A Method and System for Robot Task Planning and Anomaly Correction Based on Common Temporal Sequence

Abstract

This invention proposes a robot task planning and anomaly correction method and system based on common temporal sequence, belonging to the field of robot task planning technology. It includes: inputting user task instructions into a large language model to generate multiple candidate item sequences and constructing a common temporal sequence task graph with items as nodes and item operation order as edges; matching and pruning the task graph according to the real-world scenario state, and calculating the optimal item operation path based on a value function; inputting the user task instructions, the real-world scenario state, and the optimal item operation path into the large language model, and obtaining an executable fine-grained task sequence through task instruction penalty term correction to drive robot execution; when the robot encounters an anomaly, introducing a fine-grained correction mechanism, adjusting the task sequence through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms to select the branch path with the highest overall value to continue execution. This invention achieves high efficiency and execution stability in robot task planning.

Description

Technical Field

[0001] This invention relates to the field of robot task planning technology, and in particular to a robot task planning and anomaly correction method and system based on common timing. Background Technology

[0002] As a core component of smart homes, home service robots are expanding from simple tasks like cleaning and moving to more complex services such as delivering water and retrieving items from different rooms, becoming important intelligent terminals for enhancing home convenience. With the maturation of the smart home ecosystem, users are increasingly demanding higher levels of complexity and reliability from these robots. Achieving efficient and stable execution in the dynamic and open home environment has become a core challenge for the industry.

[0003] The existing task planning framework has obvious shortcomings: traditional planning methods rely on pre-set environment models, which are prone to planning anomalies when faced with uncertain situations such as missing items or changes in location. Even if the model is reconstructed, it is difficult to adapt to dynamic home environments due to poor real-time performance and high costs. Although large language models can understand complex tasks and generate plans, their plans are easily out of touch with the actual scenario, resulting in poor feasibility of the solutions. Moreover, after failure, they often adopt a complete replanning and lack a systematic anomaly handling mechanism, which leads to insufficient execution stability and low efficiency. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a robot task planning and anomaly correction method and system based on common temporal sequences, achieving high efficiency and execution stability in robot task planning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a robot task planning and anomaly correction method based on common temporal sequences, comprising: The user's task instructions are input into the large language model to generate multiple candidate item sequences; a common temporal task graph is constructed based on the candidate item sequences, with items as nodes and the order of item operations as edges; The task graph is matched and pruned according to the real scene state, and the optimal item operation path is calculated based on the value function on the pruned task graph. The user task instructions, real-world scenario state, and optimal item operation path are input into the large language model, and an executable fine-grained task sequence is obtained by correcting the task instruction penalty term and driving the robot to execute it. When the robot malfunctions, a fine-grained correction mechanism is introduced. This mechanism adjusts the task sequence through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms to select the branch path with the highest overall value to continue execution.

[0006] Secondly, the present invention provides a robot task planning and anomaly correction system based on common temporal sequence, comprising: The task graph construction module is configured to input user task instructions into a large language model and generate multiple candidate item sequences; and to construct a common temporal task graph based on the candidate item sequences, with items as nodes and the order of item operations as edges. The pruning module is configured to match and prune the task graph according to the real scene state, and calculate the optimal item operation path on the pruned task graph based on the value function. The fine-grained task generation module is configured to input the user task instructions, real scene state and optimal item operation path into the large language model, and obtain an executable fine-grained task sequence through the correction of task instruction penalty terms and drive the robot to execute it. The correction module is configured to introduce a fine-grained correction mechanism when the robot encounters an anomaly. This mechanism adjusts the task sequence through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms to select the branch path with the highest overall value to continue execution.

[0007] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the robot task planning and anomaly correction method based on common timing described in the first aspect.

[0008] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the robot task planning and anomaly correction method based on common timing described in the first aspect.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention generates multiple candidate item sequences using a large language model to construct a common temporal task graph. This retains the mainstream logic of task execution while providing the robot with a rich array of alternative operation paths, avoiding the limitations of single planning. By pruning the task graph in conjunction with real-world scenario states, redundant paths that are incompatible with the environment can be effectively eliminated. The optimal operation path is then calculated using a value function, improving the adaptability of task planning to real-world scenarios and reducing the probability of unexecutable plans. In the anomaly handling phase, a fine-grained correction mechanism is introduced. The task sequence is adjusted through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms. When execution anomalies occur, local alternative solutions are used for rapid correction. Backtracking only occurs when alternative solutions are exhausted. When neither alternative selection strategies nor backtracking mechanisms can find a feasible solution, the large language model is triggered to update the path again. This avoids the drawbacks of traditional methods that require a complete regeneration of the plan, significantly improving the stability and task completion efficiency of robot execution, while ensuring the robustness and continuity of task execution in complex home scenarios.

[0010] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0012] Figure 1 The main flowchart of a robot task planning and anomaly correction method based on common timing provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a robot task planning and anomaly correction method based on common timing provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the candidate item sequence generation and normalization process provided in an embodiment of the present invention; Figure 4 Common timing task diagrams provided for embodiments of the present invention; Figure 5 A schematic diagram of optimal item path selection based on a value function provided for an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating online execution and exception recovery provided in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] As mentioned in the background section, home service robots are playing an increasingly important role in modern family life, becoming essential intelligent devices to assist humans in handling daily tasks. With the increasing maturity of smart home systems, the tasks undertaken by robots have expanded from simple cleaning and moving to more complex multi-step tasks, such as retrieving items across rooms, delivering water, and organizing items.

[0015] The increasing demand for tasks has created a promising future for service robots. However, existing task planning frameworks still generally suffer from problems such as difficulty in implementing plans, weak anomaly recovery capabilities, and high online computing costs when faced with service task execution.

[0016] On the one hand, from a task planning perspective, traditional rule-based or symbol-based planning methods typically require the pre-construction of a relatively complete environmental model, action set, and constraints. The open environment referred to here is a dynamic, unstructured environment in a home setting where item positions change randomly, items are missing, states are uncertain, and interfering factors are unpredictable and cannot be exhaustively exhaustively identified in advance. While traditional methods can address local changes through remodeling, each environmental change requires manual or semi-automatic model updates and redefinition of constraints and action rules. This not only results in delayed response and low efficiency but also fails to cover the infinitely diverse unknown scenarios in open environments. Planning will still fail if new situations arise that the model does not cover. Therefore, when facing frequent and irregular dynamic changes in the home environment, abnormal planning is easily generated, with extremely poor overall generalization ability and real-time adaptability, making it difficult to meet the requirements of stable, reliable, and continuous operation in open environments.

[0017] On the other hand, while large language models demonstrate strong capabilities in task understanding and planning generation, task plans directly output by these models often lack strict matching with real-world scenarios, easily generating unexecutable or costly plans. Furthermore, when execution fails, there's no way to fine-tune the original plan or handle specific anomalies; the entire plan must be regenerated, lacking a systematic anomaly solution, leading to unstable execution and low task completion efficiency. In addition, frequent calls to large language models during online execution incur high computational overhead and latency, making it difficult to meet the real-time requirements of resource-constrained platforms.

[0018] In view of this, the present invention proposes a robot task planning and anomaly correction method, system, medium and device based on common temporal sequence. Through the construction of common temporal sequence task graph, optimal path generation and fine-grained correction mechanism, the robot can generate executable optimal task plan under real environmental constraints and quickly resume execution in abnormal situations. At the same time, it reduces the frequency of large language model calls and system costs, and improves the efficiency and execution stability of robot task planning.

[0019] Example 1 like Figure 1 As shown, this embodiment discloses a robot task planning and anomaly correction method based on common temporal sequences, including the following steps: S1: Input the user task instructions into the large language model to generate multiple candidate item sequences; construct a common temporal task graph based on the candidate item sequences, with items as nodes and the order of item operations as edges; S2: Match and prune the task graph according to the real scene state, and calculate the optimal item operation path on the pruned task graph based on the value function; S3: Input the user task instructions, real scene state and optimal item operation path into the large language model, and obtain an executable fine-grained task sequence through the correction of the task instruction penalty term and drive the robot to execute it. S4: When the robot encounters an anomaly, a fine-grained correction mechanism is introduced. The task sequence is adjusted through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms to select the branch path with the highest overall value to continue execution.

[0020] Next, combined Figure 2 This embodiment provides a detailed description of a robot task planning and anomaly correction method based on common timing.

[0021] (I) Constructing a common time-series task graph The system retrieves the user task instruction I and uses a large language model to generate K candidate item sequences based on it. Each sequence consists of several items required to complete the task, arranged in the order of operation, representing the core items and optional item sets needed to complete the task. After the candidate item sequences are generated, the item names are normalized and standardized using synonyms to merge similar items and unify node representations.

[0022] For example, such as Figure 3 As shown, based on the user task instruction I "deliver a glass of water to the bedroom", the large language model is invoked to generate K=10 candidate item sequences, including "Candidate Sequence 1: water glass, glasses, bedroom", "Candidate Sequence 2: coffee cup, nearsighted glasses, bedroom", "Candidate Sequence 3: glass, control mirror, bedroom", etc. The item names in the sequences are normalized, for example, "cup" is unified as "water glass", and "bedroom" is unified as "bedroom", etc.

[0023] Construct a directed item task graph based on the K candidate item sequences. , where node V is the set of items and edge E represents the order of operations between items.

[0024] In this embodiment, a multi-candidate item sequence is generated by a large language model and a common temporal task graph is constructed to realize a structured representation of candidate solutions, which facilitates unified evaluation and selection.

[0025] The term "common temporal sequence" refers to the process of extracting common components from multiple candidate task sequences to construct a task graph. Specifically, "common temporal sequence" describes identifying shared components with common sequential relationships among multiple item operation sequences. These common components are extracted to form a unified temporal task graph. By comparing and summarizing multiple candidate item sequences, a graph structure based on item operation order and dependencies is formed.

[0026] For example, there are multiple ways to complete the same task, and the items involved in these methods may overlap. These overlapping items are called "common" items. "Sequence" refers to the order in which items are used, reflecting a certain logical relationship during task execution.

[0027] Specifically, the task graph is pruned by introducing real-world scenario states S (including item existence, reachability, availability, etc.), removing unusable item nodes and sequence edges that do not meet constraints, and cleaning up the connectivity of the graph to obtain the pruned task graph. .

[0028] First, construct the item-task graph G based on the candidate sequences. Obtain the availability a(v) of each item from the environment state S. Set a pruning threshold. =0.2, delete a(v)< The nodes and related edges are identified; simultaneously, unreasonable order edges are removed according to hard constraint rules (e.g., the order related to "water / heating" cannot be executed before obtaining the container; if the instruction contains a mandatory target item, a feasible path containing that item must be retained). Finally, connectivity cleanup is performed to obtain the pruned graph. .

[0029] The existence of an item refers to determining whether an item exists in the environment using sensors and an environmental perception system. If an item is missing, the value of its node a(v) is set to 0, and the corresponding node in the task graph is pruned. Assuming the task is to "find a container" and "fetch water," the steps are as follows... Figure 4 As shown, the task steps are: Step 1: Find the container → Step 2: Find the water source → Step 3: Take the water → Step 4: Place the container. Assuming the container is a necessary task node, the system uses an environmental awareness system to determine if the container is in the environment. If the container is missing, the system will output a(v)=0 and delete the "container" node from the task graph. If the container does not exist, the "Find Container" node in the task graph will be pruned, and the system will be unable to continue searching for or taking water.

[0030] Item accessibility refers to assessing whether an item is reachable based on the robot's current location and navigation system. If an item is located in an inaccessible area, its accessibility a(v) value will be set to low, potentially leading to pruning. Assuming the task is "find a desk and place an item," the task graph steps are: Step 1: Find a desk → Step 2: Place item. The system assesses the desk's accessibility using the robot's navigation system. If the desk is in an area blocked by obstacles, preventing the robot from reaching it, the desk node's accessibility a(v) will be set to low, indicating inaccessibility. If the desk is inaccessible, even if it exists, the "Find a desk" node in the task graph will still be marked as having low accessibility, potentially leading to pruning.

[0031] Item availability refers to whether an item is in an operable state and can be used to perform a task. For example, if an item is damaged or unsuitable for the current task, its availability a(v) will decrease, thus affecting the pruning of the task graph. Suppose the task is "find a cup" and "pour water," with task steps 1: find the cup → step 2: pour water → step 3: place it. The system uses an environmental awareness system to determine the cup's status. If the cup is damaged, its availability will decrease, and it may become unusable for pouring water. Assume that the cup's availability a(v) drops to a very small value, close to 0, and prune the operation steps related to the cup based on the task graph's evaluation results.

[0032] Furthermore, unreasonable edges are removed according to hard constraint rules. For example, if a task instruction requires "no operations related to 'water / heating' can be performed before a container is obtained," then any edges involving that item will be removed if the container node has not yet been obtained. If the instruction contains a mandatory target item, then feasible paths containing that item must be preserved, and edges involving that item cannot be pruned.

[0033] Finally, to ensure the effectiveness of the task graph, connectivity cleanup is also necessary. Connectivity cleanup ensures that all remaining item nodes and edges remain effectively connected in the task graph, removing those that have become isolated or are unreachable from the target. After cleanup, the resulting pruned task graph is obtained. It is simpler, conforms to environmental and constraint conditions, and also ensures the feasibility of task planning.

[0034] In this embodiment, real-world scenario states are introduced, and the common time series graph is matched and pruned based on item existence, item reachability, item availability, hard constraint rules, and connectivity cleanup, thereby reducing the probability of unexecutable planning and improving the consistency between planning results and environmental constraints.

[0035] (ii) Optimal path generation Task map after pruning Constructing a value function The calculation formula is: ; Among them, nodes For an item, m represents the total number of item nodes. Represents item nodes Contribution to completing the instruction, This represents the probability of successfully obtaining / using the item in scenario S. This represents the cost of performing operations related to this item. It is the order edge weight, representing co-occurrence frequency, switching cost, etc.

[0036] like Figure 5 As shown, based on the task graph The value function is used to evaluate the candidate paths and calculate the optimal item operation path with the highest value. The calculation formula is as follows: ; in, It is the set of feasible paths from the start node START to the target node GOAL. The value function comprehensively considers factors such as the contribution of items to task completion, execution cost, and success probability.

[0037] In this embodiment, a value function is designed to select the optimal item operation path, so that the planning achieves a better balance between feasibility, cost and task contribution.

[0038] (III) Fine-grained correction mechanism The large language model is invoked based on the task instruction I, the scene state S, and the optimal item operation path. The formula for generating an action-level fine-grained task sequence T is: ; Task instructions I often contain ambiguity (e.g., "take a cup" can be interpreted as pointing to any cup) or are inconsistent with the context (e.g., the pointed-to item does not exist), causing LLM to misunderstand the intent of the instruction. Therefore, it is necessary to preprocess the input task instructions I to standardize synonyms or ambiguous descriptions in the natural language instructions.

[0039] Specifically, a semantic ambiguity penalty term S, a vagueness penalty term F, and a contextual inconsistency penalty term C are introduced.

[0040] The semantic ambiguity penalty S addresses the polysemy of words in the instruction; for example, "take a cup" could be interpreted as taking any cup, rather than a specific cup. The ambiguity penalty F addresses the multiple interpretations of the step description; for example, "grab a cup" could refer to multiple cups. The contextual inconsistency penalty C addresses contradictions between the step and the scene state, such as when the mentioned item cannot be found in the scene.

[0041] Define LLM generation steps The initial probability distribution of the action sequence is The probability after correction based on the penalty term is: ; in, , , These are the weighting coefficients. The larger the penalty, the greater the weighting of the exponent. The smaller the value, the lower the probability, making the model more inclined to choose clear and consistent task steps.

[0042] If the corrected probability If the value exceeds the threshold, the robot is driven to execute the command; otherwise, the item is matched to the item ID in the actual scene through semantic analysis mapping, or the target is confirmed with the user, and the probability is recalculated after receiving a clear instruction.

[0043] To further reduce semantic ambiguity during the robot's task execution, this embodiment also incorporates fine-grained task correction based on the scene context.

[0044] For example, will Input LLM and output action-level sequence T, such as generating executable steps in the form of "Navigate → Open → Grab → Place → Deliver", and aligning them with the preset skill interface (Navigate, Pick, Place, Open, Close, Deliver).

[0045] like Figure 6 As shown, during the task planning and execution process, when a subtask cannot be completed, the system introduces a fine-grained correction mechanism that combines alternative selection, backtracking, and cascading triggering decisions to ensure the continuity of task execution and overall efficiency.

[0046] 1. Alternative Selection Strategy First, when a local task fails, the system uses an alternative-first strategy for rapid recovery.

[0047] When the item or operation corresponding to the current task node is unavailable, the system first performs a deeper matching based on the semantic awareness network, which not only considers the basic attributes of the item and the operation requirements, but also analyzes the specific reasons for the task failure.

[0048] Specifically, when the items or operations required for the current task node are unavailable, the system first analyzes the reasons for the failure (e.g., items do not exist, path is blocked, function is incompatible, etc.), and then makes alternative selections based on the functional semantic library. The functional semantic library pre-classifies and models the functional attributes of items and paths (e.g., "container class", "support class", "movement path class"), and calculates the functional similarity between entities of the same class.

[0049] The replacement type is determined based on the reason for failure. If an item is missing, the most similar replacement item is found; if the path is unreachable, an equivalent alternative path is found, such as another navigation path to the target area.

[0050] Next, a pre-trained semantic model (such as BERT) is used to encode keywords, item descriptions, or operation steps in the task instructions into vectors, and cosine similarity or Euclidean distance is calculated between these vectors and the semantic vectors of candidate items / paths. Combining functional similarity and semantic similarity, the alternative with the highest overall score is selected.

[0051] In this embodiment, when the robot encounters an anomaly, substitutions are made based on the cause of the failure, significantly improving system robustness and task completion rate. For issues such as missing items and path obstruction, the system accurately matches functionally and semantically similar alternatives, avoiding task interruptions, reducing invalid retries, and improving execution efficiency. This differentiated strategy balances task objectives with environmental adaptability, enabling the system to operate stably in dynamic scenarios, reducing the impact of failures, ensuring task continuity, and enhancing adaptability to complex environments.

[0052] 2. Backtracking mechanism Furthermore, if all alternatives for the current node are exhausted, the system will trigger a backtracking mechanism. Unlike traditional simple backtracking, this embodiment introduces a backtracking step reduction factor to balance different backtracking depths, in order to avoid frequent backtracking causing a large waste of completed tasks.

[0053] Specifically, for a candidate path back to the k-th historical node, its comprehensive value is determined by both the original value of the node and the backtracking factor, as shown in the formula: ; in, This represents the original assessed value of the branch path corresponding to the k-th node when backtracking; k represents the number of backtracking steps; λ is the backtracking penalty coefficient, used to control the value decay caused by increasing backtracking depth. Through this mechanism, even if some earlier nodes have higher theoretical value, the system will still comprehensively consider the execution waste caused by backtracking, thereby achieving a balance between efficiency and benefit.

[0054] During the backtracking node selection process, the system not only considers the value assessment of the node, but also comprehensively incorporates the following structural information.

[0055] The first factor is the number of nodes. The more branches a node has, the greater its potential for replacement.

[0056] Secondly, there's the remaining task depth, which refers to the number of task steps required to backtrack from the current node to the target node; in other words, the "complexity" or "depth" of the remaining task. It measures the distance from the current node to the completion of the task objective. Starting from the current backtracking node, the number of task steps required to reach the target node is calculated. This number of steps is usually represented by the path length in the task graph, i.e., the length of the item operation sequence from the backtracking node to the target node. Considering that there may be multiple candidate paths in the task graph, the number of task steps for each path is calculated. Assume path P = (N1, N2, ..., N...). n ), where N1 is the current node, N n If the target node is n, then the path depth is n-1. By calculating the path depth of all candidate paths, the path with the minimum depth is selected as the priority. The formula for the remaining task depth can be expressed as: ; in, For backtracking nodes The remaining task depth, For path Number of steps (i.e. from node) Path length to the target node. For the node All candidate paths to the target node. The smaller the remaining task depth, the closer the path is to the target and should be prioritized. This factor helps the robot prioritize paths closer to the target, thereby improving task recovery efficiency.

[0057] Finally, execution stability and the degree of change in environmental constraints are considered. Together, these factors measure the stability of the path execution corresponding to the current backtracking node, as well as the impact of changes in the task environment on path execution.

[0058] Execution stability This refers to the failure rate during path execution. It is calculated by statistically analyzing the success rate of previous executions of the same path. Historical execution data or real-time feedback data can be used to estimate the stability of the path. Assuming multiple execution records, the success rate of the path can be calculated by counting the number of successful executions and the number of failures: Success Rate = Number of Successful Executions / Total Number of Executions. The closer the result is to 1, the more stable the path is.

[0059] The degree of environmental constraint change measures the impact of environmental changes (such as changes in item state, position, robot state, etc.) on the execution of the current path. Significant environmental changes can lead to unstable path execution. The impact of environmental changes is assessed by monitoring them in real time. If the path of the current backtracking node is significantly affected by environmental changes, a penalty value can be set to represent the instability of the path execution. For example, the degree of environmental change can be set as ΔE, and the path score can be adjusted based on the magnitude of the change. The specific calculation can be performed using a weighting factor based on the magnitude of the change. ; in, For backtracking nodes The degree of environmental change Weighting factors for environmental change This refers to the magnitude of environmental change.

[0060] Higher execution stability indicates that the path has been more reliable in the past and should be chosen first. Less change in environmental constraints indicates that the current path is less affected by environmental changes, resulting in higher path stability and making it a more desirable choice.

[0061] The comprehensive evaluation score for each candidate backtracking node is calculated based on the above factors, and the optimal node is selected as the new starting point for task recovery. The final comprehensive evaluation score will be a weighted average of multiple factors. The specific evaluation formula is as follows: ; in, For backtracking nodes The overall evaluation score, For backtracking nodes The number of branches, For backtracking nodes The remaining task depth, For backtracking nodes Execution stability, For backtracking nodes The degree of environmental change. For the weighting coefficients, satisfying The number of branches a node has reflects its potential for substitution; a higher number of branches enhances the flexibility of alternative solutions. After calculating the comprehensive evaluation score of all candidate backtracking nodes, the system will select the node with the highest score as the new starting point for task recovery.

[0062] By comprehensively evaluating these factors, the system can be more intelligent in selecting backtracking nodes, prioritizing shorter, more stable paths that are less affected by environmental changes, thereby improving the efficiency and stability of task recovery.

[0063] 3. Cascading Trigger Mechanism Furthermore, to reduce the cost of calling the large language model, a cascading triggering mechanism is designed. In most cases, task recovery and path selection can be quickly decided solely based on rule matching, cost evaluation, and heuristic scoring. The system only triggers the large language model to participate in the decision-making process in the following situations: the difference in the evaluated value between alternative paths is too small to distinguish their superiority; conflicts arise between task constraints; environmental updates cause the original path to become invalid as a whole; and neither backtracking nor alternative strategies can find a feasible solution.

[0064] When the triggering conditions are met, the large language model will semantically reorder the candidate paths and dynamically expand and complete the task graph according to the current scenario, generating new candidate item paths. The system then regenerates the optimal task plan based on the updated task graph and continues execution. This fine-grained correction mechanism significantly reduces computational overhead while ensuring task execution stability, improving the task recovery capability and execution efficiency of home service robots in complex and dynamic environments.

[0065] In this embodiment, a fine-grained correction mechanism is used to improve the recovery capability and execution robustness under abnormal conditions, reduce the number of large language model calls, reduce planning latency and computational costs, and improve online execution efficiency.

[0066] After fine-grained correction, the system dynamically updates the task graph structure based on the current execution status and environmental feedback. It re-evaluates the feasibility of the current path based on the original task graph and generates new candidate paths through real-time perception. For the selection among multiple candidate paths or alternatives, the system employs a fuzzy logic mechanism to optimize path selection. Each factor in path selection (cost, success probability, semantic matching degree, etc.) is defined as a fuzzy set, and paths are comprehensively evaluated using fuzzy inference rules. The system inputs the evaluation value of each candidate path into a fuzzy decision matrix and calculates the priority of each path through fuzzy inference. Based on factors such as path cost, success probability, semantic matching degree, and task complexity, the system selects the path with the highest comprehensive score. This considers not only cost, success probability, and semantic matching degree but also factors such as task complexity and execution stability, resulting in a more intelligent path selection.

[0067] Furthermore, reinforcement learning is introduced to self-optimize the selection of alternative paths. Through task history and real-time feedback, the system gradually learns which paths are more effective in similar situations and selects them as the preferred paths. By continuously optimizing the decision-making strategy, the system can select the optimal path based on environmental changes and the long-term interests of the task objectives.

[0068] This specific embodiment addresses the prominent issues of difficult task planning, weak anomaly recovery capabilities, and high online computing costs in home service robots, making comprehensive improvements across four dimensions: plan generation, scenario adaptation, anomaly handling, and computational efficiency. It utilizes a large language model to generate multiple candidate item sequences and constructs a temporal task graph based on common temporal sequences, effectively integrating natural language task understanding capabilities with stable task execution logic, avoiding the unreliability of single-planning approaches. By matching and pruning the task graph in conjunction with real-world scenario states and selecting the optimal item operation path based on a value function, it significantly improves the adaptability of planning to the actual environment and reduces ineffective and high-cost planning. A fine-grained correction mechanism is introduced during the execution phase, employing an anomaly correction strategy that combines substitution priority and backtracking. In case of anomalies, local adjustments are prioritized over a complete regeneration of the plan, greatly improving system stability and task completion efficiency. Simultaneously, this approach reduces the frequent calls to the large language model during online execution, effectively alleviating computational overhead and latency, and better adapting to resource-constrained robot platforms, providing robust, efficient, and implementable task planning and anomaly handling capabilities for open and dynamic home environments.

[0069] Example 2 This embodiment provides a robot task planning and anomaly correction system based on common temporal sequences, including: The task graph construction module is configured to input user task instructions into a large language model and generate multiple candidate item sequences; and to construct a common temporal task graph based on the candidate item sequences, with items as nodes and the order of item operations as edges. The pruning module is configured to match and prune the task graph according to the real scene state, and calculate the optimal item operation path on the pruned task graph based on the value function. The fine-grained task generation module is configured to input the user task instructions, real scene state and optimal item operation path into the large language model, and obtain an executable fine-grained task sequence through the correction of task instruction penalty terms and drive the robot to execute it. The correction module is configured to introduce a fine-grained correction mechanism when the robot encounters an anomaly. This mechanism adjusts the task sequence through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms to select the branch path with the highest overall value to continue execution.

[0070] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the robot task planning and anomaly correction method based on common timing as described in Embodiment 1 above.

[0071] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the robot task planning and anomaly correction method based on common timing as described in Embodiment 1 above.

[0072] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for task planning and anomaly correction for home service robots based on common temporal sequences, characterized in that, include: The user task instruction is input into the large language model to generate multiple candidate item sequences; a common temporal task graph with items as nodes and the order of item operations as edges is constructed based on the candidate item sequences; specifically, the user task instruction I is obtained, and the large language model is called to generate K candidate item sequences based on the user task instruction I. Each sequence consists of several items required to complete the task arranged in the order of operation, which are used to represent the core items and optional item set required to complete the task. After the candidate item sequence is generated, the item names are normalized and standardized using synonyms in order to merge similar items and unify the node representation; The task graph is matched and pruned according to the real scene state, and the optimal item operation path is calculated based on the value function on the pruned task graph. Specifically, the matching and pruning of the task graph based on the real-world scenario involves: matching and pruning the task graph based on item existence, item reachability, item availability, hard constraint rules, and connectivity cleanup; item existence refers to determining whether an item exists in the environment using sensors and an environmental perception system; if an item is missing, its node value is set to 0, and the corresponding node in the task graph is pruned; item reachability refers to assessing whether an item is reachable based on the robot's current position and the navigation system; if an item is located in an inaccessible area, its reachability value is set to low, and the corresponding node in the task graph is pruned; item availability refers to whether an item is in an operable state and can be used to perform the task; a pruning threshold is set. =0.2, deletion availability is less than The nodes and related edges; the hard constraint rule is that if the instruction contains a mandatory target item, then the feasible path containing the item must be retained, and the edges involving the item are not allowed to be cut; the connectivity cleanup is used to ensure that all remaining item nodes and edges remain validly connected in the task graph, and to delete those nodes and edges that have become isolated or cannot reach the target. The user task instructions, real-world scenario states, and optimal item operation paths are input into a large language model. Through correction of the task instruction penalty term, an executable, fine-grained task sequence is obtained and the robot is driven to execute it. Specifically, this includes: The user task instructions, the real scene state, and the optimal item operation path are input into the large language model to obtain the initial probability distribution of the action sequence. The initial probability distribution is modified by introducing a semantic ambiguity penalty term, a vagueness penalty term, and a contextual inconsistency penalty term; wherein, the semantic ambiguity penalty term targets the polysemy of words in the instruction, the vagueness penalty term targets the multiple interpretations of the instruction description, and the contextual inconsistency penalty term targets the contradiction between the instruction and the scene state. If the corrected probability distribution is greater than the threshold, the robot is driven to execute; otherwise, the item is matched to the item ID in the actual scene through semantic analysis mapping, or the target is confirmed with the user and the probability distribution is recalculated after receiving a clear instruction. When the robot malfunctions, a fine-grained correction mechanism is introduced to adjust the task sequence through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms, so as to select the branch path with the highest overall value to continue execution. The alternative selection strategy specifically involves: analyzing the reasons for failure when the item or operation corresponding to the current task node is unavailable; the reasons for failure include missing items or unreachable paths; determining the alternative type based on the reasons for failure; if the item is missing, finding the most functionally similar alternative item; if the path is unreachable, finding a functionally equivalent alternative path; wherein, based on the pre-classification of the functional attributes of items and paths in a functional semantic library, calculating the functional similarity between entities of the same type, and using a pre-trained semantic model to calculate the similarity between the keywords in the task instruction and the semantic vectors of candidate alternatives, combining functional similarity and semantic similarity to select the alternative option with the highest comprehensive score; the backtracking mechanism Specifically, backtracking is performed when alternative solutions are exhausted; the path is comprehensively evaluated based on backtracking value assessment, number of node branches, remaining task depth, execution stability, and degree of change in environmental constraints, and the branch path with the highest overall value is selected at the backtracking node to continue execution; the backtracking value assessment introduces a backtracking step reduction factor to reduce the value as the backtracking depth increases; the cascading triggering mechanism is specifically: the large language model is triggered to update the path only when the difference in the assessed value between alternative paths is less than a threshold and it is difficult to distinguish between superior and inferior paths, when conflicts arise between task constraints, when environmental updates cause the original path to fail as a whole, or when neither the alternative selection strategy nor the backtracking mechanism can find a feasible solution.

2. The method for task planning and anomaly correction of a home service robot based on common temporal sequence as described in claim 1, characterized in that, The value function is constructed based on the item's contribution to task completion, success probability, execution cost, and co-occurrence frequency. Task map after pruning Constructing a value function The calculation formula is: ； Among them, nodes For an item, m represents the total number of item nodes. Represents item nodes Contribution to completing the instruction, This represents the probability of successfully obtaining / using the item in scenario S. This represents the cost of performing operations related to this item. It is the sequential edge weight, representing the co-occurrence frequency and switching cost.

3. A task planning and anomaly correction system for home service robots based on common temporal sequences, characterized in that, The method for task planning and anomaly correction of home service robots based on common temporal sequence as described in any one of claims 1-2 includes: The task graph construction module is configured to input user task instructions into a large language model and generate multiple candidate item sequences; and to construct a common temporal task graph based on the candidate item sequences, with items as nodes and the order of item operations as edges. The pruning module is configured to match and prune the task graph according to the real scene state, and calculate the optimal item operation path on the pruned task graph based on the value function. The fine-grained task generation module is configured to input the user task instructions, real scene state and optimal item operation path into the large language model, and obtain an executable fine-grained task sequence through the correction of task instruction penalty terms and drive the robot to execute it. The correction module is configured to introduce a fine-grained correction mechanism when the robot encounters an anomaly. This mechanism adjusts the task sequence through alternative selection strategies, backtracking mechanisms, or cascading triggering mechanisms to select the branch path with the highest overall value to continue execution.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for task planning and anomaly correction of a home service robot based on common timing as described in any one of claims 1-2.

5. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for task planning and anomaly correction of a home service robot based on common timing as described in any one of claims 1-2.

Images