Robot curtain wall installation method and system based on large language model

Abstract

The invention provides a robot curtain wall installation method and system based on a large language model, and the method comprises the steps: generating a cue word of a corresponding stage based on an S-T-A-D cue word template according to an assembly stage of a to-be-assembled curtain wall; the assembling stage comprises a leveling stage or a clamping stage; the S-T-A-D cue word template comprises a scene state, a scene task, a skill library and a decision task; inputting the cue word into a pre-trained large language model to generate a decision instruction; and verifying the feasibility of the decision instruction through digital twin simulation, driving the robot to execute the decision instruction meeting the feasibility, and completing the assembly of the curtain wall and the main part. According to the method, assembly requirements are dynamically analyzed through a large language model, a decision instruction is generated based on an S-T-A-D mode, and self-adaptive control of curtain wall leveling and clamping is achieved in combination with digital twinborn verification and a hierarchical skill library. Historical experience can be automatically associated to cope with complex pendants, parameters do not need to be reconfigured, and the assembly precision and efficiency are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of building decoration installation, and in particular to a robot curtain wall installation method and system based on a large language model. Background Art

[0002] During the installation of building curtain walls, typical contact states such as leveling and snapping are crucial for ensuring precision and stability. These contact states correspond to the mechanical interactions and spatial positioning requirements between curtain wall components. For example, curtain wall panels must be precisely leveled with the main mounting components and mechanically snapped into place. Therefore, selecting appropriate robotic assembly skills for each assembly state and designing a sound autonomous decision-making system are crucial for achieving flexible curtain wall installation.

[0003] Currently, in the construction and decoration industry, curtain wall installation mostly adopts the dry hanging method, which has a complex assembly process. The existing robotic curtain wall installation system often relies on sensor fusion path planning algorithms to optimize the assembly process, but has the defect of fixed assembly strategy: when the type of curtain wall hangers or the size of the main installation parts change, the assembly parameters and environmental parameters of the robotic installation system need to be reconfigured. This fixed strategy plus manual reconfiguration mode is difficult to quickly adapt to the diverse needs of the project and has poor flexibility. It not only increases the debugging cost, but also seriously restricts the installation efficiency of the curtain wall. Summary of the Invention

[0004] In response to the problems that current robotic curtain wall installation methods lack in-depth understanding and flexible adaptability of complex scenarios, and are unable to quickly adjust assembly strategies according to real-time changes, combined with the process flow of curtain wall assembly, the present invention proposes a robotic curtain wall installation method and system based on a large language model. Through natural language interaction and dynamic decision-making optimization, it autonomously analyzes assembly requirements and makes decisions based on scene status, realizing efficient and adaptive assembly of curtain wall main components and hanging parts, and improving the assembly accuracy, flexibility and generalization of robots in complex construction scenarios.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a robot curtain wall installation method based on a large language model, comprising: According to the current assembly stage of the curtain wall to be assembled, a prompt word for the corresponding stage is generated based on a preset STAD prompt word template; the assembly stage includes a leveling stage or a locking stage; the STAD prompt word template includes a scene state, a scene task, a skill library, and a decision task; Input the prompt word into a pre-trained large language model to generate a decision instruction; The feasibility of the decision instructions is verified through digital twin simulation, and the robot is driven to execute the decision instructions that meet the feasibility to complete the assembly of the curtain wall and the main components.

[0006] In a second aspect, the present invention provides a robotic curtain wall installation system based on a large language model, comprising: The prompt word generation module is configured to generate prompt words for the corresponding stage based on a preset STAD prompt word template according to the current assembly stage of the curtain wall to be assembled; the assembly stage includes a leveling stage or a locking stage; the STAD prompt word template includes a scene state, a scene task, a skill library, and a decision task; A decision generation module is configured to input the prompt word into a pre-trained large language model to generate a decision instruction; The decision execution module is configured to verify the feasibility of the decision instructions through digital twin simulation, drive the robot to execute the decision instructions that meet the feasibility, and complete the assembly of the curtain wall and the main parts.

[0007] In a third aspect, 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 a robotic curtain wall installation method based on a large language model described in the first aspect.

[0008] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the robot curtain wall installation method based on a large language model described in the first aspect are implemented.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) Conventional technologies are difficult to adapt to diverse needs in curtain wall installation due to their fixed strategies. This invention introduces a large language model with the help of STAD prompt word templates. Based on the scene states and tasks of different assembly stages such as contact, leveling, and engagement, it flexibly calls knowledge from the skill library to generate decision instructions, breaking through the limitations of fixed strategies. When facing changes in curtain wall hanging types and main component sizes, there is no need to manually reconfigure parameters, and it can quickly adapt to diverse needs. Digital twin simulation verifies the feasibility of decision instructions, avoids assembly risks in advance, and ensures the accuracy of robot execution instructions. The two work together to reduce manual intervention, lower debugging costs, and improve the efficiency and flexibility of curtain wall installation.

[0010] (2) When faced with an unknown contact state during the leveling phase, the present invention utilizes historical experience data from the case library and retrieves similar scenarios using a cosine similarity algorithm. This allows for rapid association of six-dimensional state sequences and action logs accumulated through teaching or trial and error. These data are stored as multi-dimensional indexes, enabling precise matching of the current pendant structural features. The adaptive leveling strategy is directly reused with historical effective experience, improving leveling adaptability and efficiency and ensuring assembly stability.

[0011] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.

[0013] Figure 1 A main flow chart of a large language model-based robotic curtain wall installation method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a curtain wall to be assembled with SE-type hooks provided in an embodiment of the present invention; Figure 3 This is a curtain wall assembly diagram provided by an embodiment of the present invention; wherein (a) represents the leveling stage, and (b) represents the locking stage; Figure 4 A schematic flow chart of a large language model-based robotic curtain wall installation method provided by an embodiment of the present invention; Figure 5 A schematic diagram of a large language model prompt word template provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Example 1 like Figure 1 As shown, this embodiment discloses a robot curtain wall installation method based on a large language model, comprising the following steps: S1: Generate prompt words for the corresponding stage based on a preset STAD prompt word template according to the current assembly stage of the curtain wall to be assembled; the assembly stage includes a leveling stage or a locking stage; the STAD prompt word template includes a scene state, a scene task, a skill library, and a decision task; S2: Input the prompt word into a pre-trained large language model to generate a decision instruction; S3: Verify the feasibility of the decision instructions through digital twin simulation, drive the robot to execute the decision instructions that meet the feasibility, and complete the assembly of the curtain wall and main components.

[0016] Based on the characteristics of the curtain wall assembly process, this embodiment takes the curtain wall to be assembled with SE-type hooks as an example. Figure 2 As shown in the figure, the curtain wall to be assembled forms an assembly object with the main curtain wall component through the SE-type hook.

[0017] When the robotic arm clamps the curtain wall to be assembled and gradually advances the curtain wall main component towards the predetermined target position, the SE-type hanger comes into contact with the hook on the curtain wall main component. At this time, at least one end of the hook of the curtain wall to be assembled lightly touches the curtain wall main component. The curtain wall to be assembled, with the lightly touched end as the fulcrum, is tilted as a whole relative to the installation reference plane of the curtain wall main component, indicating that the assembly process has entered the contact stage.

[0018] After contact, the assembly enters the leveling stage: the spatial posture of the curtain wall to be assembled needs to be adjusted by the robotic arm, and the tilt angle is corrected based on the contact force and torque data fed back by the six-dimensional force sensor until the main curtain wall component and the SE hook reach a horizontal alignment state, such as Figure 3 (a) The core goal of this stage is to eliminate assembly posture deviations and achieve preliminary coupling between the main curtain wall components and hanging components through real-time posture calibration.

[0019] After the posture leveling is completed, the assembly enters the locking stage: on the basis of ensuring that the curtain wall main component and the SE hook are horizontally aligned, the control arm calculates the joint drive amount through the robot inverse kinematics model, and accurately adjusts the position parameters of the curtain wall to be assembled along the three-dimensional space. Through progressive advancement, the SE hook's slot gradually engages with the curtain wall main component hook's boss until the six-dimensional force sensor detects that the locking contact force reaches the preset threshold, indicating that the SE hook and the curtain wall main component are mechanically locked. Figure 3 (b) This stage is guided by engagement accuracy and contact force stability, and achieves precise engagement and assembly of the curtain wall main components and hanging components through position error convergence and contact force closed-loop control.

[0020] Next, combine Figure 4 , a robot curtain wall installation method based on a large language model disclosed in this embodiment is described in detail.

[0021] 1. Contact stage Firstly, the force-position hybrid control theory is adopted to construct an admittance control model in the Z-axis (assembly direction) to achieve constant force contact.

[0022]

[0023] Where, For the Z-axis position adjustment, the system first passes Adjust the Z-axis position to try to contact the main curtain wall component to achieve constant force contact control, and monitor the force signal in real time during the process . is the admittance matrix, is the expected contact force (set to 0N). Preset threshold (set to 5N), when detected When it is determined to be in contact, the six-dimensional state sequence is collected And trigger the contact state perception network.

[0024] Among them, the elements in the six-dimensional state sequence ,in, = are the three-dimensional force components, corresponding to the contact forces in the x, y, and z axis directions; = is the three-dimensional attitude angle, Roll angle, Pitch angle, The yaw angle is used to describe the spatial posture of the curtain wall to be assembled.

[0025] Use normalization method for preprocessing:

[0026] in, Receive the processed state vector, and is the training base mean and standard deviation.

[0027] In this embodiment, a contact state perception network is built based on LSTM. Input layer: dimension 6, receiving the processed state vector LSTM hidden layer: h memory units, which handle temporal dependencies through a gating mechanism. The core formula is as follows:

[0028] in, is the output of the forget gate, is the sigmoid activation function, is the input feature To the weight matrix of the forget gate, Hide the state at the last moment To the weight matrix of the forget gate, is the bias vector of the forget gate.

[0029]

[0030] in, is the input gate output, is the input feature to the weight matrix of the input gate, Hide the state at the last moment to the weight matrix of the input gate, is the bias vector of the input gate, is the cell candidate state, is the input feature To the weight matrix calculated for the candidate cell state, Hide the state at the last moment To the weight matrix calculated for the candidate cell state, is the bias vector when calculating the candidate cell state.

[0031]

[0032] in, is the cell state at the current moment, is element-wise multiplication, is the cell state at the previous moment.

[0033]

[0034] in, is the output gate output, Input features The weight matrix to the output gate, Hide the state at the last moment The weight matrix to the output gate, is the bias vector of the output gate, is the hidden state at the current moment.

[0035] Taking the contact moment as the starting point, intercept the sliding window of length n. Input the contact state perception network to obtain the contact state.

[0036] The contact status indicates the position of the curtain wall to be assembled relative to the main component, including upper forward tilt, lower forward tilt, left tilt, right tilt, upper left forward tilt, upper right forward tilt, lower left forward tilt, lower right forward tilt, leveling, no contact, fit, and engagement.

[0037] Determine the assembly stage based on the contact status. Specifically, for the leveling stage: 1. Single point contact: Only one of the four ends of the curtain wall to be assembled (such as the upper left, upper right, lower left or lower right) is in contact with the main component hanger, and the other three non-contact ends deviate from the target posture due to the overall tilt, resulting in a rotation deviation around the contact end.

[0038] 2. Double-point contact: Two of the four ends of the curtain wall to be assembled are in contact (the two ends of the upper part, the two ends of the lower part, the two ends of the left part, and the two ends of the right part). The remaining two non-contact ends deviate from the target posture due to the overall tilt, resulting in a rotation deviation around the line connecting the contact ends.

[0039] 3. Unknown contact: When the pendant configured for the curtain wall to be assembled is not SE type (such as curved hook), the state when it contacts the main part.

[0040] For the engagement phase: 1. Leveling state; 2. Non-contact state: Based on the leveling state and moving along the Z-axis, the Z-axis force signal of the six-dimensional force sensor suddenly changes, and the value drops significantly to close to 0, while the Y-axis force signal remains unchanged.

[0041] 3. Fitting state: The force signal of the six-dimensional force sensor returns to the leveling state in the first stage of the assembly operation.

[0042] 4. Engaged state: The curtain wall main component hook is engaged with the SE pendant.

[0043] As an implementation method, the contact state perception network is trained based on historical data. Starting from the historical contact moment, a sliding window of length n is intercepted to generate 12 types of contact state samples. At the same time, the state that cannot be recognized by the contact state perception network is regarded as the unknown state s, which represents the state generated when the curtain wall hanger type and the size of the main installation part change. The label set is input to form the label set. .

[0044] It should be understood that, usually, the leveling stage must be completed before entering the engaging stage; when the curtain wall to be assembled is in contact with the main component and is already in the leveling state, the engaging stage can be directly entered to perform position control operations.

[0045] (2) Leveling stage When the contact state perception network identifies that the curtain wall is in a tilted position (such as leaning to the left, leaning to the lower right, etc.), the assembly enters the leveling stage. The core goal of this stage is to adjust the curtain wall posture so that it is horizontally aligned with the SE pendant. The specific steps are as follows: 1. Scene state description and prompt word construction Based on the six-dimensional force sensor data and the contact state perception network output, the current contact type (single-point / double-point contact) and specific tilt direction (such as "upper right forward tilt") are determined.

[0046] like Figure 5 As shown, generate prompt words according to the STAD mode: Scene status (S): Currently in the leveling stage, the upper right end of the curtain wall main component is in contact, and the other ends are tilted to the lower right, obtaining six-dimensional status data.

[0047] Scenario Task (T): Level the curtain wall main component and SE pendant to eliminate posture deviation.

[0048] Skill Library (A): This library calls the "Posture Adjustment" skill library, which contains eight types of rotation actions corresponding to tilt states (e.g., "right tilt" corresponds to "left turn"). The "Posture Adjustment" skill library is shown in Table 1.

[0049] Table 1 Posture adjustment skills library

[0050] Among them, the upper forward tilt indicates the state where the two end points of the upper part of the curtain wall to be assembled are in contact with the main part, the left tilt indicates the state where the two end points of the left side of the curtain wall to be assembled are in contact with the main part, and so on for other states.

[0051] If a new type of widget is encountered (such as an arc SE widget), the case library is called to obtain the action.

[0052] Decision-making task (D): Output the specific rotation axis, angle, and execution sequence to ensure that the position requirements of the engagement phase are met after leveling.

[0053] Among them, in some decision-making scenarios, the decision task (D) is used to make precise question-oriented consulting requests to the large language model, such as "In the current environment, how should I act? Give some targeted operational guidance", so that the large language model can produce an action plan that meets the needs.

[0054] 2. Large Language Model Decision-Making and Action Generation The large language model is based on the prompt word, combined with the homogeneous transformation matrix and rotation matrix algorithm in the posture adjustment skill library (such as the rotation formula around the Z axis ), calculate the required rotation angle .

[0055] Specifically, in three-dimensional space, the posture of the main curtain wall component can be described by a homogeneous transformation matrix. Assume that the initial posture of the main curtain wall component is , after a series of translations and rotations, it reaches the target pose The homogeneous transformation matrix is ​​a 4*4 matrix, and its general form is:

[0056] Among them, R is a 3*3 rotation matrix, which describes the rotation posture of the object; t is a 3*1 translation vector, which is used to describe the position of the object.

[0057] For rotation around the coordinate axis, rotate around the Z axis Angle as an example, the rotation matrix for:

[0058] When the main curtain wall component is tilted, it is necessary to adjust its posture by rotating it so that it is level with the SE pendant. For example, when the main curtain wall component is in a left-leaning state, it is necessary to rotate it around the Z axis at a certain angle. , the rotated pose Can be achieved through Calculated.

[0059] For example, if a "left lean" is detected, the model automatically matches the "right turn" action in the skill library and generates a clockwise rotation around the Z axis. instructions.

[0060] As an implementation method, if a new type of pendant (such as a curved SE pendant) is encountered, the large language model searches the assembly case library, and uses the cosine similarity algorithm to search for similar scenarios that match the current pendant structure features. It matches historical scenarios and adjustment strategies such as "surface leveling" and "multi-point contact strategy", extracts leveling experience of similar structures (such as small-angle rotation in stages and force feedback correction), and integrates it into the current decision.

[0061] Among them, historical experience data includes six-dimensional state sequences, action logs and pendant structure feature annotations collected through teaching or trial and error, which are stored in the case library and indexed according to the structural features-assembly state-action result dimensions.

[0062] For example, for curved pendants, the model may prioritize the strategy of "multiple fine-tuning of rotation angles and real-time force monitoring". By making multiple fine-tuning adjustments and comparing them with the action results, it can avoid exceeding the contact force limit due to a one-time adjustment.

[0063] By linking historical experience data, this embodiment automatically retrieves similar scenarios from a case library when faced with non-standardized structures such as curved SE pendants, extracts leveling experience, and incorporates it into current decisions, achieving precise assembly without the need for additional specialized training. This upgrades the traditional assembly process from preset parameters and manual debugging to an intelligent perception and autonomous decision-making approach, significantly reducing parameter reconfiguration costs and improving assembly efficiency and accuracy in complex scenarios. This provides an efficient and generalized solution for automated building curtain wall construction.

[0064] 3. Action execution and feedback closed loop The robot arm performs the rotation action, and the six-axis force sensor provides real-time feedback on the contact force changes. If the force signal is abnormal during the leveling process (such as Mutation), the system triggers the digital twin feedback closed-loop mechanism: The large language model generates candidate action plans (such as "pause rotation and detect contact point offset") and calls the digital twin simulation environment to simulate and execute the decision instruction.

[0065] In the digital twin environment, the mechanical response (such as the force distribution on each axis, torque change) and posture change (rotational deviation around the contact end) of the simulation execution process are dynamically evaluated to detect whether there are risks such as "posture error expansion (such as rotational deviation exceeding the threshold) and contact anomaly (such as single-point contact turning into non-contact)".

[0066] If the simulation results do not meet the leveling stability requirements (for example, if it is predicted that continued rotation will cause the posture error to expand), the action parameters are corrected based on the mechanical response and posture data. For example, the original rotation angle of 10° is adjusted to "5° + X-axis translation compensation" to generate a corrected decision instruction.

[0067] The revised decision instructions are re-input into the large language model, the skill library is updated, and the subsequent leveling decision strategies are iteratively optimized until the six-dimensional force sensor shows that the forces on each axis are evenly distributed, and the leveling is determined to be complete.

[0068] This embodiment introduces digital twin simulation during the curtain wall leveling and assembly phase. This technology dynamically simulates the contact state between the curtain wall and the main components and hangers during the leveling process, assesses mechanical response and posture changes in real time, and accurately identifies risks such as increased posture error and contact anomalies in complex conditions such as single-point and double-point contact. When force signals are abnormal, a large language model combined with simulation rapidly generates and verifies candidate actions. Parameters are corrected based on the simulation data, and the leveling strategy is iteratively optimized to ensure even force distribution across all axes. This effectively improves leveling accuracy, avoids curtain wall damage and construction delays caused by physical trial and error, and makes the leveling process more intelligent, efficient, and precise.

[0069] (3) Engagement stage After leveling is completed, the assembly enters the locking phase, the goal of which is to achieve mechanical locking of the SE hook and the curtain wall main component through precise position control. In this embodiment, it is assumed that different types of hooks perform the same assembly action in the locking phase. The specific steps are as follows: 1. Position Error Calculation and Skill Invocation Based on the robot inverse kinematics model, the target position of the robot end effector in Cartesian space is , the current location is , then the position error is ,in, To engage the target position, is the current location.

[0070] Generate prompt words according to STAD mode: Scene status (S): Leveling completed, current position is [ ], the target position is [ ], the six-dimensional force signal is stable.

[0071] Scenario Task (T): Accurately adjust the position along the three-dimensional space to complete the engagement of the SE hook and the main component.

[0072] Skill Library (A): Calls the "Position Control" skill library, which includes actions such as "Move Forward," "Move Down," and "Stop." The details of the "Position Control" skill library are shown in Table 2.

[0073] Table 2 Position control skill library

[0074] Decision-making task (D): 2. Large Language Model Decision-Making and Action Generation The large language model calculates the adjustment angle based on the prompt words and the position control skill library. For a robot with n joints, the end effector position can be expressed as the joint angle Function , using inverse kinematics to solve Joint angle By collecting the position information of the robot's end effector in real time, calculating the position error and its rate of change, and driving the robot to accurately track the target position, the main curtain wall component and the SE pendant are engaged. The following Newton-Raphson iteration formula is used to calculate the angles of each joint and achieve position error convergence:

[0075] in, is the joint angle at the kth iteration, It's the robot The Jacobian matrix at .

[0076] 3. Action execution and feedback closed loop The robot arm moves forward gradually along the assembly direction (Z axis), and the six-axis force sensor monitors the engagement contact force in real time. When the preset threshold (e.g., 50N) is reached and stabilized, the engagement is considered successful. If the force signal is abnormal (e.g., a sudden drop or exceeding the limit), the digital twin feedback closed-loop mechanism is triggered: The large language model calls the digital twin simulation system to simulate the execution of the current action instructions, such as continuing the Z-axis advancement to reproduce the engagement process between the pendant and the main part.

[0077] In the digital twin environment, the mechanical response (such as engagement contact force distribution, axial force fluctuation) and posture changes (such as pendant offset, main component deformation) of the simulation execution process are dynamically evaluated to detect whether there are risks such as "engagement interference (such as lateral collision), uneven contact force (such as local stress concentration)".

[0078] If the simulation shows a deviation in the engagement path (for example, if it is predicted that continued advancement will cause a side collision), the action parameters are automatically adjusted based on the mechanical response and posture data (for example, adding a Y-axis micro-shift to avoid a collision), and a corrected action command is generated.

[0079] The corrected action instructions are used as iterative data to update the skill library, and the large language model is re-entered to optimize the subsequent decision-making logic to ensure that the engagement process is stable and there is no abnormal force interference.

[0080] This embodiment incorporates digital twin simulation during the curtain wall assembly phase, effectively detecting risks such as engagement interference and uneven contact force. Based on the simulation results, motion parameters are automatically adjusted to correct the engagement path. This corrected data is then fed back to update the skill library, iteratively optimizing the decision logic to ensure a stable engagement process, improve engagement success rates, and avoid component damage caused by engagement errors, ultimately driving the evolution of intelligent and precise assembly.

[0081] After the contact stage, leveling stage and locking stage, the assembly of the curtain wall and the main components is completed.

[0082] This specific embodiment utilizes the STAD prompt system and a large language model to analyze scene states in real time. Combined with a posture adjustment library, a position control library, and a case library based on historical experience, it dynamically generates leveling and engagement strategies, avoiding the inflexibility inherent in traditional systems due to fixed strategies. A digital twin simulation verification mechanism ensures the mechanical feasibility of decision instructions, and a closed-loop feedback loop continuously optimizes action plans to enhance assembly stability.

[0083] Example 2 This embodiment provides a robotic curtain wall installation system based on a large language model, including: The prompt word generation module is configured to generate prompt words for the corresponding stage based on a preset STAD prompt word template according to the current assembly stage of the curtain wall to be assembled; the assembly stage includes a leveling stage or a locking stage; the STAD prompt word template includes a scene state, a scene task, a skill library, and a decision task; A decision generation module is configured to input the prompt word into a pre-trained large language model to generate a decision instruction; The decision execution module is configured to verify the feasibility of the decision instructions through digital twin simulation, drive the robot to execute the decision instructions that meet the feasibility, and complete the assembly of the curtain wall and the main parts.

[0084] Example 3 This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the robot curtain wall installation method based on a large language model as described in the first embodiment above are implemented.

[0085] 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, the steps of the robot curtain wall installation method based on a large language model as described in the first embodiment above are implemented.

[0086] The steps or modules involved in Examples 2 to 4 above correspond to those in Example 1. For detailed implementations, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media that includes one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and cause the processor to perform any method of the present invention.

[0087] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A robot curtain wall installation method based on a large language model, characterized in that: include: According to the current assembly stage of the curtain wall to be assembled, the prompt words of the corresponding stage are generated based on the preset STAD prompt word template; The assembly stage includes a leveling stage or a locking stage; the STAD prompt word template includes scene status, scene task, skill library and decision task; Input the prompt word into a pre-trained large language model to generate a decision instruction; The feasibility of the decision instructions is verified through digital twin simulation, and the robot is driven to execute the decision instructions that meet the feasibility to complete the assembly of the curtain wall and the main components.

2. A robot curtain wall installation method based on a large language model as claimed in claim 1, characterized in that: The confirmation process of the assembly stage of the curtain wall to be assembled is specifically as follows: Based on the contact events between the curtain wall to be assembled and the main component, the sensor data at the end of the robot arm holding the curtain wall to be assembled is collected; Inputting the sensor data into a pre-trained contact state perception network to determine the current contact state, and determining the assembly stage according to the contact state; Among the contact states, the upper forward leaning, lower forward leaning, left leaning, right leaning, upper left forward leaning, upper right forward leaning, lower left forward leaning, lower right forward leaning and unknown states correspond to the leveling stage, and the unknown state is a state not recognized by the contact state perception network; leveling, no contact, fitting and engagement correspond to the engagement stage.

3. A robot curtain wall installation method based on a large language model as claimed in claim 2, characterized in that: The contact event is specifically: when the contact force between the curtain wall to be assembled and the main component exceeds a preset threshold, it is determined to be contact.

4. The method for installing a curtain wall using a robot based on a large language model according to claim 2, wherein: The prompt words for the corresponding stage are generated based on the preset STAD prompt word template, wherein: The scene state corresponds to the contact state; Scenario tasks correspond to the leveling phase or the snapping phase; The skill library corresponds to the posture adjustment library, case library or position control library; The decision-making tasks correspond to the actions in the skill library. Specifically, the posture adjustment library and case library contain the robot actions required in the leveling stage, and the position control library contains the robot actions required in the engagement stage.

5. The method for installing a curtain wall using a robot based on a large language model according to claim 2, wherein: For the leveling phase, if the contact status is unknown: Based on the unknown state, automatically associating historical experience data in the case library, and retrieving similar scenes that match the current pendant structure features through a cosine similarity algorithm; The leveling strategy corresponding to similar historical scenarios is used as a decision-making task; The historical experience data includes six-dimensional state sequences, action logs and pendant structural feature annotations collected through teaching or trial and error, which are stored in the case library and indexed according to the structural features, assembly status and action result dimensions.

6. The method for installing a curtain wall using a robot based on a large language model according to claim 1, wherein: Verifying the feasibility of the decision instructions through digital twin simulation, driving the robot to execute the decision instructions that meet the feasibility, and completing the assembly of the curtain wall and the main components specifically include: The feasibility of the decision-making instructions is verified through digital twin simulation. If the verification is passed and the current stage is the leveling stage, the system automatically enters the engagement stage and repeats the prompt word generation, large language model decision and simulation verification steps; if the verification is passed and the current stage is the engagement stage, the robot is driven to execute the decision-making instructions to complete the assembly of the curtain wall and the main components.

7. The method for installing a curtain wall using a robot based on a large language model according to claim 1, wherein: The digital twin simulation verification is specifically as follows: Simulating the execution of the decision instructions in a digital twin simulation environment; Dynamically evaluate the mechanical response and posture changes during simulation execution to detect risks such as excessive assembly force, increased posture error, or abnormal contact. If the simulation results do not meet the assembly stability requirements, the action parameters are corrected based on the mechanical response and posture data feedback to obtain the corrected decision instructions; The revised decision instructions are then re-input into the large language model to iteratively optimize subsequent decision strategies.

8. A robotic curtain wall installation system based on a large language model, characterized in that: include: The prompt word generation module is configured to generate prompt words for the corresponding stage based on a preset STAD prompt word template according to the current assembly stage of the curtain wall to be assembled; the assembly stage includes a leveling stage or a locking stage; the STAD prompt word template includes a scene state, a scene task, a skill library, and a decision task; A decision generation module is configured to input the prompt word into a pre-trained large language model to generate a decision instruction; The decision execution module is configured to verify the feasibility of the decision instructions through digital twin simulation, drive the robot to execute the decision instructions that meet the feasibility, and complete the assembly of the curtain wall and the main parts.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, the steps of the robot curtain wall installation method based on a large language model as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the robot curtain wall installation method based on a large language model as described in any one of claims 1 to 7 are implemented.

Images