Bayesian-based progressive intent alignment teammate adaptation method and system

Abstract

The application discloses a teammate adaptive method and system based on a Bayesian progressive intention alignment, fuses natural language priori and likelihood estimation of teammate interaction behavior in a unified Bayesian framework, realizes rapid convergence of teammate intention and efficient alignment in a single round of interaction, fully utilizes high-level semantic information in language description, and overcomes hysteresis and ambiguity when intention is inferred by relying on a state-action sequence in a traditional way. The application effectively solves intention confusion and low-efficiency adaptation problems of an existing method in a multi-intention scene. Overall, the application has the advantages of fast adaptation speed, high inference precision and no need of additional demonstration, and has a wide application prospect in a human-machine cooperation scene.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system technology, specifically to a teammate adaptation method in a multi-agent reinforcement learning environment, and more particularly to a Bayesian-based progressive intent alignment teammate adaptation method and system. Background Technology

[0002] Collaborative multi-agent reinforcement learning has made significant progress in recent years in solving complex tasks involving multi-agent interactions, and has shown great application potential in fields such as autonomous driving, resource scheduling, and embodied intelligence. Under the traditional "closed-world" assumption, a fixed group of agents is usually jointly trained to optimize a specific task. Such methods, such as those based on value decomposition or policy gradients, perform well on standard benchmark tests.

[0003] However, in real-world "open-world" scenarios, intelligent agents often need to temporarily team up with previously unknown and unseen teammates to complete collaborative tasks; this scenario is known as zero-shot collaboration. For example, in autonomous driving scenarios, self-driving vehicles must coordinate with human drivers with varying behavioral habits; in the field of service robots, robots need to quickly adapt to the operating habits of different users. In such scenarios, the ability to quickly adapt to teammates and adaptively adjust strategies is key to achieving efficient collaboration.

[0004] In recent years, teammate population generation training methods have provided new theoretical and technical pathways for zero-shot collaboration. These methods generate a teammate population with diverse behavioral characteristics during the training phase, attempting to cover the types of teammates that may be encountered during testing, thereby enabling the strategy to adapt to unseen teammates. However, these methods typically assume that the task has only a single collaborative goal, and perform poorly when faced with complex tasks with "multimodal collaborative equilibrium." That is, when the task has multiple incompatible valid collaborative goals (the teammates' own "intent"), existing methods often fail to correctly infer the teammates' intentions and complete the collaboration.

[0005] While existing teammate modeling and intent inference methods offer feasible approaches to understanding the intentions behind teammate behavior, current frameworks still lack the ability for rapid intent recognition and efficient alignment when faced with collaborative tasks involving complex semantic intents. This stems from the fact that existing methods primarily rely on low-dimensional, low-information-density low-level behavioral cues (such as teammate states and action sequences) for inference, neglecting the crucial role of high-level semantic priors, such as natural language, in revealing intent. This results in significant lag and uncertainty in the intent inference process. Therefore, constructing a unified framework that organically integrates high-level semantic priors and low-level interaction evidence to achieve progressive intent alignment from fuzzy linguistic descriptions to precise behavioral strategies has become a key scientific and engineering challenge for efficient human-machine collaboration in open environments. Summary of the Invention

[0006] Purpose of the Invention: This invention aims to address the problems of low efficiency in intent inference, slow adaptation speed, and weak generalization ability to unseen strategies in existing teammate adaptation methods. This invention presents a Bayesian-based progressive intent alignment teammate adaptation method and system, achieving rapid intent recognition and collaboration within a single round of interaction. This method organically integrates language and teammate behavior as prior probabilities and likelihood probabilities using Bayesian inference, providing a new approach for accurate and rapid inference of teammate intent.

[0007] Technical Solution: A Bayesian-based progressive intent alignment method for adaptive teammate alignment achieves rapid identification and efficient alignment of teammate intents in single-round interactions by organically integrating linguistic priors and interaction evidence. Specifically, it includes the following steps: Step 1: Construct a multi-intent teammate strategy population. For multi-intent collaboration scenarios, first define the intent pool. Each intention This represents a possible collaborative goal. For each intent... Constructing using multilayer perceptrons Teammate Strategy Network And optimize the rewards independently using different random number seeds. This generates a teammate population that covers a variety of behavioral styles. Then, by utilizing the strategies within the team's population, they collected information on each intention. Multi-track data ,in Indicating the intention to complete The A collaborative trajectory.

[0008] Step 2: Construct a prior estimator for intent description. This is done by summarizing cue words on multi-trajectory data. The trajectories in the data are summarized to obtain a semantic summary for each trajectory, and further extracted prompt words are combined with this summary. Semantic summary extraction of trajectories to extract corresponding intents Description text This describes the high-level semantic features used to represent intent. Then, before actual collaboration begins, teammates provide a description of their intent in natural language. This invention calculates natural language intent descriptions by matching prompt words. The semantic similarity between each intent description and the prior probability distribution is obtained by normalization: in and These represent the normalization process and the matching process, respectively.

[0009] Step 3: Distill teammate strategies and construct a likelihood estimator for interaction behaviors. Given an intent... Teammate interaction history ,in These represent the environmental state and teammate actions, respectively. The probability of the interaction history occurring is also mentioned. It can be decomposed into state-action pair probabilities. The product of the state-action pair probabilities through the intention Lower Teammates Each possible strategy The average value is used to calculate the result, as shown below: in For normalization term, Intent The number of strategies in the next teammate population. Because... The number of possibilities is often large, and directly using the above formula to estimate the likelihood would result in huge costs. Therefore, this invention aims to achieve the same goal. The following strategy family distillation is a single strategy Accelerate the likelihood estimation process and optimize distillation losses: in Indicates sampling intents from the intent pool and from the corresponding trajectory data Extraction status After distillation, the likelihood estimator only needs to call the distillation strategy once to calculate the likelihood probability: Step 4: Construct a Bayesian intent predictor. Combining the prior probabilities obtained in Step 2 and the likelihood probabilities obtained in Step 3, this invention uses a Bayesian update rule to estimate the teammate's current intent in real time. For teammate interaction history... Its posterior probability is calculated using Bayes' theorem as follows: The posterior distribution is continuously updated during the interaction process, enabling it to gradually converge from an ambiguous intent to a single intent. Intent entropy is used to determine whether the intent has converged. judge: when Greater than the threshold If the intention is still unclear, the predictor gives an empty intention state; otherwise, it is considered that the intention has converged, and the intention with the highest posterior probability is selected as the predicted intention.

[0010] At this point, we can organically integrate intent descriptions and teammate behavior to accurately predict teammate intents. We still need to build a strategy for our own intelligent agents to collaborate with teammates during deployment.

[0011] Step 5: Train the inference and cooperation strategies. To avoid biased behaviors that affect inference when the agent's intent has not yet converged, an intent-independent inference strategy is constructed. This strategy does not target any specific intent in action selection, making teammate behavior more reflective of their intentions. When training this inference strategy, reinforcement learning can be used to maximize the following objectives: Finish: in express Two strategies in intent The collaboration rewards are as follows. Secondly, intent-related collaboration strategies are constructed for scenarios with clearly defined intentions. This strategy, during training, is consistent with the intention of... By leveraging teammate strategies and maximizing collaborative rewards through reinforcement learning, our agent can maximize team gains when the intention is known. This concludes the training steps for the Bayesian-based progressive intent alignment adaptive method proposed in this invention. The testing and deployment process is described below.

[0012] Step Six: Given an unknown teammate and a description of their intentions. The description is transformed using a prior estimator. .

[0013] Step 7: Use inference strategies Take a step with an unknown teammate and collect their interaction history. Use a Bayesian intent predictor to infer the teammate's intent.

[0014] Step 8: Repeat Step 7 until the intent predictor no longer gives empty intents and predicts specific intents. .

[0015] Step Nine: Then use the corresponding collaboration strategy. To achieve efficient collaboration with teammates.

[0016] A Bayesian-based progressive intent alignment teammate adaptation system includes the following modules: The teammate population generation module is used to generate a diverse range of teammate populations with different behaviors for different intentions. The language prior extraction module is used to map the teammate's intention description in natural language form into an intention prior vector; The interactive information extraction module is used to extract features from the observation information during the collaboration process; The intent inference module is used to estimate the distribution of teammates' intents based on Bayesian inference mechanisms combined with linguistic priors and interaction evidence. The collaboration strategy generation module is used to generate corresponding collaborative actions based on the inferred intentions of teammates; The teammate population generation module mainly consists of an intent pool and a teammate policy network, wherein the intent pool... It includes various collaborative tasks that teammates might expect, and the teammate policy network is implemented using a multilayer perceptron. This module generates [instructions] for each possible intent in the intent pool. The various collaborative strategies that can accomplish the task constitute the corresponding teammate population. .

[0017] The aforementioned language prior extraction module is mainly implemented through a large language model, consisting of three types of prompts: summarizing, extracting, and matching. The first two types of prompts can reveal the collaborative trajectory under the corresponding intent. In summary And extract it into the corresponding intent language. The last type of prompt can describe the natural language intent. Matching with different intent languages ​​and mapping them to prior probabilities based on similarity. .

[0018] The interactive information extraction module mainly consists of a distillation network and a likelihood estimator. The distillation network distills multiple intention policy populations into a single policy. This accelerates likelihood extraction; the likelihood estimator then utilizes a distillation network to compute the teammate interaction history under intent T. probability .

[0019] The intent inference module mainly consists of a posterior calculator and an intent predictor. The posterior calculator calculates the prior probability. and likelihood probability The posterior probability is calculated using Bayes' theorem. The intent predictor calculates the intent entropy based on the posterior probability. When the entropy is below a certain threshold, the one with the highest posterior probability is predicted as the true intent; otherwise, the intent is predicted as empty.

[0020] The cooperative strategy generation module mainly consists of an inference strategy and a cooperative strategy, both of which are multilayer perceptrons. It possesses the characteristic of being intent-independent (empty intent), which can avoid interfering with intent inference behavior; collaborative strategy. It possesses intent-related characteristics, enabling it to efficiently collaborate with unseen teammates to complete tasks when given an intent.

[0021] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the Bayes-based progressive intent alignment teammate adaptation method as described above.

[0022] A computer-readable storage medium having a computer program / instructions thereon that, when executed by a processor, implements the steps of the Bayes-based progressive intent alignment teammate adaptation method as described above.

[0023] Beneficial Effects: Compared with existing technologies, the proposed Bayesian-based progressive intent alignment teammate adaptation method integrates natural language priors and likelihood estimation of teammate interaction behavior within a unified Bayesian framework, achieving rapid convergence of intent inference and efficient alignment within a single round of interaction. This method fully utilizes high-level semantic information in language descriptions, overcoming the lag and ambiguity of traditional intent inference methods that rely solely on state-action sequences. Furthermore, it reduces likelihood computation costs through teammate policy distillation, enabling the system to maintain stable and reliable real-time inference capabilities even when facing unseen teammates. This invention effectively solves the problems of intent confusion and inefficient adaptation in multi-intent scenarios in existing methods. Overall, this invention possesses advantages such as fast adaptation speed, high inference accuracy, and no need for additional demonstrations, making it widely applicable in human-computer collaboration scenarios. Attached Figure Description

[0024] Figure 1 This is a flowchart of a Bayesian-based progressive intent alignment teammate adaptation method in an embodiment of the present invention. Figure 2 This is the overall framework of the Bayesian-based progressive intent alignment adaptive teammate system in this embodiment of the invention. Figure 3 This is a schematic diagram of the teammate population generation module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the language prior extraction module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the interactive information extraction module in an embodiment of the present invention; Figure 6 This is a schematic diagram of the intent prediction module in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0026] The teammate adaptation problem can be modeled as a cooperative Markov game. ,in This is the state space, which includes possible state information in the environment; The action space refers to the actions that teammates and our intelligent agents can take. Let be the transition probability function, representing the state. The probability of taking an action to transition to another state; This is the intent pool, representing all possible intents; For the reward function, i.e., intention The following collaborative goals; The discount factor is used. The goal of the teammate adaptation problem is to learn the optimal strategy for our team. To maximize and a certain intention of teammates Global expected cumulative return In addition, this invention makes the additional assumption that teammates can use natural language. Describe their intentions. For this scenario, the optimal strategy for our side needs to organically integrate natural language and teammate behavior to infer the teammate's intentions and to conduct efficient collaboration based on those intentions.

[0027] To achieve rapid intent recognition and collaboration within a single round of interaction, Bayesian inference can be used to organically fuse language and teammate behavior as prior probabilities and likelihood probabilities, accurately and quickly inferring teammate intent. The Bayesian-based progressive intent alignment teammate adaptive system mainly consists of six modules: The teammate population generation module is used to generate a diverse range of teammate populations with different behaviors for different intentions. The language prior extraction module is used to map the teammate's intention description in natural language form into an intention prior vector; The interactive information extraction module is used to extract features from the observation information during the collaboration process; The intent inference module is used to estimate the distribution of teammates' intents based on Bayesian inference mechanisms combined with linguistic priors and interaction evidence. The collaboration strategy generation module is used to generate corresponding collaborative actions based on the inferred intentions of teammates; The teammate population generation module mainly consists of an intent pool and a teammate policy network, wherein the intent pool... It includes various collaborative tasks that teammates might expect, and the teammate policy network is implemented using a multilayer perceptron. This module generates [instructions] for each possible intent in the intent pool. The various collaborative strategies that can accomplish the task constitute the corresponding teammate population. .

[0028] The language prior extraction module is mainly implemented through a large language model with prompts, consisting of three types of prompts: summarizing, extracting, and matching. The first two types of prompts can reveal the collaborative trajectory under the corresponding intent. In summary And extract it into the corresponding intent language. The last type of prompt can describe the natural language intent. Matching with different intent languages ​​and mapping them to prior probabilities based on similarity. .

[0029] The interactive information extraction module mainly consists of a distillation network and a likelihood estimator. The distillation network distills a population of multiple intention policies into a single policy. This accelerates likelihood extraction; the likelihood estimator then utilizes a distillation network to compute the teammate interaction history under intent T. probability .

[0030] The intent inference module mainly consists of a posterior calculator and an intent predictor. The posterior calculator calculates the prior probability. and likelihood probability The posterior probability is calculated using Bayes' theorem. The intent predictor calculates the intent entropy based on the posterior probability. When the entropy is below a certain threshold, the one with the highest posterior probability is predicted as the true intent; otherwise, the intent is predicted as empty.

[0031] The collaborative strategy generation module mainly consists of inference strategies and collaborative strategies, both of which are multilayer perceptrons. It possesses the characteristic of being intent-independent (empty intent), which can avoid interfering with intent inference behavior; collaborative strategy. It possesses intent-related characteristics, enabling it to efficiently collaborate with unseen teammates to complete tasks when given an intent.

[0032] A Bayesian-based progressive intent alignment method for adaptive teammate alignment achieves rapid identification and efficient alignment of teammate intents in single-round interactions by organically integrating linguistic priors and interaction evidence. Specifically, it includes the following steps: Step 1: Construct a multi-intent teammate strategy population. For multi-intent collaboration scenarios, first define the intent pool. Each intention This represents a possible collaborative goal. For each intent... Constructing using multilayer perceptrons Teammate Strategy Network And optimize the rewards independently using different random number seeds. This generates a teammate population that covers a variety of behavioral styles. Then, by utilizing the strategies within the team's population, they collected information on each intention. Multi-track data .

[0033] Step 2: Construct a prior estimator for intent description. This is done by summarizing cue word pairs. The trajectories in the data are summarized to obtain a semantic summary for each trajectory, and further extracted prompt words are combined with this summary. Extracting the corresponding intent from the summary of the trajectory Description text This describes the high-level semantic features used to represent intent. Then, before actual collaboration begins, teammates provide a description of their intent in natural language. This invention calculates the semantic similarity between natural language instructions and each intent description by matching prompt words, and obtains the prior probability distribution through normalization: in and These represent the normalization process and the matching process, respectively.

[0034] Step 3: Distill teammate strategies and construct a likelihood estimator for interaction behaviors. Given an intent... Below is the teammate interaction history. The probability of its occurrence is: in For normalization term, Intent The number of strategies in the next teammate population, due to The number of possibilities is often large, and directly using the above formula to estimate the likelihood would result in huge costs. Therefore, this invention aims to achieve the same goal. The following strategy family distillation is a single strategy Accelerate the likelihood estimation process and optimize distillation losses: in Indicates sampling intents from the intent pool and from the corresponding trajectory data Extraction status After distillation, the likelihood estimator only needs to call the distillation strategy once to calculate the likelihood probability: Step 4: Construct a Bayesian intent predictor. Combining the prior probabilities obtained in Step 2 and the likelihood function obtained in Step 3, this invention uses a Bayesian update rule to estimate the teammate's current intent in real time. For teammate interaction history... Its posterior probability is calculated using Bayes' theorem as follows: This posterior distribution is continuously updated during the interaction process, enabling the system to gradually converge from fuzzy intentions to a single intention. Intention entropy is used to determine whether the intention has converged. judge: when Greater than the threshold If the intention is still unclear, the predictor gives an empty intention state; otherwise, it is considered that the intention has converged, and the intention with the highest posterior probability is selected as the predicted intention.

[0035] At this point, we can organically integrate intent descriptions and teammate behavior to accurately predict teammate intents. We still need to build a strategy for our own intelligent agents to collaborate with teammates during deployment.

[0036] Step 5: Train the inference and cooperation strategies. To avoid biased behaviors that affect inference when the agent's intent has not yet converged, an intent-independent inference strategy is constructed. This strategy does not target any specific intent in action selection, making teammate behavior more reflective of their intent. During training, it maximizes collaborative rewards with all generated strategies through reinforcement learning. in This indicates that the two strategies are in the intention The collaboration rewards are as follows. Secondly, intent-related collaboration strategies are constructed for scenarios with clearly defined intentions. This strategy, during training, is consistent with the intention of... By leveraging teammate strategies and maximizing collaborative rewards through reinforcement learning, our agent can maximize team gains when the intention is known. This concludes the training steps for the Bayesian-based progressive intent alignment adaptive method proposed in this invention. The testing and deployment process is described below.

[0037] Step Six: Given an unknown teammate and a description of their intentions. The description is transformed using a prior estimator. .

[0038] Step 7: Use inference strategies Collaborate with them to collect teammate interaction history. Use a Bayesian intent predictor to infer the teammate's intent.

[0039] Step 8: Repeat Step 7 until the intent predictor no longer gives empty intents and predicts specific intents. .

[0040] Step Nine: Then use the corresponding collaboration strategy. To achieve efficient collaboration with teammates.

[0041] It is evident that those skilled in the art will understand that the steps of the Bayesian-based progressive intent alignment teammate adaptation method or the modules of the Bayesian-based progressive intent alignment teammate adaptation system described in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using device-executable program code, thereby storing them in a storage device for execution by the computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. A Bayesian-based progressive intent alignment teammate adaptive method, characterized in that, By organically integrating language priors and interaction evidence, rapid identification and efficient alignment of teammate intentions in single-round interactions are achieved, specifically including the following steps: Step 1: Construct a multi-intent teammate strategy population; Step 2: Construct a prior estimator for intent description; summarize the trajectory by summarizing cue words, and further extract the corresponding intent using extracted cue words. Description text Given an intent description At that time, the semantic similarity between natural language instructions and each intent description is calculated by matching prompt words, and the prior probability distribution is obtained by normalization; Step 3: Distill teammate strategies and construct a likelihood estimator for interaction behaviors; given the intent, the likelihood probability of a teammate's interaction history occurring is... It can be estimated through the teammate population; to reduce the estimation cost, a distillation strategy is used. Replace the teammate population and quickly calculate the likelihood probability; Step 4: Construct a Bayesian intent predictor; estimate the posterior probability using Bayes' theorem based on the prior and likelihood probabilities obtained in Steps 2 and 3; predict the intent based on the entropy of the posterior probability. Step 5: Train the inference and collaboration policies; construct an intent-independent inference policy by training to maximize the collaboration rewards with all generated policies through reinforcement learning. ; Step Six: Given an unknown teammate and a description of their intentions. The description is transformed using a prior estimator. ; Step 7: Use inference strategies Collaborate with unknown teammates in one step and collect teammate interaction history. Use a Bayesian intent predictor to infer the intent of teammates; Step 8: Repeat Step 7 until the intent predictor no longer gives empty intents and predicts specific intents. ; Step Nine: Use the corresponding collaboration strategy To achieve efficient collaboration with teammates.

2. The Bayes-based progressive intent alignment teammate adaptation method according to claim 1, characterized in that, In step one, for multi-intent collaboration scenarios, an intent pool is first set up. Each intention Indicate a possible collaborative goal; for each intention Constructing using multilayer perceptrons Teammate Strategy Network And optimize the rewards independently using different random number seeds. This generates a teammate population that covers a variety of behavioral styles. ; Then, by utilizing strategies within the team's population, each intention was collected. Multi-track data ,in Indicating the intention to complete The A collaborative trajectory.

3. The Bayes-based progressive intent alignment teammate adaptation method according to claim 1, characterized in that, In step two, the multi-track data is analyzed by summarizing the prompt words. The trajectories are summarized to obtain a semantic summary for each trajectory, and the corresponding intent is extracted based on the semantic summary. Description text Next, before the actual collaboration begins, teammates describe their intentions in natural language. ; Calculate natural language intent description by matching prompt words The semantic similarity between each intent description and the prior probability distribution is obtained by normalization: in and These represent the normalization process and the matching process, respectively.

4. The Bayes-based progressive intent alignment teammate adaptation method according to claim 1, characterized in that, In step three, the intention is given. Teammate interaction history ,in These represent the state in the environment and the actions of teammates, respectively; the probability of the interaction history occurring. It can be decomposed into state-action pair probabilities. The product of the state-action pair probabilities through the intention Lower Teammates Each possible strategy The average value is used to calculate the result, as shown below: in For normalization term, Intent The number of strategies within the next teammate population. Furthermore, the same intention can be... The following strategy family distillation is a single strategy To accelerate the likelihood estimation process and optimize distillation losses: in Indicates sampling intents from the intent pool and from the corresponding trajectory data Extraction status After distillation, the likelihood estimator only needs to call the distillation strategy once to calculate the likelihood probability: 。 5. The Bayes-based progressive intent alignment teammate adaptation method according to claim 1, characterized in that, In step four, combining the prior probability obtained in step two and the likelihood probability obtained in step three, a Bayesian update rule is used to estimate the teammate's current intention in real time; for teammate interaction history... Its posterior probability is calculated using Bayes' theorem as follows: The posterior distribution is continuously updated during the interaction process, gradually converging from a vague intent to a single intent; this is achieved using intent entropy. Determine whether the judgment intent has converged: when Greater than the threshold If the intention is still unclear, the predictor gives an empty intention state; otherwise, it is considered that the intention has converged, and the intention with the highest posterior probability is selected as the predicted intention.

6. The Bayes-based progressive intent alignment teammate adaptation method according to claim 1, characterized in that, In step five, to avoid our agent taking biased actions that affect inference before the intent has converged, an intent-independent inference strategy is constructed. This inference strategy does not target any specific intention in action selection, making teammate behavior more reflective of their intent. During training, this inference strategy can be maximized through reinforcement learning, achieving the following objectives: Finish: in express Two strategies in intent The rewards of collaboration; Secondly, construct intent-related collaboration strategies for scenarios with clearly defined intents. This collaborative strategy, during training, is aligned with the intention of... By leveraging teammate strategies and maximizing collaborative rewards through reinforcement learning, our agent can maximize team gains when the intention is known. 。 7. A Bayesian-based progressive intent alignment teammate adaptive system, characterized in that, Includes the following modules: The teammate population generation module is used to generate a diverse range of teammate populations with different behaviors for different intentions. The language prior extraction module is used to map the teammate's intention description in natural language form into an intention prior vector; The interactive information extraction module is used to extract features from the observation information during the collaboration process; The intent inference module is used to estimate the distribution of teammates' intents based on Bayesian inference mechanisms combined with linguistic priors and interaction evidence. The collaboration strategy generation module is used to generate corresponding collaborative actions based on the inferred intentions of teammates.

8. The Bayes-based progressive intent alignment teammate adaptive system according to claim 7, characterized in that, The teammate population generation module mainly consists of an intent pool and a teammate policy network, wherein the intent pool... It includes various collaborative tasks that teammates might expect, and the teammate policy network is implemented using a multilayer perceptron; this module generates [instructions] for each possible intent in the intent pool. The possible collaborative strategies for completing the task constitute the corresponding teammate types; The aforementioned language prior extraction module is mainly implemented through a large language model, consisting of three types of prompts: summarizing, extracting, and matching. The first two types of prompts can reveal the collaborative trajectory under the corresponding intent. In summary And extract it into the corresponding intent language. The last type of prompt can describe the natural language intent. Matching with different intent languages ​​and mapping them to prior probabilities based on similarity. ; The interactive information extraction module mainly consists of a distillation network and a likelihood estimator. The distillation network distills multiple intention policy populations into a single policy. This accelerates likelihood extraction; the likelihood estimator then utilizes a distillation network to compute the teammate interaction history under intent T. probability ; The intent inference module mainly consists of a posterior calculator and an intent predictor. The posterior calculator calculates the prior probability. and likelihood probability The posterior probability is calculated using Bayes' theorem. The intent predictor calculates the intent entropy based on the posterior probability. When the entropy is below a certain threshold, the one with the highest posterior probability is predicted as the true intent; otherwise, the intent is predicted as empty. The cooperative strategy generation module mainly consists of an inference strategy and a cooperative strategy, both of which are multilayer perceptrons. It possesses the characteristic of being unrelated to intent, thus avoiding interference with intent inference behavior when inferring intent; Collaboration strategy It possesses intent-related characteristics, enabling it to efficiently collaborate with unseen teammates to complete tasks when given an intent.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the Bayes-based progressive intent alignment teammate adaptation method as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instruction is executed by the processor, it implements the steps of the Bayes-based progressive intent alignment teammate adaptation method as described in any one of claims 1-6.

Images