Information processing device, planning method, and program

Abstract

To provide an information processing device that can suppress the generation of plans that include logically contradictory or unnecessary actions. [Solution] The system comprises: a decomposition unit that decomposes an input task into subtasks using a first large-scale language model; an ordering unit that determines the execution order of the subtasks using a second large-scale language model; and a determination unit that determines, according to the determined order, whether the subtasks can be executed by actuators in relation to the current state of the task object using a third large-scale language model. The decomposition unit decomposes the subtask if the first large-scale language model determines that the subtask is not executable.

Description

Technical Field

[0001] The present disclosure relates to an information processing apparatus, a planning method, and a program. Background Art

[0002] In recent years, large language models (LLM: Large Language Model) have exhibited high inference capability, and have attracted growing attention for application as agents in various decision-making processes. When given a task (linguistic instruction), an LLM agent outputs an action (plan) for solving the task. Prior Art Literature Non-Patent Literature

[0003] Non-Patent Literature 1 Shangzi Xue, Zhenya Huang, Jiayu Liu, Xin Lin, Yuting Ning, Binbin Jin, Xin Li, Qi Liu,“Decompose, Analyze and Rethink: Solving Intricate Problems with Human-like Reasoning Cycle”,The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024),September 2024 Summary of the Invention Problem to be Solved by the Invention

[0004] In planning using LLM, a sequence of actions (plan) is generated to achieve the task (objective) based on the current state. In such planning, as the task becomes more complex, the search space increases, and it may become impossible to maintain state consistency within the same inference process. As a result, problems may arise in the generated plan, such as the inclusion of logically contradictory or unnecessary actions. For example, an action to separate objects that have already been separated may be proposed, or an action to connect objects back together may be proposed after an action to separate them has been proposed.

[0005] Non-limiting embodiments of this disclosure contribute to providing information processing devices, planning methods, and programs that can prevent the generation of plans that include logically contradictory or unnecessary actions. [Means for solving the problem]

[0006] An information processing device according to one embodiment of the present disclosure includes: a decomposition unit that decomposes an input task into subtasks using a first large-scale language model; an ordering unit that determines the execution order of the subtasks using a second large-scale language model; and a determination unit that determines, according to the determined order, the feasibility of the subtasks being executed by actuators for the current state of the task object using a third large-scale language model, wherein the decomposition unit decomposes the subtask if the first large-scale language model determines that the subtask is not executable.

[0007] A planning method according to one embodiment of the present disclosure involves an information processing device that decomposes an input task into subtasks using a first large-scale language model, determines the execution order of the subtasks using a second large-scale language model, determines the feasibility of the subtasks by actuators for the current state of the task object using a third large-scale language model according to the determined order, and decomposes the subtask if the first large-scale language model determines that the subtask is not executable.

[0008] A program according to one embodiment of the present disclosure causes a computer to perform the following processes: decompose an input task into subtasks using a first large-scale language model; determine the execution order of the subtasks using a second large-scale language model; determine, according to the determined order, whether the subtasks are executable by actuators in relation to the current state of the task object using a third large-scale language model; and, if the first large-scale language model determines that the subtasks are not executable, decompose the subtasks.

[0009] These comprehensive or specific embodiments may be implemented as a system, apparatus, method, integrated circuit, computer program, or recording medium, or as any combination of a system, apparatus, method, integrated circuit, computer program, and recording medium. [Effects of the Invention]

[0010] According to one embodiment of the present disclosure, it is possible to suppress the generation of plans that include logically contradictory or unnecessary actions.

[0011] Further advantages and effects of one embodiment of this disclosure will be made apparent from the specification and drawings. Such advantages and / or effects are provided by several embodiments and features described in the specification and drawings, but not all of them are necessarily provided in order to obtain one or more identical features. [Brief explanation of the drawing]

[0012] [Figure 1] A diagram illustrating a bottom-up planning method. [Figure 2] Diagram illustrating the overview of this disclosure. [Figure 3] Block diagram of an information processing device [Figure 4] Diagram illustrating operation 1 [Figure 5] Diagram explaining operation 2 [Figure 6] Diagram explaining operation 3 [Figure 7] Diagram for explaining operation 4 [Figure 8] Diagram for explaining operation 5 [Figure 9] Diagram for explaining operation 6 [Figure 10] Diagram for explaining operation 6 [Figure 11] Diagram for explaining operation 7 [Figure 12] Diagram for explaining operation 7 [Figure 13] Diagram for explaining operation 8 [Figure 14] Flowchart showing the operation of the information processing apparatus [Figure 15] Flowchart showing the operation of the information processing apparatus [Figure 16] Flowchart showing the operation of the information processing apparatus [Figure 17] Flowchart showing the operation of the information processing apparatus [Figure 18] Hardware configuration diagram of the information processing apparatus DESCRIPTION OF EMBODIMENTS

[0013] Hereinafter, embodiments of the present disclosure will be described in detail with appropriate reference to the drawings. However, excessive detailed description may be omitted in some cases. For example, detailed description of already well-known matters and redundant description of substantially the same configuration may be omitted. This is to avoid unnecessary redundancy in the following description and facilitate understanding for those skilled in the art.

[0014] The accompanying drawings and the following description are provided for those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter recited in the claims thereby.

[0015] <Terms used in the present disclosure> · Actuator An actuator refers to an entity that receives and executes a subtask generated by the information processing device 1. For example, an actuator may be a robot, an agent that performs operations (an LLM-based agent), or a human. Furthermore, an actuator is not necessarily limited to a real, existing entity; it may also be a virtual robot or the like that performs simulations in a virtual environment such as a digital twin. ·task A task is a linguistic instruction input to the information processing device 1, and represents the objective, or goal, in this disclosure. A task may include one or more objectives (instructions). Subtasks A subtask is a smaller unit of processing that is broken down to accomplish a larger task. Subtasks, plans, and actions can be used interchangeably. • Necessity of execution The necessity of execution is an indicator used to determine whether or not a particular subtask needs to be performed. If the current state already meets the objective of the subtask, then it is determined that the subtask does not need to be performed. ·situation The "state" refers to the environmental information of the task object, including information such as its position, orientation, relationships, and attributes. The term "state" may also be interpreted as "situation." Feasibility Feasibility is an indicator used to determine whether a subtask can be executed in the current state, given that there are no physical or logical constraints. For example, if the current state satisfies the prerequisites for executing a subtask, that subtask is determined to be feasible. ·Wood structure A tree structure is a data structure that hierarchically represents tasks and subtasks. Each node in a tree structure holds information (status) such as the task and subtask, the execution order, the result of the execution necessity assessment, the result of the execution feasibility assessment, and whether or not a subtask has been executed (e.g., subtask completed).

[0016] <Overview> Figure 1 illustrates a bottom-up planning method. The bottom-up planning method determines the next action to take to achieve the objective, starting from the current state. For example, as shown in Figure 1, the bottom-up planning method determines the next action to take towards the goal (task) from the initial state. Therefore, as the task becomes more complex, the search space becomes enormous (for example, the number of arrows shown in Figure 1 increases), making it easier for contradictions or hallucinations between plans to occur.

[0017] In contrast, the planning method described in this disclosure determines the actions to be taken based on the objective (the given task). In other words, the planning method described in this disclosure employs a top-down procedure.

[0018] The planning method for this disclosure will employ a top-down procedure, as well as the following methods. (1) Break down the task or subtask into a tree structure. (2) Determine the feasibility of each decomposed task or subtask at any given time. If it is determined that the task or subtask cannot be executed, further decompose the task or subtask (by moving down one level in the tree structure).

[0019] Figure 2 is a diagram illustrating the overview of this disclosure. In Figure 2, we consider a stack of blocks. We assume that the blocks are manipulated one by one.

[0020] Initially, a yellow block is placed on the table, and a blue block is placed on top of that block. Also, initially, a red block is placed on the table, and an orange block is placed on top of that block.

[0021] In Figure 2, "O" represents orange, "R" represents red, "B" represents blue, and "Y" represents yellow. A brief explanation is provided here; a more detailed explanation will be given in the <Operation Example> below.

[0022] The planning method described herein uses LLM to decompose tasks into subtasks (plans). For example, the planning method described herein decomposes tasks into subtasks as shown in S1, S2, and S5 in Figure 2. S1, S2, and S5 in Figure 2 are subtasks at the first level of the tree structure. A task is the root node.

[0023] The planning method of this disclosure uses LLM to determine whether a decomposed subtask is executable in the current state (in this case, the initial state). For example, in the initial state, the orange block can be placed on the table. Therefore, the planning method of this disclosure determines that S1 is executable. In other words, the planning method of this disclosure determines that a subtask is executable if the current state satisfies the preconditions for the execution of that subtask (for example, conditions such as there being no block on top of the orange block and the orange block being able to be placed on the table).

[0024] The planning method described in this disclosure changes the current state according to the subtask that has been determined to be executable, if it is determined to be executable. The planning method described in this disclosure sequentially determines whether the subtasks (S1, S2, S5) in the first level are executable, and changes the state if they are determined to be executable. If the task is completed through this operation, the process ends.

[0025] On the other hand, if there are subtasks that cannot be executed at the first level, the disclosure planning method further breaks down those subtasks (lowers the tree hierarchy). For example, if the disclosure planning method determines that S1 and S2 in Figure 2 can be executed but subtask S5 in Figure 2 cannot be executed, it breaks down subtask S5 as shown in S3 and S4. S3 and S4 in Figure 2 are subtasks at the second level of the tree structure.

[0026] The same processing is applied to subtasks at the second level and below as to the first level. Once all tasks at a lower level are executable, the process returns to the level above, and once all tasks at that level are executable, it returns to the level above. If it can return to the root, the task is completed. If it cannot return to the root, the task is not completed.

[0027] According to the planning method of this disclosure, the state in which blocks should be placed (subtasks) is generated from the task using LLM. If the subtasks generated by LLM are not executable, those subtasks are further broken down into subtasks. This suppresses the expansion of the search space and reduces inconsistencies or hallucinations between plans.

[0028] <Block diagram> Figure 3 is a block diagram of the information processing device 1. The planning method of this disclosure is performed by the information processing device 1. The information processing device 1 may be, for example, a personal computer, a server, a tablet, or a smartphone. As shown in Figure 3, the information processing device 1 includes an input unit 11, a decomposition unit 12, a sequencing unit 13, a management unit 14, a determination unit 15, and an execution unit 16.

[0029] The input unit 11 accepts tasks via an input device such as a keyboard.

[0030] The decomposition unit 12 uses LLM to decompose the given task or a subtask determined to be unexecutable into smaller subtask groups in a tree structure.

[0031] The ordering unit 13 determines the execution order of subtasks using the LLM. For example, the ordering unit 13 instructs the LLM to determine an order in which all subtasks included in the same hierarchical level of the tree structure are executed without contradiction. The ordering unit 13 may also instruct the LLM to determine an order that takes into account the initial state of the multiple subtasks before they are executed.

[0032] The management unit 14 manages tasks, subtasks generated by the decomposition unit 12, and status as a tree structure, not using LLM, but according to predefined rules (a program describing the rules). For example, the management unit 14 generates nodes corresponding to tasks and each subtask, and has each node hold a string representing the task or subtask. More specifically, each node has a list (for example, a list type in Python) that holds child nodes, and adds child nodes corresponding to subtasks generated by the decomposition of the task corresponding to that node to this list. The root node of the tree structure is the node corresponding to the task. The tree structure may be constructed in the memory of the information processing device 1, or it may be constructed in an external memory separate from the information processing device 1.

[0033] The determination unit 15 uses LLM to determine whether a subtask needs to be executed in relation to the current state. For example, suppose the current state is that there are no blocks on top of the red block. Suppose the subtask is to "place the block on top of the red block onto the table". In this case, since there are no blocks on top of the red block, there is no need to place the block on top of the red block onto the table. In this case, the determination unit 15 determines that there is no need to execute the subtask. In other words, if the state that the subtask aims to achieve has already been achieved by the current state, it is determined that there is no need to execute the subtask.

[0034] If the determination unit 15 determines that there is a need for execution, it uses LLM to determine whether the subtask is executable in the current state. If the current state satisfies the prerequisites for the subtask to be executed, it is determined that it is executable.

[0035] On the other hand, if the determination unit 15 determines that there is no need to perform the action, it does not perform the feasibility determination. This is because, as described above, if there is no need to perform the action, the state that is to be achieved by the execution of the subtask has already been achieved by the current state.

[0036] The execution unit 16 executes subtasks that the determination unit 15 has determined to be executable. The execution unit 16 also prevents the actuator from executing subtasks that it has determined do not need to be executed. The execution unit 16 updates the current state by executing subtasks.

[0037] The LLM used by each part may be a different session (dialogue state) of a single LLM, or it may be a different LLM. In this embodiment, the LLM used by each part is not executed sequentially in the same inference context, but rather a configuration is adopted in which each process is executed as an independent inference. The LLM tends to output a response that corresponds to the context (inference context) based on the dialogue history until the session is interrupted. Therefore, when using a single LLM in each part, the inference can be made independent for each process of each part by having different sessions. Alternatively, by using different LLMs in each part, the inference context can be made independent for each LLM used by each part.

[0038] Each part of the process has different underlying state information, inference objectives, and evaluation criteria. For example, plan generation and subtask decomposition involve inference based on assumed future states, while subtask ordering and feasibility assessment require evaluation of logical consistency and feasibility based on the current or established state. If these processes with different premises are executed within the same inference context, assumed states and states being verified may become mixed, potentially leading to unnecessary re-proposals of actions or judgments inconsistent with the state. Therefore, in this embodiment, each process is separated as an independent inference unit, and each performs inference under clear underlying states, thereby improving the overall logical consistency and reliability of the planning.

[0039] If it is difficult to prepare multiple inference contexts due to implementation or computational resource constraints, each process may be executed sequentially using the same LLM. When using the same LLM, the inference contexts may be made independent by clearly stating the assumptions such as state information, the purpose of inference, and evaluation criteria in the prompts to the LLM, or by instructing the LLM to discard all past inference results and start a new inference via prompts.

[0040] Furthermore, the LLM function may be provided by the information processing device 1, or it may be provided by a separate device such as a server.

[0041] <Example of operation> This example illustrates the planning process for a task that involves arranging multiple blocks of different colors into a stack arranged in a predetermined manner. In this example, the following prompt is input to the information processing device 1.

[0042] I'm playing with a set of building blocks that requires me to stack them up and organize them into a stack. My ultimate goal is to reach the desired state as efficiently as possible, with as few failures as possible. My goal is to stack the blocks so that the top layer is yellow, the middle layer is red, and the bottom layer is orange. The initial state is as follows: The blue block is on top of the yellow block, and the orange block is on top of the red block. Also, the red and yellow blocks are on the table. Please tell me what actions are necessary to achieve the above goal, provided that the following conditions are met. (1) The task must be directly accomplished if all the proposed actions are performed. Please tell me all the necessary conditions. (2) The possible actions are to place a block on top of another block, to place a block on the table, or to place the block on top of another block on the table. There are no other conditions. (3) Please specify which block you are going to manipulate. (4) When placing a block on top of another block, there must be nothing on top of the block being placed.

[0043] The above is the prompt that is input to the information processing device 1. In Figures 4 to 13 described below, "O" indicates orange, "R" indicates red, "B" indicates blue, and "Y" indicates yellow. In Figures 4 to 13, "Sx" (where x is a number) indicates the execution order of the subtasks. The execution order of the subtasks "Sx" may be changed during the planning process (see, for example, operation 5 below).

[0044] • Action 1 (Subtask decomposition and determination of execution order) Figure 4 is a diagram illustrating operation 1. Among the prompts above, "Stack blocks in yellow at the top, red in the middle, and orange at the bottom" corresponds to the task. The information processing device 1 extracts the task from the input prompts using LLM and breaks it down into subtasks. In other words, the information processing device 1 gives the LLM the task, and the LLM generates subtasks to solve that task.

[0045] LLM generates the following subtasks from the given task (see also Figure 4). Place the orange block on the table. Place the red block on the orange block. Place the yellow block on the red block.

[0046] The information processing device 1 determines the execution order of the generated (decomposed) subtasks using LLM. For example, LLM determines the following execution order: S1: Place the orange block on the table. S2: Place the red block on the orange block. S3: Place the yellow block on the red block.

[0047] • Action 2 (Determination of necessity and feasibility of execution) Figure 5 illustrates operation 2. The information processing device 1 uses LLM to determine the necessity of executing a subtask given the current state of the block.

[0048] For example, the information processing device 1 uses LLM to determine whether it is necessary to execute the subtask S1 shown in Figure 5. The current state (initial state) does not satisfy the condition of the subtask S1, "Place the orange block on the table". Therefore, LLM determines that it is necessary to execute the subtask S1.

[0049] If the information processing device 1 determines that it is necessary to execute a subtask, it uses LLM to determine the feasibility of the subtask.

[0050] For example, the information processing device 1 uses LLM to determine the feasibility of the subtask S1 shown in Figure 5. In the current state (initial state), no blocks are placed on the orange block, and the orange block can be placed on the table. Therefore, LLM determines that the subtask S1 is feasible (OK).

[0051] • Action 3 (Execution) Figure 6 illustrates operation 3. When the information processing device 1 determines that the subtask of S1 is executable, it instructs the actuator to execute the subtask of S1. After the execution of the subtask of S1, the current state of the block becomes the "state after execution of S1" shown in Figure 6.

[0052] • Action 4 (Repeat actions 2 and 3 at the same level) Figure 7 is a diagram illustrating operation 4. The information processing device 1 repeats operations 2 and 3, namely the determination of the necessity and feasibility of execution, and the execution by the actuator, at the same hierarchical level of the tree structure, according to the execution order determined in operation 1.

[0053] For example, the current state (the "state after execution of S1" in Figure 7) does not satisfy the condition of the S2 subtask "Place the red block in the orange block". Therefore, LLM determines that the S2 subtask needs to be executed.

[0054] For example, in the current state, there is no red block on top of the orange block, and a red block can be placed on top of the orange block. Therefore, LLM determines that the subtask S2 is executable (OK).

[0055] • Action 5 (Subtask breakdown and execution order determination) Figure 8 illustrates operation 5. Figure 8 shows the state after the execution of the S2 subtask (see "State after execution of S2" in Figure 8).

[0056] If there are subtasks that cannot be executed, the information processing device 1 uses LLM to decompose those subtasks and deepens the hierarchy of the tree structure by one level.

[0057] For example, in the current state (Figure 8, "State after execution of S2"), the blue block is placed on top of the yellow block. Therefore, the subtask "Place the yellow block on the red block" in S5 (the reason for the change from S3 to S5 will be explained later) in Figure 8 is not executable (NG). Accordingly, the information processing device 1 decomposes the subtask "Place the yellow block on the red block" in S5 using LLM.

[0058] For example, LLM generates the following subtasks from a given subtask in S5. Place the block on top of the red block onto the table. Place the block on top of the yellow block onto the table.

[0059] The information processing device 1 determines the execution order of the generated (decomposed) subtasks using LLM. For example, LLM determines the following execution order: S3: Place the block on top of the red block onto the table. S4: Place the block on top of the yellow block onto the table.

[0060] In other words, LLM changes the order of parent nodes so that after the execution of the child node's subtask (S3, S4 in Figure 8), the parent node that has that child node, i.e., the subtask that was determined to be unexecutable (S5 in Figure 8), is executed.

[0061] For example, the subtask that was S3 in the operations shown in Figures 4 to 7 becomes a subtask of S5 in Figure 8. The subtasks S1, S2, and S5, their execution order, and status are stored in a node one level below the root node (task), while the subtasks S3 and S4, and their execution order, are stored in a node two levels below the root node.

[0062] • Action 6 (Repeat actions 2 and 3 at the same level) Figure 9 is a diagram illustrating operation 6. The information processing device 1 repeats operations 2 and 3, namely the determination of the necessity and feasibility of execution, and execution by the actuator, in a one-level deeper hierarchy (one level lower).

[0063] For example, in the current state (the "state after execution of S2" in Figure 9), no block is placed on top of the red block, and the state of the S3 subtask "Place the block on top of the red block on the table" is already satisfied. Therefore, the LLM determines that it does not need to execute the S3 subtask, i.e., it has already been executed (OK), and the information processing device 1 does not execute the S3 subtask with the actuator.

[0064] Next, the information processing device 1 uses LLM to determine whether it is necessary to execute the subtask S4, "Place the block on top of the yellow block on the table." For example, the current state does not satisfy the condition for the subtask S4, "Place the block on top of the yellow block on the table." Therefore, LLM determines that it is necessary to execute the subtask S4.

[0065] If the information processing device 1 determines that it is necessary to execute the subtask of S4, it uses LLM to determine the feasibility of executing the subtask of S4. In the current state, a blue block is placed on top of a yellow block, and the blue block can be placed on the table. Therefore, LLM determines that the subtask of S4 is executable (OK).

[0066] If the information processing device 1 determines that the subtask of S4 is executable, it instructs the actuator to execute the subtask of S4. After the execution of the subtask of S4, the current state of the block becomes the "State after execution of S4" shown in Figure 10.

[0067] Furthermore, if there are subtasks that cannot be executed in a child node, the information processing device 1 repeats operations 5 and 6 while descending one level in the tree structure hierarchy.

[0068] • Operation 7 (Repeat of operations 2 and 3 on the parent node) Figure 11 is a diagram illustrating operation 7. When the information processing device 1 has executed all subtasks at the same hierarchical level of a child node, it executes operations 2 and 3 on the subtasks of the parent node of that child node.

[0069] For example, the information processing device 1 uses LLM to determine whether it is necessary to execute the subtask "Place yellow blocks in red blocks" of S5, which is the parent node of S3 and S4. The current state (the state after execution of S4) does not satisfy the condition for the subtask "Place yellow blocks in red blocks" of S5. Therefore, LLM determines that it is necessary to execute the subtask of S5.

[0070] If the information processing device 1 determines that it is necessary to execute the subtask S5, it uses LLM to determine the feasibility of executing the subtask S5. In the current state, there are no yellow blocks on top of the red blocks, and it is possible to place yellow blocks on top of the red blocks. Therefore, LLM determines that the subtask S5 is executable (OK).

[0071] If the information processing device 1 determines that the subtask of S5 is executable, it instructs the actuator to execute the subtask of S5. After the execution of the subtask of S5, the state of the current block becomes the "state after execution of S5" shown in Figure 12.

[0072] • Action 8 (End of planning) Figure 13 illustrates operation 8. The information processing device 1 repeats operations 2 and 3 from the child node to the root node. When the information processing device 1 determines that the root node, i.e., the task execution is unnecessary or feasible, it terminates the planning process.

[0073] For example, the information processing device 1 determines whether the task that is the parent node of the subtasks S1, S2, and S5 needs to be executed. In the current state (the state after execution of S5), yellow blocks, red blocks, and orange blocks are placed from top to bottom, satisfying the task condition "stack blocks with yellow on top, red in the middle, and orange at the bottom." Therefore, LLM determines that the task does not need to be executed, i.e., it has already been executed (OK), and the information processing device 1 terminates task planning.

[0074] <Flowchart> Figures 14 to 17 are flowcharts illustrating the operation of the information processing device 1. In the flowcharts shown in Figures 14 to 17, processing continues with A to E as connection points. As shown in Figures 14 to 17, the operation of the information processing device 1 is broadly composed of tree structure construction, execution process, and parent node execution process.

[0075] The information processing device 1 causes the LLM to generate subtasks from the input task (S101).

[0076] The information processing device 1 causes another LLM (for example, another session) to generate the execution order of the subtasks generated by the LLM (S102).

[0077] The information processing device 1 converts the subtask strings generated by the LLM into a list (S103).

[0078] The information processing device 1 holds one element of the list (subtask) in each child node (S104).

[0079] The information processing device 1 determines the necessity of executing each subtask in relation to the current state using LLM (S105).

[0080] If the information processing device 1 determines that there is no need to perform the task (NO in S106), it adds it to the completed list as a subtask that has already been completed (S107). Then, the information processing device 1 changes the status of the subtask to "completed" (S108).

[0081] On the other hand, if the information processing device 1 determines that there is a need to perform the task (YES in S106), it uses LLM to determine whether the subtask is executable in the current state (S109).

[0082] If the information processing device 1 determines that the subtask is not executable (NO in S110), it sets the status of the subtask to "Not executable" (S111).

[0083] On the other hand, if the information processing device 1 determines that the subtask is executable (YES in S110), it causes the actuator to execute the subtask (S112).

[0084] After any of the processes in S108, S111, and S112 are executed, the information processing device 1 determines whether or not there has been a change in the status of each subtask (S113). If there has been a change in the status (YES in S113), the information processing device 1 proceeds to process S105.

[0085] On the other hand, if there is no change in status (NO in S113), that is, if the determination process for the necessity or feasibility of execution has been completed for all subtasks at the same hierarchical level, the information processing device 1 decomposes (subdivides) the subtasks with a status of "not executable" using LLM (S114).

[0086] The information processing device 1 determines whether the status of all child nodes is completed or not (S115). If the status of all child nodes is not completed (NO in S115), the information processing device 1 considers the decomposition to have failed, lowers LLM parameters such as Temperature, and restarts the decomposition of the parent node (S116). Then, the information processing device 1 moves the process to S114.

[0087] On the other hand, if the status of all child nodes is completed (YES in S115), the information processing device 1 adds the parent node whose child nodes are all "completed" to the completed list (S117).

[0088] The information processing device 1 sets the status of the parent node's subtask to "completed" (S118).

[0089] The information processing device 1 uses LLM to determine the necessity of executing each subtask of the parent node in relation to the current state (S119).

[0090] If the information processing device 1 determines that there is no need to perform the action (NO in S120), it skips the execution by the actuator (S121).

[0091] On the other hand, if the information processing device 1 determines that there is a need to perform the task (YES in S120), it causes the actuator to perform the subtask (S122).

[0092] The information processing device 1 determines whether the iterative processing from S117 to S122 has reached the root node (S123). If the iterative processing from S117 to S122 has not reached the root node (NO in S123), the information processing device 1 proceeds to S117 in order to repeat the processing for the "cannot be executed" subtask in the higher level because the current state has changed.

[0093] On the other hand, if the iterative processing of S117 to S122 reaches the root node (YES in S123), the information processing device 1 determines whether the status of the root node is "completed" or not (S124). If the status of the root node is not "completed" (NO in S124), the information processing device 1 determines that the planning has failed.

[0094] On the other hand, if the status of the root node is "completed" (YES in S124), the information processing device 1 determines that the planning was successful and terminates the process. At this stage, the completed list contains a history of the series of subtasks that the actuator performed to accomplish the task, and the task can be accomplished by executing the subtasks as recorded in the completed list. Therefore, this completed list can be used as the result (plan) of the planning.

[0095] <Hardware Configuration Diagram> Figure 18 is a hardware configuration diagram of the information processing device 1. As shown in Figure 18, the information processing device 1 includes a processor 21, a communication unit 22, and a storage unit 23.

[0096] The processor 21 is composed of components such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 21 realizes the functions of each component described in Figure 3 by executing programs stored in the memory unit 23.

[0097] The communication unit 22 communicates with other devices based on the control of the processor 21. The communication unit 22 communicates with other devices via wired and wireless connections.

[0098] The memory unit 23 stores the programs executed by the processor 21. The memory unit 23 also stores data generated by the processor 21 and data used in the processor 21's processing. The tree structure is constructed by the memory unit 23.

[0099] <Summary of this disclosure> The information processing device 1 decomposes the input task into subtasks using LLM and determines the execution order of the subtasks using LLM. The information processing device 1 then uses LLM to determine the feasibility of the subtasks by actuators in relation to the current state of the task object, according to the determined order. If the LLM determines that a subtask is not executable, the information processing device 1 decomposes the subtask.

[0100] In this way, the information processing device 1 provides a task to the LLM and causes the LLM to generate subtasks to solve that task. Then, if the LLM determines that the execution of the generated subtask by the actuator is impossible, the information processing device 1 decomposes that subtask. This suppresses the generation of plans that include logically contradictory or unnecessary actions. It also suppresses the expansion of the search space and suppresses contradictions or hallucinations between plans.

[0101] Furthermore, because tasks and subtasks are managed as a tree structure, the connections between tasks and subtasks become easier for the user to see, improving the explainability of the planning process.

[0102] Furthermore, if the LLM determines that a subtask is unexecutable, the information processing device 1 further decomposes the subtask. Then, the information processing device 1 deepens the tree structure hierarchy by one level and manages the further decomposed subtasks in the deepened hierarchy. In this way, subtasks determined to be unexecutable are further decomposed and managed in lower levels, improving the accuracy of the planning.

[0103] While embodiments have been described above with reference to the drawings, this disclosure is not limited to such examples. It will be apparent to those skilled in the art that various modifications or alterations can be conceived within the scope of the claims. Such modifications or alterations are also understood to fall within the technical scope of this disclosure. Furthermore, the components in the embodiments may be combined in any way without departing from the spirit of this disclosure.

[0104] In the above-described embodiment, the notation "...part" used for each component may be replaced with other notations such as "...circuitry", "...assembly", "...device", "...unit", or "...module".

[0105] This disclosure can be implemented in software, hardware, or software in conjunction with hardware. Each functional block used in the description of the above embodiments may be implemented in part or in whole as an integrated circuit (LSI), and each process described in the above embodiments may be controlled in part or in whole by a single LSI or a combination of LSIs. An LSI may consist of individual chips, or it may consist of a single chip that includes some or all of the functional blocks. An LSI may have data inputs and outputs. Depending on the degree of integration, LSIs may be referred to as ICs, system LSIs, super LSIs, or ultra LSIs.

[0106] The method of integration is not limited to LSIs; it may also be implemented using dedicated circuits, general-purpose processors, or dedicated processors. Furthermore, FPGAs (Field Programmable Gate Arrays) that can be programmed after LSI manufacturing, or reconfigurable processors that allow for the reconfiguration of the connections and settings of circuit cells within the LSI, may also be used. This disclosure may be implemented as digital or analog processing.

[0107] Furthermore, if advancements in semiconductor technology or other derived technologies lead to the emergence of integrated circuit technologies that replace LSIs, then naturally, it would be possible to use those technologies to integrate functional blocks. The application of biotechnology, for example, is a possibility. [Industrial applicability]

[0108] This disclosure is useful for techniques that suppress inconsistencies or hallucinations between plans and generate highly reliable plans in planning using LLM. [Explanation of Symbols]

[0109] 1. Information Processing Device 11 Input section 12 Disassembly section 13. Sequencing section 14 Management Department 15 Judgment section 16. Execution Unit

Claims

1. A decomposition unit that decomposes the input task into subtasks using a first large-scale language model, An ordering unit that determines the execution order of the aforementioned subtasks using a second large-scale language model, A determination unit that determines, in accordance with the determined order, the feasibility of the actuator performing the subtask for the current state of the task object, using a third large-scale language model, Equipped with, The decomposition unit, when the first large-scale language model determines that the subtask is unexecutable, decomposes the subtask. Information processing device.

2. The system further comprises a management unit that manages the aforementioned tasks and subtasks using a tree structure, The management unit manages the decomposed subtasks at a hierarchy one level below the hierarchy that manages the subtasks before they are decomposed. The information processing apparatus according to claim 1.

3. The determination unit, The necessity of the actuator executing the subtask in relation to the current state of the task object is determined using the third large-scale language model. The feasibility of the subtasks deemed necessary to perform is determined using the third large-scale language model. The information processing apparatus according to claim 1.

4. The system further comprises a management unit that manages the aforementioned tasks and subtasks using a tree structure, The management unit determines that the subtask is completed if the third large-scale language model determines that it is not necessary to perform the subtask, or if the third large-scale language model determines that the subtask is executable. If the determination unit determines that all subtasks in the first level have been completed, it uses the third large-scale language model to determine the necessity of executing subtasks in the second level, which is one level above the first level, and uses the third large-scale language model to determine the feasibility of executing the subtasks that have been determined to have a necessity. The information processing apparatus according to claim 3.

5. If the determination unit determines that there is no need to execute the subtask, it will not perform a determination on the feasibility of executing the subtask. The information processing apparatus according to claim 3.

6. Information processing device, The input task is broken down into subtasks using the first large-scale language model. The execution order of the aforementioned subtasks is determined using a second large-scale language model. In accordance with the determined order, the feasibility of the actuators performing the subtasks for the current state of the task object is determined using a third large-scale language model. If the first large-scale language model determines that the subtask is unexecutable, the subtask is decomposed. Planning methods.

7. On the computer, The input task is broken down into subtasks using the first large-scale language model. The execution order of the aforementioned subtasks is determined using a second large-scale language model. In accordance with the determined order, the feasibility of the actuators performing the subtasks for the current state of the task object is determined using a third large-scale language model. If the first large-scale language model determines that the subtask is unexecutable, the subtask is decomposed. A program that executes a process.

Images