Cleaning robot closed-loop variable impedance control method based on large language model parameter tuning
Patent Information
- Application Number
- CN202610734356.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-26
- Publication Date
- 2026-08-18
AI Technical Summary
然而，传统阻抗控制的惯性、阻尼和刚度参数通常需要依靠工程师的经验进行繁琐的手动试凑，且参数一旦确定便无法适应环境刚度和摩擦力的动态变化
[0029]本发明构建出清洁机器人的自适应阻抗控制模型，根据接触式作业反馈的实际接触力，基于预设物理语义映射规则构建大语言模型提示词，输入大语言模型得到优化后阻抗参数并更新自适应阻抗控制模型；利用大语言模型的自然语言理解与物理法则推理能力，唤醒机器人的智能，使其在复杂不确定环境下实现免训练的阻抗参数自主演化；消除了传统控制中人工调参的依赖，提高清洁机器人在高刚度接触环境下的瞬态抗冲击能力与稳态力跟踪精度。
Smart Images
Figure CN122593009A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot intelligent control technology, and in particular to a closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning. Background Technology
[0002] With the widespread application of service robots and industrial robots in modern society, robots are increasingly performing contact tasks in scenarios such as glass cleaning and curved surface polishing. Performing such tasks not only requires the robotic arm to have precise trajectory tracking capabilities, but also requires its end effector to maintain a smooth contact force with the environment to prevent damage to the environmental surface or the robotic arm.
[0003] Traditional impedance control techniques are key to achieving compliant force control in robots. However, the inertia, damping, and stiffness parameters of traditional impedance control typically require tedious manual trial and error based on engineers' experience, and once determined, these parameters cannot adapt to dynamic changes in environmental stiffness and friction. In recent years, although adaptive impedance parameter methods based on neural networks or reinforcement learning (RL) have emerged, these methods generally suffer from drawbacks such as long training times, the need for massive amounts of interactive data, difficulties in designing reward functions, and a lack of explicit physical interpretability.
[0004] The emergence of Large Language Models (LLMs) has endowed robots with powerful contextual understanding and zero-shot reasoning capabilities. Therefore, how to leverage the natural language understanding and physical law reasoning capabilities of LLMs to awaken the intelligence of robots and enable them to achieve training-free autonomous evolution of impedance parameters in complex and uncertain environments is an important research direction for improving the force control performance and human-robot co-operation safety of cleaning robots. Summary of the Invention
[0005] To address the above problems, this invention proposes a closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning. An adaptive impedance control model for the cleaning robot is constructed. Based on the actual contact force feedback from contact-type operations, prompt words for the large language model are built according to preset physical semantic mapping rules. These prompts are input into the large language model to obtain optimized impedance parameters and update the adaptive impedance control model. This eliminates the reliance on manual parameter tuning in traditional control methods. By utilizing the natural language understanding and physical law reasoning capabilities of the large language model, the robot's intelligence is awakened, enabling it to achieve training-free autonomous evolution of impedance parameters in complex and uncertain environments.
[0006] The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning is described in the following steps:
[0007] S1, Construct an adaptive impedance control model for the cleaning robot in a contact-type operation scenario; the adaptive impedance control model uses impedance parameters as adjustable parameters, its input is the force error signal after disturbance rejection compensation, and its output is the position deviation.
[0008] S2 generates the desired pose through trajectory planning, and controls the cleaning robot to perform contact operations by combining the position deviation generated by the adaptive impedance control model.
[0009] S3 calculates the maximum force overshoot and steady-state error based on the actual contact force feedback from the contact operation; based on the preset physical semantic mapping rules, constructs prompt words to guide the output of the optimized impedance parameters of the large language model according to the impedance parameters, maximum force overshoot and steady-state error of the current adaptive impedance control model;
[0010] S4. Input the prompt word into the large language model to obtain the optimized impedance parameters. Use the optimized impedance parameters to update the adaptive impedance control model. Repeat S3 and S4 until the iteration termination condition is reached.
[0011] Preferably, the desired pose is generated through trajectory planning, and a cosine smooth transition curve is used as the flexible and safe approximation trajectory; the cosine smooth transition curve is expressed as:
[0012] ;
[0013] in, The pose at time t; This is the initial safe suspension position. For the target contact location, This is the set approximate time constant.
[0014] Preferably, the adaptive impedance control model is expressed as:
[0015] ;
[0016] in, This is the positional deviation. For speed deviation, This refers to the acceleration deviation. This represents the force error signal after disturbance rejection compensation. To actually come into contact with external forces, For the desired contact force, This represents the total disturbance. , and Indicates the impedance parameter. For the desired quality, For desired damping, Let be the desired stiffness.
[0017] Preferably, the total disturbance is obtained by using an extended state observer in linear active disturbance rejection control to subtract the actual contact force from the expected force obtained from trajectory planning, thereby obtaining a real-time estimated total disturbance.
[0018] Preferably, the step of constructing prompt words to guide the output of optimized impedance parameters by the large language model based on preset physical semantic mapping rules, according to the impedance parameters, maximum force overshoot, and steady-state error of the current adaptive impedance control model, specifically includes:
[0019] Rigid excess semantic mapping rule: if the maximum force overshoot If the stiffness exceeds a preset excess threshold, a first prompt word is generated to indicate to the large language model that the expected stiffness of the current adaptive impedance control model "needs to be significantly reduced". At the same time, "reduce" the expected quality. To reduce system inertia;
[0020] Underdamped semantic mapping rule: If high-frequency force oscillations occur during the actual contact force response process, a second cue word is constructed to prompt the large language model to separately "increase" the expected damping of the current adaptive impedance control model. To absorb vibration energy by utilizing the shock absorber function;
[0021] Insufficient support semantic mapping rule: If the steady-state error e ss If the value remains too high, a third cue word is constructed to suggest "appropriately increasing" the expected stiffness of the current adaptive impedance control model. With expected damping To enhance force tracking accuracy.
[0022] Preferably, the iteration termination condition is as follows:
[0023] A cost function is constructed using the maximum force overshoot and steady-state error. The iteration terminates when the calculated cost function value is less than a set optimal threshold; or, when the maximum force overshoot... The overshoot is less than the preset maximum force target threshold, and the steady-state error is... The iteration terminates when the error falls below the preset steady-state error target threshold.
[0024] Preferably, the cost function is expressed as:
[0025] ;
[0026] in, The cost function; for Weighting coefficients; for Weighting coefficients; This is the maximum force overshoot. This represents the steady-state error.
[0027] Preferred, Value greater than .
[0028] Preferably, after inputting the large model prompt words into the large language model to obtain the optimized impedance parameters, the method further includes: performing boundary constraint checks on the optimized impedance parameters, specifically: setting physical hard constraint boundaries for the optimized impedance parameters output by the large language model to limit the expected quality. Expected damping and desired stiffness The exploration range is defined, and if the optimized output impedance parameter exceeds the limit, boundary trimming is forcibly performed. Compared with the prior art, the present invention has the following beneficial effects:
[0029] This invention constructs an adaptive impedance control model for a cleaning robot. Based on the actual contact force feedback from contact-type operations, a large language model prompt word is constructed based on preset physical semantic mapping rules. The prompt word is input into the large language model to obtain optimized impedance parameters and update the adaptive impedance control model. By utilizing the natural language understanding and physical law reasoning capabilities of the large language model, the robot's intelligence is awakened, enabling it to achieve training-free autonomous evolution of impedance parameters in complex and uncertain environments. This eliminates the reliance on manual parameter tuning in traditional control and improves the cleaning robot's transient impact resistance and steady-state force tracking accuracy in high-stiffness contact environments. Attached Figure Description
[0030] The present invention will now be described in further detail with reference to the accompanying drawings;
[0031] Figure 1 This is a flowchart of a closed-loop variable impedance control method for a cleaning robot based on large language model parameter tuning, according to an embodiment of the present invention.
[0032] Figure 2 This is a block diagram of the closed-loop control architecture of the closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to an embodiment of the present invention.
[0033] Figure 3 The diagram shows the flexible approach trajectory curve in the safety protection mechanism of the closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning in this embodiment of the invention; where (a) represents the position cosine smooth approximation process; and (b) represents the velocity bell-shaped transition curve.
[0034] Figure 4 This is a large language model closed-loop three-dimensional parameter space convergence trajectory diagram of the cleaning robot closed-loop variable impedance control method based on large language model parameter tuning according to an embodiment of the present invention.
[0035] Figure 5The graph shows a comparison of the end contact force error response of the cleaning robot before and after the large language model closed-loop optimization of the closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning, as presented in this embodiment of the invention.
[0036] Figure 6 The graph shows the physical limit detection verification curves of the self-disturbance rejection impedance control of the closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning in this embodiment of the invention at different high-speed wiping frequencies. Detailed Implementation
[0037] The present invention will be further described below through specific embodiments.
[0038] like Figure 1 As shown, the closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning has the following specific steps:
[0039] S1. Construct an adaptive impedance control model for cleaning robots in contact-based work scenarios.
[0040] To address environmental uncertainties, high-frequency force feedback noise, and frictional interference in contact-based cleaning robot operations, a cascade variable impedance control model based on low-pass filtering and active disturbance rejection observation compensation is constructed. The adaptive impedance control model uses impedance parameters as adjustable parameters, takes the force error signal after disturbance rejection compensation as input, and outputs the position deviation.
[0041] S11. Cleaning robots experience variable nonlinear frictional forces and unmodeled dynamic disturbances during contact operations such as wiping. This embodiment directly introduces the Extended State Observer (ESO) from Linear Active Disturbance Rejection Control (LADRC) to measure the actual contact force. The total disturbance of the system is estimated in real time by subtracting the expected force. .
[0042] S12, based on the observed disturbance, an adaptive impedance control model considering environmental influences in the force constraint direction is established as follows:
[0043] ;
[0044] in, These are the expected mass parameters, damping parameters, and stiffness parameters of the cleaning robot, respectively. This is the positional deviation in the direction of the force constraint. Desired contact force. Through dynamic adjustment. The resulting set of impedance parameters can alter the compliance of the cleaning robot's end effector. This embodiment abandons manual trial and error, instead allowing a large language model to autonomously optimize this parameter set.
[0045] S2 generates the desired pose through trajectory planning, and controls the cleaning robot to perform contact operations by combining the position deviation generated by the adaptive impedance control model.
[0046] like Figure 2 The control architecture diagram shown in this embodiment consists of an upper-level large language model intelligence layer and a lower-level robot control execution layer. The actual environmental contact force feedback signal is transmitted upward to the state feature extraction module through a closed-loop channel, while the optimal physical parameters issued by the large model inference center are directly injected into the lower-level adaptive variable impedance controller, forming a complete cross-layer closed-loop feedback system.
[0047] S3 calculates the maximum force overshoot and steady-state error based on the actual contact force feedback from the contact operation; based on the preset physical semantic mapping rules, constructs prompt words to guide the output of the optimized impedance parameters of the large language model according to the impedance parameters, maximum force overshoot and steady-state error of the current adaptive impedance control model.
[0048] To enhance the robot's intelligence and optimize the impedance parameter set, a joint cost function is constructed to extract control quality features, while a safety protection mechanism based on a flexible approach trajectory is added.
[0049] To prevent the cleaning robot from generating enormous transient destructive forces when approaching high-rigidity glass surfaces at high speeds, a safety mechanism is added to the S31. A flexible, safe approach trajectory (Guard Approach) is designed, reshaping the traditional step position command into a smooth cosine transition curve. ;
[0050] in, This is the initial safe suspension position. Let τ be the target contact position, and τ be the set approach time constant. Taking the derivative of equation (3), it can be seen that this mechanism ensures that the transient velocity and acceleration of the cleaning robot at the moment of contact with the surface are close to zero.
[0051] Combination Figure 3 As shown in the flexible approach trajectory curve, in the upper part (a), the desired position command at the end exhibits a smooth approach process; while in the lower part (b), the desired velocity command forms a perfect bell-shaped transition curve. At the instant of contact with the high-stiffness physical boundary, the velocity command is smoothly suppressed to zero, thereby completely weakening the potentially destructive impact energy from the kinetic energy source.
[0052] S32, constructing a joint cost function for parameter tuning of a large language model. Extracting core state features characterizing cleaning operation quality, including dynamic contact impact and steady-state wiping accuracy: ;
[0053] in, This is the maximum force overshoot at the moment of contact; This is the steady-state error; Assign weighting coefficients. Considering the extremely low tolerance of glass wiping to impact, set... Greater than This guides large language models to prioritize finding impedance parameters that achieve the ultimate compliance yielding.
[0054] This paper utilizes a large language model to solve the optimization problem in a high-dimensional continuous parameter space. Zero-shot closed-loop inference of the impedance control strategy is driven by robot physical cues (Prompts) and semantic mapping rules. Unlike traditional reinforcement learning that relies on massive amounts of data to update neural network weights, this embodiment directly uses a large language model as an intelligent optimizer. The specific process is as follows:
[0055] Single-round simulation and feature extraction: Initializing the current network Parameters are set, and the underlying simulation model of the cleaning robot is run. After execution, the overshoot for this round is extracted. With steady-state error .
[0056] Semantic mapping and structured cue word construction: The current impedance parameter is concatenated with the force control performance index, and the robot's physical impedance law is embedded. The semantic mapping rules of the physical law are as follows:
[0057] 1) Rule A: If overshoot Excessive stiffness triggers the semantic concept of "excessive rigidity in the cleaning system," suggesting that large language models need to significantly reduce their stiffness. Even approaching zero, while reducing mass. To reduce system inertia;
[0058] 2) Rule B: If high-frequency force oscillations occur during the response process, triggering the semantic meaning of "underdamped cleaning system", it indicates that the large language model needs to increase the damping separately. To absorb vibration energy by utilizing the shock absorber function;
[0059] 3) Rule C: If the steady-state error The consistently large size of the large language model triggers the semantic sense of "insufficient support," suggesting that the large language model should be appropriately improved. and To enhance force tracking accuracy.
[0060] Zero-shot physics inference: The text containing the aforementioned states and rules is input as a Prompt to the large language model API. The large language model infers based on contextual logic and outputs the next-generation optimal impedance parameters in pure JSON format. .
[0061] S4. Input the prompt word into the large language model to obtain the optimized impedance parameters. Use the optimized impedance parameters to update the adaptive impedance control model. Repeat S3 and S4 until the iteration termination condition is reached.
[0062] Adjusting the relevant constraints and evaluation parameters involved in the optimization process ensures that the optimal impedance control strategy can be output, thereby significantly improving the force control performance of the cleaning robot.
[0063] S41 sets the physical hard constraints for the parameter tuning actions of the large language model. To reduce useless exploration time and ensure absolute safety, the mass is set. The exploration range is [0.05, 1.0], damping [5, 500], stiffness The range is [0, 200]. If the output of a large model exceeds the limit, boundary clipping will be forced.
[0064] S42 sets the automatic iteration termination condition. This occurs when the cost function (Cost) is less than the set optimal threshold, or when the overshoot condition is met. Less than the target threshold and steady-state error When the value is extremely small, the intelligent parameter tuning is determined to be converged, the closed-loop iteration process is terminated in advance, and the current optimal impedance parameter set is locked.
[0065] S43 sends the optimal parameters locked by the final inference to the underlying ADRC impedance controller.
[0066] Implementation Results and High-Fidelity Physical Simulation Verification Analysis: To verify the effectiveness and industrial applicability of the closed-loop control strategy proposed in this embodiment in improving the force tracking performance of the robotic arm, this embodiment abandons the idealized pure mathematical model and instead adopts a joint architecture of MATLAB / Simscape Multibody and Python to build a high-fidelity robot physical dynamics simulation platform. The environmental contact mechanics model is set to a maximum of 10. 5 With a high environmental stiffness of N / m, and by introducing a nonlinear friction mechanism, a highly challenging high-stiffness hard contact scenario was constructed.
[0067] The optimization process of large language models is highly interpretable. For example... Figure 4 As shown in the three-dimensional parameter space convergence trajectory diagram, the large language model starts from an extremely rigid high-cost starting point and accurately understands the physical law of "reducing damping and inertia to improve compliance". It drives the parameter coordinates to continuously explore along a smooth downward trajectory in three-dimensional space, and finally falls precisely into the global optimal valley marked in green in the figure. ).
[0068] Comparison of contact force error response before and after closed-loop optimization of large language model Figure 5As shown. In the unoptimized initial state ( =100.0, =0.5, =10.0), the system exhibits extremely high contact stiffness, with the maximum reverse impact force overshoot exceeding 20N at the moment of contact. After initiating intelligent optimization, the large model accurately identifies physical defects and rapidly softens the parameters. Under the optimal impedance strategy ( =15.0, =0.1, At a value of 0.5, the contact overshoot decreased by nearly 50%, and the oscillation converged rapidly. Even more remarkably, the system's steady-state force tracking error decreased from the initial 10... -2 The N-level (0.0100N) converges to 10. -4 The N-level (0.0001N) achieves a steady-state accuracy leap of up to two orders of magnitude, enabling smooth and damage-free high-quality cleaning operations.
[0069] To further investigate the stability boundary of this embodiment under extreme conditions, a high-speed wiping physical limit test was performed on the optimized parameter set after optimization. The results are as follows: Figure 6 As shown. In Figure 6 In the low to medium speed operating range (wiping frequency 1 to 10 rad / s), thanks to the precise dynamic compensation of the ESO for nonlinear disturbances, the red data curve representing the root mean square error (RMSE) of the global force tracking error is almost a horizontal straight line, maintaining an extremely stable low level of 2.2N. Only when the frequency exceeds the system's physical bandwidth limit does the error exhibit a nonlinear upward trend, reaching the system's collapse tolerance threshold at extreme frequencies. This fully verifies that the architecture proposed in this method not only possesses high-precision intelligent optimization capabilities but also excellent high-frequency disturbance rejection and dynamic robustness.
[0070] In summary, the method proposed in this invention does not require complex model training. It can endow cleaning robots with physical intuition and intelligence through only a small number of natural language closed-loop interaction iterations, which significantly improves the robot's compliant operation adaptability and force control accuracy.
[0071] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.
Claims
1. A closed-loop variable impedance control method for a cleaning robot based on large language model parameter tuning, characterized in that, Includes the following steps: S1, Construct an adaptive impedance control model for the cleaning robot in a contact-type operation scenario; the adaptive impedance control model uses impedance parameters as adjustable parameters, its input is the force error signal after disturbance rejection compensation, and its output is the position deviation. S2 generates the desired pose through trajectory planning, and controls the cleaning robot to perform contact operations by combining the position deviation generated by the adaptive impedance control model. S3, calculate the maximum force overshoot and steady-state error based on the actual contact force feedback from the contact operation; Based on the preset physical semantic mapping rules, prompt words are constructed to guide the output of the optimized impedance parameters of the large language model according to the impedance parameters, maximum force overshoot and steady-state error of the current adaptive impedance control model. S4. Input the prompt word into the large language model to obtain the optimized impedance parameters. Use the optimized impedance parameters to update the adaptive impedance control model. Repeat S3 and S4 until the iteration termination condition is reached.
2. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 1, characterized in that, The desired pose is generated through trajectory planning, and a cosine smooth transition curve is used as a flexible and safe approximation trajectory; the cosine smooth transition curve is expressed as follows: ； in, The pose at time t; This is the initial safe suspension position. For the target contact location, This is the set approximate time constant.
3. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 1, characterized in that, The adaptive impedance control model is expressed as follows: ； in, This is the positional deviation. For speed deviation, This refers to the acceleration deviation. This represents the force error signal after disturbance rejection compensation. To actually come into contact with external forces, For the desired contact force, This represents the total disturbance. , and Indicates the impedance parameter. For the desired quality, For desired damping, Let be the desired stiffness.
4. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 3, characterized in that, The total disturbance is obtained by using the extended state observer in linear active disturbance rejection control to subtract the actual contact force from the expected force obtained from trajectory planning, thus obtaining a real-time estimated total disturbance.
5. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 3, characterized in that, The method, based on preset physical semantic mapping rules, constructs prompt words to guide the output of optimized impedance parameters by the large language model according to the impedance parameters, maximum force overshoot, and steady-state error of the current adaptive impedance control model. Specifically, this includes: Rigid excess semantic mapping rule: if the maximum force overshoot If the stiffness exceeds a preset excess threshold, a first prompt word is generated to indicate to the large language model that the expected stiffness of the current adaptive impedance control model "needs to be significantly reduced". At the same time, "reduce" the expected quality To reduce system inertia; Underdamped semantic mapping rule: If high-frequency force oscillations occur during the actual contact force response process, a second cue word is constructed to prompt the large language model to separately "increase" the expected damping of the current adaptive impedance control model. To absorb vibration energy by utilizing the shock absorber function; Insufficient support semantic mapping rule: If the steady-state error e ss If the value remains too high, a third cue word is constructed to suggest "appropriately increasing" the expected stiffness of the current adaptive impedance control model. With expected damping To enhance force tracking accuracy.
6. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 1, characterized in that, The iteration termination condition is specifically as follows: A cost function is constructed using the maximum force overshoot and steady-state error. The iteration terminates when the calculated cost function value is less than a set optimal threshold; or, when the maximum force overshoot... The overshoot is less than the preset maximum force target threshold, and the steady-state error is... The iteration terminates when the error falls below the preset steady-state error target threshold.
7. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 6, characterized in that, The cost function is expressed as: ； in, The cost function; for Weighting coefficients; for Weighting coefficients; This is the maximum force overshoot. This represents the steady-state error.
8. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 7, characterized in that, Value greater than .
9. The closed-loop variable impedance control method for cleaning robots based on large language model parameter tuning according to claim 1, characterized in that, After inputting the large model prompts into the large language model to obtain the optimized impedance parameters, the process further includes: performing boundary constraint checks on the optimized impedance parameters, specifically: setting physical hard constraint boundaries for the optimized impedance parameters output by the large language model to limit the expected quality. Expected damping and desired stiffness The exploration range is limited, and if the optimized impedance parameters of the output exceed the limit, boundary pruning is forcibly performed.