reward of -1. The distance from the previous agent to the end point minus the distance from the current agent to the end point indicates that the agent can get a positive reward if it is close to the end point, otherwise the opposite. Therefore, the agent is encouraged to move towards the destination. t The last item is a time penalty to encourage the agent to reach the destination as quickly as possible. Through the above reinforcement learning design, the agent performs PPO reinforcement learning training in the constructed simulation dataset, so it can learn social navigation capabilities with intention perception.

[0093] In general, existing social navigation simulation datasets only construct human motion trajectories or generate a small number of human actions manually. Therefore, current social navigation simulations are not close enough to reality, and there are bottlenecks in the cost and efficiency of generating datasets. At the same time, since the existing social navigation datasets only contain human trajectories or only a small number of human actions, the designed method does not include an intent understanding module, resulting in the lack of intent understanding ability of the intelligent agent. The present invention proposes an automatic human action construction framework, which generates a variety of real 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. At the same time, the present invention embeds a human action perception module in the basic navigation module to give the intelligent agent the ability to understand the intention of human actions.

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

[0095] (1) Traditional methods usually rely on manually crafted human actions in social navigation datasets, resulting in small dataset size, high construction cost, and low generation efficiency. The method of the present invention only uses existing deep learning methods (large language model and text action generation model) to automatically generate a large number of diverse and real human actions, thus having significant advantages in dataset construction scale, construction cost, and construction efficiency.

[0096] (2) Existing methods do not consider adding human action intentions due to the scarcity of human actions in the data set, or the lack of data sets leads to weak generalization ability of intelligent agents in social navigation. The method of the present invention integrates the human action intention module into the traditional navigation module and uses a large and diverse set of real human action data for training. Therefore, the intelligent agent has significant advantages and generalization ability in intention understanding navigation.

[0097] Example 2

[0098] An embodiment of the present invention further provides an electronic device, the electronic device comprising a processor and a memory, the memory storing 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 the following Figure 1 A social navigation approach towards human intent perception and understanding is shown.

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

[0100] The processor may include one or more processing cores. The processor uses various interfaces and lines to connect the various parts of the entire server, and executes various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory, and calling data stored in the memory. Optionally, the processor can be implemented in at least one hardware form of digital signal processing (DSP), field programmable gate array (FPGA), and programmable logic array (PLA). The processor can integrate one or a combination of a central processing unit (CPU) and a modem. Among them, the CPU mainly processes the operating system and application programs; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor, but implemented separately through a chip.

[0101] Since the 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 the electronic device is similar to that of the method, the implementation of the electronic device can refer to the implementation process of the above-mentioned method embodiment, and the repeated parts will not be repeated.

[0102] Example 3

[0103] The embodiment of the present invention further provides a computer-readable storage medium, wherein the storage medium stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by a processor to implement the following Figure 1 A social navigation approach towards human intent perception and understanding is shown.

[0104] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0105] Since the storage medium is a 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 solving the problem by the storage medium is similar to that of the method, the implementation of the storage medium can refer to the implementation process of the above-mentioned method embodiment, and the repeated parts will not be repeated.

[0106] Example 4

[0107] In some possible implementations, various aspects of the method of the embodiments of the present invention may also be implemented in the form of a program product, which includes a program code. When the program product is run on a computer device, the program code is used to enable the computer device to execute the steps of a social navigation method for human intention perception and understanding according to various exemplary embodiments of the present application described above in this specification. Among them, the executable computer program code or "code" for executing various embodiments may be written in a high-level programming language 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 the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0109] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0110] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A social navigation method for human intention perception and understanding, characterized in that: The following steps are involved: Automatically build a training set for human action navigation through a large language model and text-action generation model; Embed the intention predictor module in the basic navigation model to build an intelligent agent navigation model with intent understanding capability; Construct reinforcement learning reward function and auxiliary loss function, and train the intelligent agent navigation model through reinforcement learning.

2. A social navigation method for human intention perception and understanding according to claim 1, characterized in that: The method of automatically constructing a training set for human action navigation by using a large language model and a text action generation model includes: Obtain text descriptions of action generation through a preset large language model; According to the text description, the humanoid agent action clips are automatically generated through the preset text action generation model; Obtain the walking trajectory of the humanoid agent, generate walking actions based on the trajectory, splice action clips and walking actions, obtain complete human actions, and construct a training set.

3. A social navigation method for human intention perception and understanding according to claim 1, characterized in that: The step of obtaining a text description of the action generated by using a preset large language model includes: Obtain the object categories and coordinates that can interact with the humanoid agent in the 3D simulation scene, and construct the object set O D , the set O D Provided as candidates to the large language model to ensure that the generated activities can be executed in the scenario; The large language model generates templates based on preset actions to ensure that the human-like agent described in the text performs meaningful actions in two different places.

4. A social navigation method for 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 characterized as where x i represents a single posture, and N is the length of the action sequence; The MDM model is a denoising diffusion model that assumes T noise steps modeled by a random process: In the formula, α t is a cumulative scaling factor, which represents the degree of change of the data from the initial state to the state at time t, and I is the unit matrix; The MDM model is given a noisy action X t , time step t and the text description c encoded by CLIP to predict the denoised action At each time step t, a denoised action is predicted and then noised to X t-1 , repeating from t=T until denoising obtains X0; therefore, the MDM model can generate human actions that meet the description based on the noisy actions and text descriptions.

5. The social navigation method for human intention perception and understanding according to claim 1, characterized in that: The step of obtaining the walking trajectory of the humanoid intelligent body, generating walking actions according to the trajectory, and splicing the action clips and the 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 A* algorithm is used to calculate the collision-free trajectory of the humanoid agent walking in the scene; Based on the obtained collision-free trajectory and text description of walking, the PriorMDM model is used to generate walking actions; The generated action clips and walking movements are stitched together through linear interpolation to form complete human movements, which are integrated into the 3D simulation environment according to the coordinates, ensuring that humans can perform meaningful movements and purposeful walking in the 3D environment.

6. A social navigation method for human intention perception and understanding according to claim 1, characterized in that: The basic navigation model includes a mapping network and a policy network; The mapping network is used to map observations in the environment to feature space; a convolutional neural network is used to extract features for image observations, and a multi-layer fully connected layer is used to extract features for category observations; The strategy network is implemented using an LSTM network, which is used to predict the action taken by the agent at the current moment based on the output features of the mapping network at the current moment, the action at the previous moment, and the LSTM hidden layer features; The input of the intention predictor module is the human action observed in the environment, including the third-person perspective image I of the human action t , objects that human actions are interacting with Human skeleton P predicted from human motion t And the historical trajectory of human movement T t-5～t ; In order to model the intrinsic connection between human observation and human intention, a transformer architecture is constructed to convert intention prediction into a sequence-to-sequence conversion problem; then three additional tags are embedded after the extracted features, namely, CLS obj 、CLS int and REG tra , to reinforce information and make predictions; CLS after the transformer network obj 、CLS int and REG tra , respectively predict the intended objects through the linear network Human activity category Y act And the future trajectory of mankind t+1～t+K ; Finally, the output of the intention predictor module and the agent observation are input into the mapping network together.

7. The social navigation method for human intention perception and understanding according to claim 1, characterized in that: The constructing of the auxiliary loss function comprises: During the training process, real data corresponding to the action intention prediction results are obtained from the environment, and the auxiliary loss is calculated for supervised learning. The auxiliary loss formula is as follows: Where P obj , P act , are the predicted values ​​of intended objects, activity categories, and future trajectories, respectively. Y act , T t～t+K are the true values ​​of the intended object, activity category, and future trajectory, respectively; The supervision loss of the intended object and activity category is calculated using the cross entropy loss, and the supervision loss of the future trajectory is calculated using the mean square error loss; the final auxiliary loss is the weighted sum of the three types of losses, and is directly added to the loss of the PPO algorithm and optimized along with reinforcement learning.

8. The social navigation method for human intention perception and understanding according to claim 1, characterized in that: The construction of the reinforcement learning reward function includes: During the training process, the agent is encouraged to reach the destination efficiently, and according to the human-oriented navigation requirements, the agent should not collide with humans during navigation, so the reward function is designed as follows: In the formula, R t is the reward at time t, 1 indicates task completion, 0 indicates task failure, and when the agent collides with a human The value is 1, otherwise it is 0; The distance from the previous agent to the end point minus the distance from the current agent to the end point indicates that the agent can get a positive reward if it is close to the end point, otherwise the opposite; Δ(b t ,h t ) represents the distance between the agent and the target point at the current moment.

9. 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 described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The storage medium 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 according to any one of claims 1 to 8.

Images