Robot decision reasoning method based on unified thinking map and related equipment

Abstract

The invention discloses a robot decision reasoning method based on a unified thinking map and related equipment, and relates to the technical field of robot search, and the method comprises the steps: inputting environment data and task description into a large language model, and obtaining a plurality of first candidate suggestions output by the large language model; performing mandatory criterion adjustment and non-mandatory criterion adjustment on each first candidate proposal by using an evaluator independent of the large language model to obtain a plurality of second candidate proposals; sorting the second candidate proposals according to evaluation scores of the evaluator on the second candidate proposals to obtain a sorted proposal list; and making a decision on the control of the robot by using the sorting proposal list. Each candidate proposal in the sorting proposal list not only provides clear guidance for robot decision, but also provides structured feedback for the large language model, so that continuous optimization of the reasoning ability of the large language model is promoted, and meanwhile, the situation that the large language model needs to be utilized for oriented training of a specific task is avoided.

Description

Technical Field

[0001] The present application relates to the field of robot search technology, and in particular to a robot decision-making reasoning method and related equipment based on a unified mind map. Background Art

[0002] With the recent development of large language models (LLMs), enhancing their reasoning and problem-solving capabilities has become a research hotspot. To achieve this goal, various technical frameworks have been proposed, including the Chain of Thought (CoT), Tree of Thoughts (ToT), and Diagram of Thought (DoT). These frameworks significantly improve the performance of LLMs in complex robotics tasks through structured and systematic reasoning processes.

[0003] However, existing technologies still have problems such as low efficiency and inflexible solutions. Summary of the Invention

[0004] In view of this, the embodiments of the present application provide a robot decision-making reasoning method and related equipment based on a unified mind map to improve the efficiency and flexibility of robot search decision-making reasoning.

[0005] One aspect of an embodiment of the present application provides a robot decision-making and reasoning method based on a unified mind map, the method comprising the following steps:

[0006] Inputting the environment data and the task description into the large language model to obtain a plurality of first candidate proposals output by the large language model;

[0007] Using an evaluator independent of the large language model, first performing mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals;

[0008] sorting the second candidate proposals according to the evaluation scores of the evaluators on the second candidate proposals to obtain a sorted proposal list;

[0009] The ranked proposal list is used to make decisions about controlling the robot.

[0010] In some embodiments, before inputting the environmental data and the task description into the large language model, the method further comprises the step of segmenting the entire police force in the environmental information, wherein the step of segmenting the entire police force in the environmental information comprises the following steps:

[0011] The panoramic image is evenly divided into an odd number of segments; wherein the middle segment of each segment is the segment of the current direction of the robot.

[0012] In some embodiments, inputting the environment data and the task description into the large language model to obtain a plurality of first candidate proposals output by the large language model comprises the following steps:

[0013] Input each of the segments and the task description into the large language model, and obtain the first candidate proposal corresponding to each search direction output by the large language model; wherein one segment corresponds to one search direction.

[0014] In some embodiments, the step of first performing mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals using an evaluator independent of the large language model to obtain a plurality of second candidate proposals includes the following steps:

[0015] Using the evaluator independent of the large language model to adjust the mandatory criteria of format and quantity of each of the first candidate proposals, so that the format and quantity of each of the first candidate proposals meet the requirements;

[0016] Calculate a safety penalty, a repeated exploration penalty, and a direction change penalty for each of the first candidate proposals to obtain a comprehensive penalty;

[0017] The evaluator adjusts the non-mandatory criteria of each of the first candidate proposals after the mandatory criteria are adjusted according to the comprehensive penalty to obtain each of the second candidate proposals.

[0018] In some embodiments, calculating a safety penalty, a repeated exploration penalty, and a direction change penalty for each of the first candidate proposals to obtain a comprehensive penalty comprises the following steps:

[0019] The security penalty is calculated as follows:

[0020] ;

[0021] in, For the said security penalty, is the distance from the robot to the nearest obstacle, is a predefined safety distance; if , no penalty is imposed, otherwise, as the distance from the robot to the nearest obstacle decreases, the safety penalty increases quadratically;

[0022] The repeated exploration penalty is calculated as follows:

[0023] ;

[0024] in, For the repeated exploration penalty, represents the area of ​​the total free space covered by the lidar, Indicates the overlapping area with the explored area;

[0025] The direction change penalty is calculated as follows:

[0026] ;

[0027] in, is the direction change penalty, Propose a search direction corresponding to the first candidate, is the current direction of the robot;

[0028] The combined penalty is calculated as follows:

[0029] ;

[0030] in, For the comprehensive penalty, Indicates the first candidate proposal In the sorted proposal list The original index in ; is the weight corresponding to each penalty.

[0031] In some embodiments, ranking each of the second candidate proposals according to the evaluator's evaluation score of each of the second candidate proposals to obtain a ranked proposal list includes the following steps:

[0032] using the comprehensive penalty obtained by the evaluator adjusting the non-mandatory criteria for each of the first candidate proposals as the evaluation score corresponding to the second candidate proposal;

[0033] The second candidate proposals are sorted from small to large according to the evaluation scores to obtain the sorted proposal list.

[0034] In some embodiments, using the ranked proposal list to make a decision on controlling the robot comprises the following steps:

[0035] The second candidate proposal ranked first in the ranked proposal list is selected to make a decision on the control of the robot.

[0036] Another aspect of the present application further provides a robot decision-making and reasoning device based on a unified mind map, the device comprising:

[0037] a proposal generation unit, configured to input the environment data and the task description into the large language model, and obtain a plurality of first candidate proposals output by the large language model;

[0038] a proposal optimization unit, configured to use an evaluator independent of the large language model to perform mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals;

[0039] a proposal ranking unit, configured to rank each of the second candidate proposals according to an evaluation score of the evaluator on each of the second candidate proposals, to obtain a ranked proposal list;

[0040] A control decision unit is used to make a decision on the control of the robot using the sorted proposal list.

[0041] Another aspect of the embodiments of the present application further provides an electronic device, including a processor and a memory;

[0042] The memory is used to store programs;

[0043] The processor executes the program to implement any of the above methods.

[0044] Another aspect of the embodiments of the present application further provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement any of the above methods.

[0045] This application has at least the following beneficial effects:

[0046] This application can input environmental data and task descriptions into a large language model to obtain multiple first candidate proposals output by the large language model; use an evaluator independent of the large language model to first adjust the mandatory criteria and non-mandatory criteria of each first candidate proposal to obtain multiple second candidate proposals; sort each second candidate proposal according to the evaluator's evaluation score of each second candidate proposal to obtain a ranked proposal list; use the ranked proposal list to make decisions on the control of the robot. This application uses an evaluator independent of the large language model to evaluate each first candidate proposal in parallel, and integrates the evaluation results into a ranked proposal list with corresponding scores. Each second candidate proposal in the ranked proposal list not only provides clear guidance for the robot's decision-making, but also provides structured feedback to the large language model, thereby promoting the continuous optimization of the large language model's reasoning ability, while avoiding the need to use the large language model for targeted training of specific tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0048] Figure 1 A schematic diagram of a flow chart of a robot decision-making and reasoning method based on a unified mind map provided in an embodiment of the present application;

[0049] Figure 2 An example flowchart of a robot decision-making and reasoning method based on a unified mind map provided in an embodiment of the present application;

[0050] Figure 3 An example diagram of a search task provided in an embodiment of the present application;

[0051] Figure 4 This is a structural block diagram of a robot decision-making and reasoning device based on a unified mind map provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0053] Before describing the embodiments of the present application in detail, some of the related technologies involved in the embodiments of the present application are first described as follows:

[0054] CoT is a method for enhancing the reasoning capabilities of large language models by breaking down complex problems into multiple intermediate steps, gradually guiding the model to generate the final answer. This process emphasizes clear logic and structured reasoning, enabling the model to self-evaluate at each step. Specifically, CoT first identifies the key elements of the problem and then gradually builds a chain of reasoning, ensuring that each intermediate result is closely connected to the subsequent steps. Through this systematic approach, CoT significantly improves the model's performance in complex tasks, especially in scenarios requiring multi-level logical reasoning and in-depth analysis, demonstrating higher accuracy and consistency.

[0055] By generating multiple reasoning paths and forming a tree-like structure, ToT enables the model to simultaneously explore multiple possible solutions. Unlike traditional linear reasoning, ToT allows the model to generate multiple reasoning paths in parallel when faced with complex problems and evaluate each path independently. This approach not only enhances the model's flexibility but also enables it to dynamically adjust its reasoning strategy to adapt to changing circumstances. Ultimately, the model integrates the intermediate results of each path to form a more comprehensive and accurate answer, thereby excelling in handling complex tasks and demonstrating greater creativity and in-depth analytical capabilities.

[0056] DoT is a reasoning enhancement framework for LLMs. It enhances LLM reasoning capabilities by simulating human cognitive processes and integrating multiple specialized roles within the LLM. These roles include a proposer that generates candidate proposals, a critic that evaluates and optimizes proposals, and a summarizer that integrates verified reasoning steps and generates the final output. This multi-role structure works synergistically across all reasoning stages, enabling more precise and complex logical reasoning, resulting in superior performance in complex decision-making tasks.

[0057] Taking DoT as an example, there are the following problems:

[0058] 1) Reasoning time: The iterative nature of DoT requires multiple reasoning cycles for each task, which will cause decision delays and reduce the robot's response speed, thus limiting its real-time performance.

[0059] 2) Loss of integrity: Changing the search objective or adapting to a new one often requires reconfiguring the judge’s evaluation criteria. This frequent modification of the LLM’s internal structure can undermine its integrity, impacting its flexibility during deployment and potentially leading to unpredictable consequences.

[0060] 3) Strong application coupling: In robotics-related tasks, the judge's evaluation process relies on external resources, such as maps, and requires frequent access to external memory. This dependency increases system complexity and reduces reasoning efficiency.

[0061] 4) Insufficient dynamic response: Robot application scenarios are full of uncertainty. The reasoning output of DoT is a single proposal after merging, which has a strong timeliness and is difficult to cope with time-varying dynamic environments. This hinders the robot's rapid response to dynamic changes in the external environment.

[0062] In summary, although DoT significantly improves the reasoning ability of LLM by simulating human cognitive processes, it still faces challenges in robotic applications, such as time-consuming reasoning, impaired LLM integrity, strong application coupling, and insufficient dynamic response. These problems limit the application of DoT in real-time robot decision-making.

[0063] To overcome these limitations, this application proposes an LLM-enhanced reasoning scheme suitable for real-time robot reasoning: the Diagram of Unified Thought (DoUT). DoUT is designed specifically for real-time robot decision-making. DoUT continues DoT's role division strategy, enhancing the accuracy of LLM reasoning through collaboration between different roles. However, unlike DoT, DoUT does not directly modify the LLM's internal structure. Instead, it uses the LLM as a proposer, and only the proposer can generate candidate proposals based on the task description. DoUT also introduces an independent evaluator, located outside the LLM, replacing the internal judge and summarizer roles in DoT. The evaluator performs parallel evaluations based on the candidate proposals generated by the LLM and aggregates the evaluation results into a ranked list of proposals with corresponding scores. This ranked proposal not only provides clear guidance for robot decision-making but also provides structured feedback to the LLM, thereby promoting the continuous optimization of its reasoning capabilities while avoiding targeted training for specific tasks.

[0064] Reference Figure 1 , the embodiment of the present application provides a robot decision-making reasoning method based on a unified mind map, specifically comprising the following steps S100~S130:

[0065] S100: Inputting environmental data and task description into a large language model to obtain a plurality of first candidate proposals output by the large language model;

[0066] S110: Using an evaluator independent of the large language model, perform mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals;

[0067] S120: Sort each of the second candidate proposals according to the evaluation score of each of the second candidate proposals by the evaluator to obtain a ranked proposal list;

[0068] S130: Utilize the ranked proposal list to make a decision on the control of the robot.

[0069] Optionally, before inputting the environmental data and the task description into the large language model, the method further comprises the step of segmenting the entire police force in the environmental information, wherein the step of segmenting the entire police force in the environmental information comprises the following steps:

[0070] The panoramic image is evenly divided into an odd number of segments; wherein the middle segment of each segment is the segment of the current direction of the robot.

[0071] Optionally, inputting the environment data and the task description into the large language model to obtain a plurality of first candidate proposals output by the large language model comprises the following steps:

[0072] Input each of the segments and the task description into the large language model, and obtain the first candidate proposal corresponding to each search direction output by the large language model; wherein one segment corresponds to one search direction.

[0073] Optionally, the step of using an evaluator independent of the large language model to first perform mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals comprises the following steps:

[0074] Using the evaluator independent of the large language model to adjust the mandatory criteria of format and quantity of each of the first candidate proposals, so that the format and quantity of each of the first candidate proposals meet the requirements;

[0075] Calculate a safety penalty, a repeated exploration penalty, and a direction change penalty for each of the first candidate proposals to obtain a comprehensive penalty;

[0076] The evaluator adjusts the non-mandatory criteria of each of the first candidate proposals after the mandatory criteria are adjusted according to the comprehensive penalty to obtain each of the second candidate proposals.

[0077] Optionally, calculating a safety penalty, a repeated exploration penalty, and a direction change penalty for each of the first candidate proposals to obtain a comprehensive penalty comprises the following steps:

[0078] The security penalty is calculated as follows:

[0079] ;

[0080] in, For the said security penalty, is the distance from the robot to the nearest obstacle, is a predefined safety distance; if , no penalty is imposed, otherwise, as the distance from the robot to the nearest obstacle decreases, the safety penalty increases quadratically;

[0081] The repeated exploration penalty is calculated as follows:

[0082] ;

[0083] in, For the repeated exploration penalty, represents the area of ​​the total free space covered by the lidar, Indicates the overlapping area with the explored area;

[0084] The direction change penalty is calculated as follows:

[0085] ;

[0086] in, is the direction change penalty, Propose a search direction corresponding to the first candidate, is the current direction of the robot;

[0087] The combined penalty is calculated as follows:

[0088] ;

[0089] in, For the comprehensive penalty, Indicates the first candidate proposal In the sorted proposal list The original index in ; is the weight corresponding to each penalty.

[0090] Optionally, sorting the second candidate proposals according to the evaluator's evaluation scores of the second candidate proposals to obtain a sorted proposal list comprises the following steps:

[0091] using the comprehensive penalty obtained by the evaluator adjusting the non-mandatory criteria for each of the first candidate proposals as the evaluation score corresponding to the second candidate proposal;

[0092] The second candidate proposals are sorted from small to large according to the evaluation scores to obtain the sorted proposal list.

[0093] Optionally, the making a decision on controlling the robot using the ranked proposal list comprises the following steps:

[0094] The second candidate proposal ranked first in the ranked proposal list is selected to make a decision on the control of the robot.

[0095] Next, the solution of the embodiment of the present application will be introduced and explained in detail with reference to specific application examples.

[0096] Reference Figure 2 This embodiment provides an example flowchart of a robot decision-making and reasoning method based on a unified mind map. The DoUT in this embodiment integrates unified evaluation criteria through an external evaluator, ranks candidate LLM propositions based on mandatory and advisory criteria, and provides structured feedback to improve decision-making and reasoning efficiency. DoUT utilizes parallel decision-making to accelerate the reasoning process and avoid iterative reasoning, making it suitable for real-time robot decision-making tasks.

[0097] Specifically, this embodiment, called DoUT, provides an LLM reasoning enhancement solution for real-time robotic applications. By simulating human cognitive processes, it introduces two specialized roles within the LLM: a collaborative proposer and an independent evaluator. This enhances the LLM's reasoning capabilities. The LLM, acting as the proposer, receives multimodal inputs from the robot's sensors, including images, lidar, text, and voice descriptions. Leveraging its powerful reasoning and comprehension capabilities, the LLM reasoned and summarized these multimodal sensor inputs, combined with the robot's task requirements, and outputs a list of candidate proposals.

[0098] The evaluator then evaluates each proposal in this list of candidate proposals, primarily using a comprehensive evaluation of the candidate proposals using mandatory criteria and advisory criteria. Mandatory criteria include specifications such as proposal format and anomaly detection. Any candidate proposal that violates the mandatory criteria is immediately flagged, and feedback is provided to the proposer to enforce necessary changes. Only proposals that meet these criteria proceed to further evaluation using the advisory criteria. Mandatory criteria ensure that even without any task-specific pre-training, the LLM can autonomously adjust its output to meet the robot's decision-making requirements based on the evaluator's feedback, thereby enhancing the system's robustness. Advisory criteria focus on ranking the candidate proposals, typically including factors such as safety, efficiency, and task-specific objectives. By ranking the candidate proposals derived by the proposer based on these advisory criteria and formatting feedback, the proposer (i.e., the LLM) is implicitly guided to learn the evaluator's evaluation criteria, thereby self-learning and improving its reasoning accuracy. Over time, this process can significantly improve the LLM's reasoning accuracy and adaptability.

[0099] The beneficial effects of the DoUT of this embodiment include:

[0100] 1) Integrity: By centralizing the evaluation process to run independently outside the LLM, the modification of the LLM internal evaluation standards is avoided, thus maintaining its integrity.

[0101] 2) Decoupling: Evaluators have centralized and unified access to external storage, eliminating direct LLM access to external storage. This design decouples the system, reduces complexity, and mitigates the risk of unauthorized or inconsistent storage access.

[0102] 3) Parallel Acceleration: DoUT replaces iterative judgment with parallel evaluation, significantly reducing the number of reasoning cycles, thereby improving decision-making speed and accelerating the robot's responsiveness.

[0103] 4) Flexible output: Through the scoring mechanism, the evaluator provides a ranked list of candidate proposals, enabling the robot to dynamically adjust its decision-making based on changes in the actual environment.

[0104] The DoUT of this embodiment provides a simplified and adaptive solution, which addresses the limitations of the existing DoT in real-time response reasoning and enhances the efficiency, stability and adaptability of LLM in robotic applications.

[0105] More specifically, this embodiment includes the following technical solutions:

[0106] In the search task in a large-scale unknown environment, combined with the solution of this embodiment, the specific example is as follows:

[0107] For example, Figure 3 This is an example diagram of the search task of this embodiment. Figure 3 (a) and Figure 3 (b) in the figure presents the task description and the environment panorama input respectively. Figure 3 (c1–c3) in the figure show the candidate proposals generated by the proposer. Figure 3 (d1–d3) in is the proposer’s feedback on the corresponding proposal. Figure 3 (d1–d2) in is the feedback for violating the mandatory criteria and guides the proposer to make corrections. Figure 3 (d3) provides optimization suggestions to the proposer based on non-mandatory criteria.

[0108] At the moment , the robot is located at position , the onboard panoramic camera captures the panoramic view of the environment .like Figure 3 As shown in (b) in the figure, the panorama is divided into multiple segments, denoted as , each segment corresponds to a possible movement direction of the robot. This division method not only facilitates LLM reasoning, but also enhances the interpretability of the decision process. and The intersection of the robot and the surrounding images will cause the robot to frequently change its direction and reduce the search efficiency. The GET framework will Divided into fragments, of which , ensuring that the robot's current direction is always in the middle segment In addition, since the environment panorama is captured in the robot's camera coordinate system and the middle segment of the panorama segmentation Located directly in front of the robot's movement direction, the LLM can reason directly based on the current state without complex coordinate transformations. This intuitive input format improves the LLM's reasoning accuracy and simplifies the decision-making process.

[0109] In the reasoning search process of the GET framework, the task description ( Figure 3 (a) in the figure) and a panoramic view of the environment ( Figure 3 (b) in the figure is passed to DoUT as input. LLM, as the proposer, uses its powerful semantic understanding and reasoning capabilities to analyze the panoramic image and combine it with the task description to infer the direction that is most likely to contain the search object, and finally outputs candidate proposals. .

[0110] However, at the start of the task, the proposer in DoUT has not been pre-trained for the search task, and its reasoning accuracy in the search task needs to be improved. The output candidate proposals may not meet the specific proposal format requirements of the evaluator during automatic evaluation. In addition, the panoramic image of the environment is captured using an equirectangular projection, which introduces projection distortion and makes it difficult for the LLM to intuitively interpret the distance to surrounding obstacles. Therefore, GET adopts the following mandatory criteria to guide the proposer to generate proposals that meet the format requirements. The non-mandatory criteria are used to guide the proposer to reason about accurate candidate proposals and correct some inappropriate decisions.

[0111] In conjunction with the search task in a large-scale unknown environment, the following mandatory and non-mandatory criteria can be used:

[0112] 1) Mandatory guidelines:

[0113] a) Proposal Format:

[0114] At the beginning of the task execution, the proposer may not be able to generate a proposal that meets the format required by the evaluator. For example, in GET, the evaluator needs the format The formatted proposal list is a list of proposals, so that the evaluator can extract key information and make correct evaluations through regular expression matching. However, the proposals given by the proposer may not meet the format requirements, such as Figure 3 In this case, the evaluator will directly provide feedback specifying the expected format to guide the proposer to generate proposals that meet the format, such as Figure 3 This process is repeated until the proposal conforms to the desired structure.

[0115] b) Proposed quantity:

[0116] To ensure that the sorted proposal list output by DoUT contains enough proposals for the robot to make decisions, the proposer infers all the fragments in the input panorama and sorts them according to the probability of finding the search target. This also means that the number of candidate proposals obtained by inference must be the same as the number of fragments in the panorama segmentation. Consistent, if the proposer generates a candidate proposal list The number of proposals and If there is no match, the evaluator will directly reject the evaluation and provide corrective feedback, e.g. Figure 3 (c2) and Figure 3 As shown in (d2) in .

[0117] Once all mandatory criteria are met, the evaluator will evaluate and rank the candidate proposals based on the following non-mandatory criteria to generate a ranked list of proposals: .

[0118] 2) Non-mandatory guidelines:

[0119] a) Security Penalties:

[0120] This criterion allows the evaluator to access external map storage and combine environmental information to impose penalties on proposals that guide the robot to approach obstacles. The real-time point cloud of the sensor is divided into regions, each region uniquely corresponds to a candidate proposal in the environment panorama For each area, the distance to the nearest obstacle is , candidate Safety penalties for:

[0121] ;(1)

[0122] in is a predefined safety distance. If , no penalty is imposed. Otherwise, the penalty grows quadratically as the distance to the nearest obstacle decreases.

[0123] b) Repeated exploration penalty:

[0124] This criterion aims to penalize proposals that lead the robot to areas that have already been explored. Based on this, by calculating the Penalty is calculated based on the overlap ratio between the covered area and the explored area :

[0125] ;(2)

[0126] in, represents the area of ​​the total free space covered by the lidar, Represents the area of ​​overlap with the explored area.

[0127] c) Direction change penalty:

[0128] This criterion is used to penalize proposals that will cause the robot to change its current direction of movement and make large angle turns. Corresponding direction and the current robot's moving direction Penalize the candidate proposal to avoid frequent direction changes during the robot search, which may lead to low efficiency or high energy consumption. Punishment for:

[0129] ;(3)

[0130] d) Comprehensive Penalty:

[0131] Based on the above points, the evaluator proposes Comprehensive punishment The calculation is as follows:

[0132] ;(4)

[0133] in express In the candidate proposal list generated by the proposer The original index in , encourages prioritization based on its initial order. It is the weight to balance the influence of each criterion.

[0134] According to formula (4), each candidate proposal The evaluator evaluates each candidate proposal based on its score right Sort in ascending order to generate a sorted list of proposals and their corresponding score lists . These ranked proposal lists Not only as a decision basis for robot search, but also as a corresponding score list The proposed information is fed back to the proposer to promote LLM to conduct targeted learning for the search task and improve the reasoning accuracy.

[0135] In summary, this embodiment has the following technical effects:

[0136] 1. DoUT achieves decoupling and parallelization of the reasoning process by using a large language model (LLM) as a proposer and introducing an external independently running evaluator.

[0137] 2. Evaluators use mandatory criteria (such as proposal format, anomaly detection) and non-mandatory criteria (such as security, efficiency, and task-specific goals) to comprehensively evaluate candidate proposals.

[0138] 3. Generate a ranked list of proposals through a scoring mechanism, enabling the robot to flexibly adjust its decisions according to dynamic changes in the environment.

[0139] 4. Through the feedback from the evaluators, LLM is encouraged to conduct self-learning and optimization, gradually improving the reasoning accuracy and adaptability.

[0140] In summary, this embodiment includes the following technical solutions:

[0141] 1. The overall architectural design of DoUT, including the division of roles between proposers and evaluators, the parallel evaluation mechanism, and the feedback optimization mechanism.

[0142] 2. By constructing evaluation criteria, parallel evaluation, and feedback mechanisms, we guide LLM to generate proposals that meet the robot's decision-making needs.

[0143] 3. The protection feedback mechanism is designed to realize the self-learning of LLM.

[0144] The beneficial effects of this embodiment include:

[0145] a) Improving the robot's real-time decision-making capabilities: The Unified Map of Thought (DoUT) significantly reduces the number of reasoning cycles by introducing external evaluators and adopting a parallel evaluation mechanism, thereby significantly improving the robot's response speed and real-time decision-making capabilities. This design enables the robot to quickly adjust its decisions in dynamic environments and adapt to the needs of complex scenarios.

[0146] b) Maintaining the integrity of the Large Language Model (LLM): DoUT centralizes the evaluation process outside the LLM and runs it independently, avoiding modifications to the LLM's internal structure and thus maintaining its integrity and stability. This design not only enhances the robustness of the system but also reduces the risk of unpredictability caused by frequent modifications to the LLM's internal structure.

[0147] c) Enhanced system decoupling and flexibility: By unifying access to external storage and preventing LLMs from directly accessing external resources, DoUT achieves system decoupling and reduces complexity. This design not only improves inference efficiency but also reduces the risk of unauthorized or inconsistent storage access, enhancing system flexibility and scalability.

[0148] d) Optimizing Reasoning Accuracy and Adaptability: The DoUT comprehensively evaluates candidate proposals using both mandatory and non-mandatory criteria to ensure they meet the robot's decision-making requirements. Simultaneously, the evaluation results are fed back to the LLM, prompting it to engage in self-learning and optimization, thereby gradually improving reasoning accuracy and adaptability. This mechanism enables the LLM to quickly adapt to new tasks without task-specific pre-training.

[0149] e) Flexible output in dynamic environments: DoUT generates a ranked list of proposals through a scoring mechanism, enabling the robot to flexibly adjust its decisions based on dynamic changes in the actual environment. This design not only improves the adaptability of its decisions but also enhances the robot's ability to perform tasks in complex environments.

[0150] f) Simplifying the Decision-Making Process and Enhancing Interpretability: DoUT simplifies the reasoning process of LLM and enhances the interpretability of the decision-making process through intuitive input formats (such as panorama partitioning) and a structured feedback mechanism. This design enables users to better understand the model's reasoning logic, thereby increasing their trust in the system's generated results.

[0151] Reference Figure 4 , the embodiment of the present application provides a robot decision-making and reasoning device based on a unified mind map, including:

[0152] a proposal generation unit, configured to input the environment data and the task description into the large language model, and obtain a plurality of first candidate proposals output by the large language model;

[0153] a proposal optimization unit, configured to use an evaluator independent of the large language model to perform mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals;

[0154] a proposal ranking unit, configured to rank each of the second candidate proposals according to an evaluation score of the evaluator on each of the second candidate proposals, to obtain a ranked proposal list;

[0155] A control decision unit is used to make a decision on the control of the robot using the sorted proposal list.

[0156] It can be understood that the contents of the above method embodiments are all applicable to the present device embodiments, the functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0157] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flow chart of the present application are provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0158] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present application. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the routine skills of an engineer. Therefore, a person skilled in the art can implement the present application as set forth in the claims using ordinary techniques without undue experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0159] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0160] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0161] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0162] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0163] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0164] Although the embodiments of the present application have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and intent of the present application, and that the scope of the present application is defined by the claims and their equivalents.

[0165] The above is a specific description of the preferred implementation of the present application, but the present application is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present application, and these equivalent modifications or substitutions are all included in the scope defined by the claims of the present application.

Claims

1. A robot decision-making reasoning method based on a unified mind map, characterized in that: The method comprises the following steps: Inputting the environment data and the task description into the large language model to obtain a plurality of first candidate proposals output by the large language model; Using an evaluator independent of the large language model, first performing mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals; sorting the second candidate proposals according to the evaluation scores of the evaluators on the second candidate proposals to obtain a sorted proposal list; The ranked proposal list is used to make decisions about controlling the robot.

2. The robot decision-making and reasoning method based on a unified mind map according to claim 1 is characterized in that: Before inputting the environmental data and the task description into the large language model, the method further includes the step of segmenting the entire police force in the environmental information, wherein the step of segmenting the entire police force in the environmental information includes the following steps: The panoramic image is evenly divided into an odd number of segments; wherein the middle segment of each segment is the segment of the current direction of the robot.

3. The robot decision-making and reasoning method based on a unified mind map according to claim 2 is characterized in that: Inputting the environment data and the task description into the large language model to obtain a plurality of first candidate proposals output by the large language model includes the following steps: Input each of the segments and the task description into the large language model, and obtain the first candidate proposal corresponding to each search direction output by the large language model; wherein one segment corresponds to one search direction.

4. The robot decision-making and reasoning method based on a unified mind map according to claim 1 is characterized in that: The method of using an evaluator independent of the large language model to first perform mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals includes the following steps: Using the evaluator independent of the large language model, adjusting the mandatory criteria of format and quantity of each of the first candidate proposals so that the format and quantity of each of the first candidate proposals meet the requirements; Calculate a safety penalty, a repeated exploration penalty, and a direction change penalty for each of the first candidate proposals to obtain a comprehensive penalty; The evaluator adjusts the non-mandatory criteria of each of the first candidate proposals after the mandatory criteria are adjusted according to the comprehensive penalty to obtain each of the second candidate proposals.

5. The robot decision-making and reasoning method based on a unified mind map according to claim 4 is characterized in that: The step of calculating a safety penalty, a repeated exploration penalty, and a direction change penalty for each of the first candidate proposals to obtain a comprehensive penalty includes the following steps: The security penalty is calculated as follows: ； in, For the said security penalty, is the distance from the robot to the nearest obstacle, is a predefined safety distance; if , no penalty is imposed, otherwise, as the distance from the robot to the nearest obstacle decreases, the safety penalty increases quadratically; The repeated exploration penalty is calculated as follows: ； in, For the repeated exploration penalty, represents the area of ​​the total free space covered by the lidar, Indicates the overlapping area with the explored area; The direction change penalty is calculated as follows: ； in, is the direction change penalty, Propose a search direction corresponding to the first candidate, is the current direction of the robot; The combined penalty is calculated as follows: ； in, For the comprehensive penalty, Indicates the first candidate proposal In the sorted proposal list The original index in ; is the weight corresponding to each penalty.

6. The robot decision-making and reasoning method based on a unified mind map according to claim 1 is characterized in that: Sorting the second candidate proposals according to the evaluator's evaluation scores of the second candidate proposals to obtain a ranked proposal list includes the following steps: using the comprehensive penalty obtained by the evaluator adjusting the non-mandatory criteria for each of the first candidate proposals as the evaluation score corresponding to the second candidate proposal; The second candidate proposals are sorted from small to large according to the evaluation scores to obtain the sorted proposal list.

7. The robot decision-making and reasoning method based on a unified mind map according to claim 6 is characterized in that: The step of making a decision on controlling the robot using the ranked proposal list comprises the following steps: The second candidate proposal ranked first in the ranked proposal list is selected to make a decision on controlling the robot.

8. A robot decision-making and reasoning device based on a unified mind map, characterized in that: The device comprises: a proposal generation unit, configured to input the environment data and the task description into the large language model, and obtain a plurality of first candidate proposals output by the large language model; a proposal optimization unit, configured to use an evaluator independent of the large language model to perform mandatory criteria adjustment and non-mandatory criteria adjustment on each of the first candidate proposals to obtain a plurality of second candidate proposals; a proposal ranking unit, configured to rank each of the second candidate proposals according to an evaluation score of the evaluator on each of the second candidate proposals, to obtain a ranked proposal list; A control decision unit is used to make a decision on the control of the robot using the sorted proposal list.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The storage medium stores a program, and the program is executed by a processor to implement the method according to any one of claims 1 to 7.

Images