User intention alignment robot task planning method based on large language model

Abstract

The invention relates to the technical field of robots, in particular to a user intention alignment robot task planning method based on a large language model, and the method comprises the steps: firstly receiving human task voice, carrying out the textualization processing of the voice, carrying out the human instruction ambiguity elimination through combining with a robot task environment, and generating a feasible instruction candidate set; and then calling the maintained user feature memory module by the large language model, and further screening the candidate instructions in combination with human user preferences. And finally, the large language model decomposes and plans the reserved task instruction, guides the robot to execute the task, collects human feedback information in the task process, and is used for optimizing and updating the user feature memory module and guaranteeing the task planning continuity of the large language model driven by reflection feedback. According to the method, when the service robot based on the large language model carries out task planning, the instruction intention of the human user is analyzed and aligned more accurately, and customized services are better provided for the human user.

Description

Technical Field

[0001] The present invention relates to the field of robotics technology, and in particular to a user intention alignment robot task planning method based on a large language model. Background Art

[0002] Robot task planning is an essential part of robot autonomous planning, which aims to enable the robot to plan a reasonable action sequence based on the environment, task objectives and its own capabilities. Service robots refer to autonomous or semi-autonomous robots used in non-manufacturing industries with service as the core. They can maintain all-weather, stable quality and efficient service operations. At present, the application scenarios covered by service robots have covered many fields such as hotels, medical care, education, logistics, etc., and they are playing an increasingly important role in modern society.

[0003] Traditional robot task planning methods are mainly implemented by defining a language framework in the planning domain. In recent years, large language models (LLMs) with strong ability to understand and generate human language have been rapidly developed using a large amount of corpus data for training and fine-tuning. Large language models have strong reasoning ability and rich general knowledge. At present, a series of studies have used large language models for robot task planning and achieved good results. However, for service robots with relatively complex task sequences and frequent interactions with human users, the paradigm of using large language models for task planning is difficult to meet the robot task requirements. Specifically, the human users faced by service robots may have uncertainty in the wording of their intentions when issuing instructions, and different human users have their own preference characteristics. In addition, during the process of robot execution, human users will generate rich task feedback. Robots need to comprehensively utilize the above information related to human users to better serve human users, complete specific user instructions, and meet user needs in a targeted manner. Summary of the invention

[0004] In order to solve the above technical problems existing in the prior art, the present invention proposes a user intention alignment robot task planning method based on a large language model, and its specific technical solution is as follows: A user intention alignment robot task planning method based on a large language model, comprising: Step 1: The robot receives voice commands from human users and performs text parsing; Step 2: Disambiguate the intent of the parsed voice command in combination with the robot task environment to generate a command candidate set; Step 3: The large language model calls the user feature memory module maintained to store user preference information, and screens the candidate instructions in combination with human user preferences, retaining the precise instructions that best match the current user preferences; Step 4: The large language model decomposes and plans the retained task instructions; Step 5: The robot performs the decomposed and planned tasks and collects feedback from human users during the tasks for optimization and updating of the user feature memory module; Step 6: Set up a large language model task planning continuity guarantee mechanism driven by reflection feedback.

[0005] Furthermore, in step 1, the robot receives voice commands from human users and uses a voice conversion module to transcode the user voice into a natural language text format as input to the robot's large language model.

[0006] Furthermore, in step 2, the large language model is combined with the robot task environment to eliminate ambiguity in the wording and description of human user instructions, so as to parse out alternative instructions, i.e., a candidate set of instructions, that meet both the current task scenario and human user needs.

[0007] Furthermore, in step 3, the large language model selects the precise instruction that best matches the current user's preference from the alternative instructions parsed in the previous step based on the user feature memory module that is continuously learned and maintained, combined with the preference characteristics of the current user being served, and performs user preference fusion on the precise instructions.

[0008] Furthermore, in step 4, the large language model decomposes the precise instructions into tasks, plans each sub-task after the decomposition, and then outputs a detailed sequence of task steps to be executed, guiding the embodied robot to perform task actions according to the planned sequence.

[0009] Furthermore, in step 5, the robot based on the large language model continuously collects feedback information on whether the human user determines that the task is successful during the execution of the task, and then uses the feedback information to update and maintain the user feature memory module for autonomous learning and continuous optimization.

[0010] Furthermore, in step 6, the large language model uses a reflective architecture to utilize the task execution status and human feedback information collected during the task process. When the task execution fails, it autonomously uses the feedback information to generate a reflective text, and then adds this reflective text to the next round of task re-planning to correct the erroneous behaviors during the task process and perform continuous self-optimization.

[0011] Beneficial effects: The method of the present invention enables the service robot based on the large language model to more accurately parse and align the human user's instruction intentions when performing task planning, so as to better provide customized services for human users. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a flow chart of a method for aligning robot task planning with user intention based on a large language model of the present invention; Figure 2 is a detailed example schematic diagram of human user intention alignment according to an embodiment of the present invention; Figure 3 This is an example diagram of a reflective feedback text according to an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical scheme and technical effect of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0014] like Figure 1 As shown, the present invention proposes a user intention alignment robot task planning method based on a large language model, wherein the robot receives voice commands from human users and performs text processing on the voice commands; the voice commands are ambiguous in intent in combination with the robot task environment to generate a command candidate set; the large language model calls the maintained user feature memory module, and further screens the candidate commands in combination with human user preferences to retain the precise commands that best match the current user preferences; the language model decomposes and plans the retained task commands; the robot executes the decomposed and planned tasks, and collects human feedback information during the task process for optimization and updating of the user feature memory module; at the same time, a large language model task planning continuity guarantee mechanism driven by reflective feedback is set. Specifically, the method includes the following steps: Step 1: Receiving and parsing user instructions: The robot receives the human user's voice and parses the voice into instructions in the form of natural language text, that is, parses to obtain the instruction text.

[0015] Specifically, the robot receives voice commands from human users and uses a speech conversion module to transcode the user's speech into a natural language text format as input to a large language model that acts as the robot's brain.

[0016] Step 2: User intent disambiguation: The large language model is combined with the robot's specific task environment to disambiguate human user instructions in order to parse out alternative instruction sets that meet both the current task scenario and human user needs.

[0017] Specifically, after receiving the transcoded human user instructions, the large language model robot combines the current task environment and scenario, uses its rich knowledge and reasoning ability to reason and analyze the uncertain descriptions in the human user instructions, so that the user's task intention is aligned with the specific environment, so as to generate a task candidate set that is allowed by the current task scenario and can meet the top-level needs of human users. For example, "drinks" are concretized into "water, juice, and wine" that exist in the current task environment, and a command candidate set that meets the task environment is generated.

[0018] Step 3: User preference learning and fusion: Based on the large language model, the user customize memory module is called in real time to further filter the instructions of the candidate set parsed in the previous step, and finally retain the precise instructions that best meet the current specific user preferences, perform user preference fusion on the precise instructions, and then use the filtered instructions for the next step of task decomposition planning. The user customize memory module is used to store user preference information.

[0019] The large language model continuously learns and maintains a user feature memory module, collects user preference information during interaction with human users, and dynamically updates and iterates this memory module to continuously learn the characteristics and preferences of human users, and provide customized services for interacting human users during task planning and instruction parsing.

[0020] Step 4: Large language model task decomposition and planning: Based on the large language model, the task instructions are decomposed after ambiguity elimination and user preference matching, and the human instructions are decomposed into further sub-tasks for planning. Then a detailed sequence of task steps to be executed is output to guide the embodied robot to perform task actions according to the planned sequence.

[0021] Specifically, the large language model first decomposes the received and processed human top-level instructions into several sub-task sequences, and then plans the action steps for each sub-task sequence in detail. Finally, according to this logical mechanism, it guides the embodied service robot to complete the action sequence step by step to plan and execute the overall task instructions. For example, "help me heat up some cookies" is decomposed into the sub-task sequence "go to the kitchen, find cookies, heat up cookies", and then a more detailed action sequence is planned for each sub-task, such as decomposing "find cookies" into "approach the dining table, grab the bowl, grab the cookies, and put the cookies in the bowl". After completing the task decomposition and planning, the large language model guides the embodied robot body to execute the planned task sequence.

[0022] Step 5: Collect user feedback: The robot based on the large language model collects feedback information on whether the user judges the task to be successful through its own task actions, and uses this feedback information to subsequently optimize the user feature memory module and conduct reflection and error correction on the robot task sequence.

[0023] Specifically, during the execution of tasks, the robot continuously collects feedback signals of task success or failure, and collects human user feedback at each task step to update and adjust the user-customized memory module and task planning logic details in real time. Under this mechanism, the large language model robot continuously learns and self-optimizes to better align with human user intention preferences and provide user-customized services.

[0024] Step 6: Set up a feedback-enhanced task continuity guarantee mechanism: When the robot fails to perform a task, it collects task feedback from human users through the robot sensor and processes the feedback information into text format as a reflection text for the large language model to re-plan the task. The large language model uses the reflection text information and human user feedback to correct errors and failures in the task execution process, so that the robot has the ability to recover from errors autonomously, ensuring the task continuity of the robot. The overall process of large language model planning and embodied robot execution of tasks constitutes a closed loop for task continuity guarantee.

[0025] For example, in one embodiment, if the robot fails to perform the subtask, the large language model will generate text-based reflective feedback based on the collected feedback information, such as "The degree of heating of the biscuits failed to satisfy the user, and I will heat the biscuits with a higher fire in the next round." The large language model uses this feedback text to re-plan the failed subtask to perform autonomous correction and recovery of the error. At the same time, the feedback information of human users will also be used to optimize user feature memory, such as "This user prefers hotter and sweeter flavors."

[0026] The core technical content of the method of the present invention is the alignment of human user intentions. Figure 2 As shown, in one embodiment, when aligning user intent, the large language model robot first receives human voice instructions, processes and parses them into text format, and then the large language model eliminates the ambiguity of human wording according to the specific task scenario, and generates a task instruction set that conforms to the environment. Then the large language model calls the user feature memory module, such as "the user currently served never drinks and needs to supplement vitamin C", and further eliminates the ambiguity of the task instruction in combination with the specific user preferences and needs, such as "bring me a drink" is finally disambiguated into "heat a glass of orange juice". This precise task instruction is open to the user through an interactive interface, allowing human users to confirm. After confirming the specific task instruction, the large language model takes this task instruction as input, decomposes it into several subtasks, and then performs detailed action planning for each subtask, such as "go to the kitchen, open the refrigerator, take out orange juice, heat orange juice, and go to the bedroom to give it to the user". Finally, the human user gives a judgment of satisfaction and detailed feedback based on the task planning and execution results. The large language model receives the above feedback for updating and iteration of user feature memory, as well as reflection and replanning of failed actions.

[0027] The method of the present invention as a whole constitutes a feedback-enhanced task continuity guarantee closed loop, such as Figure 3 Figure 2 shows an example of a large language model reflecting on text.

[0028] The above is only a preferred implementation case of the present invention and does not limit the present invention in any form. Although the implementation process of the present invention is described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the above examples, or replace some of the technical features therein with equivalents. All modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A user intention alignment robot task planning method based on a large language model, characterized in that: include: Step 1: The robot receives voice commands from human users and performs text parsing; Step 2: Disambiguate the intent of the parsed voice command in combination with the robot task environment to generate a command candidate set; Step 3: The large language model calls the user feature memory module maintained to store user preference information, and screens the candidate instructions in combination with human user preferences, retaining the precise instructions that best match the current user preferences; Step 4: The large language model decomposes and plans the retained task instructions; Step 5: The robot performs the decomposed and planned tasks and collects feedback from human users during the tasks for optimization and updating of the user feature memory module; Step 6: Set up a large language model task planning continuity guarantee mechanism driven by reflection feedback.

2. The method for robot task planning based on user intention alignment of large language model as claimed in claim 1, characterized in that: In step 1, the robot receives human user voice commands and uses a speech conversion module to transcode the user voice into a natural language text format as input to the robot's large language model.

3. The user intention alignment robot task planning method based on a large language model as claimed in claim 2, characterized in that: In step 2, the large language model combines the robot task environment to eliminate ambiguity in the wording and description of human user instructions, so as to parse out alternative instructions, i.e., the instruction candidate set, that meet both the current task scenario and human user needs.

4. The method for robot task planning based on user intention alignment of large language model as claimed in claim 3, characterized in that: In step 3, the large language model selects the precise instruction that best matches the current user's preference from the alternative instructions parsed in the previous step based on the user feature memory module that is continuously learned and maintained, combined with the preference characteristics of the current user being served, and performs user preference fusion on the precise instructions.

5. The method for robot task planning based on user intention alignment of large language model as claimed in claim 4, characterized in that: In step 4, the large language model decomposes the precise instructions into tasks, plans each sub-task after the decomposition, and then outputs a detailed sequence of task steps to be executed, guiding the embodied robot to perform task actions according to the planned sequence.

6. The method for robot task planning based on user intention alignment of large language model as claimed in claim 5, characterized in that: In step 5, the robot based on the large language model continuously collects feedback information on whether the human user determines the task is successful during the execution of the task, and then uses the feedback information to update and maintain the user feature memory module for autonomous learning and continuous optimization.

7. The method for robot task planning based on user intention alignment of large language model as claimed in claim 6, characterized in that: In step 6, the large language model uses a reflective architecture to utilize the task execution status and human feedback information collected during the task. When the task fails, it autonomously uses the feedback information to generate a reflective text, and then adds this reflective text to the next round of task re-planning to correct the erroneous behaviors during the task and perform continuous self-optimization.

Images