A control method and system for a collaborative robot arm

Abstract

The present application relates to the technical field of robot control, in particular to a control method and system of a collaborative robot mechanical arm. The method comprises: collecting multi-modal observation data such as vision, force and torque, touch, joint state, etc., outputting operation intention, task stage and uncertainty index after time alignment and feature fusion; calculating safety constraints and constructing feasible region according to the above information, adaptively updating impedance parameters and generating control instructions, and executing after constraint projection; at the same time, an online self-calibration mechanism triggered by residual error is introduced to update parameters such as sensing zero offset, friction and TCP. The method can balance the compliance, operation accuracy and safety of human-machine collaboration, reduce the risk of impact, collision and jamming, and improve the long-term operation stability of the mechanical arm.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a control method and system for a collaborative robot arm. Background Technology

[0002] Collaborative robots, in tasks such as assembly, insertion, grinding, and handling, need to achieve safe, compliant, and stable interactive control in complex contact environments. Existing technologies typically employ fixed or rule-based impedance parameter settings, making it difficult to simultaneously address the dynamic requirements of different task stages.

[0003] On the other hand, control based on single perception is subject to uncertainty due to factors such as occlusion, reflection, workpiece tolerance, and environmental changes. If the controller fails to effectively utilize the uncertain information, it is prone to shocks and jamming caused by excessive rigidity, or trajectory deviation and efficiency reduction caused by excessive compliance.

[0004] Therefore, there is an urgent need for a control scheme that can integrate multimodal information to infer intent and contact phases, and incorporate inference uncertainty into impedance parameter updates and safety constraint calculations, so as to improve the robustness and safety of collaborative robots in real working conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a control method and system for a collaborative robot arm, in order to solve the technical problems in the prior art, such as the inability of impedance parameters to be adaptively adjusted with uncertainty, robot parameter drift, lack of safety constraints in control commands, and difficulty in balancing compliance, operational accuracy and operational safety in human-robot collaborative scenarios.

[0006] The technical solution of this invention is: a control method for a collaborative robot arm, comprising:

[0007] Multimodal observation data of the collaborative robot arm is collected, the multimodal observation data is processed, and the data is input into the intention inference model to obtain the operation intention, task stage, and uncertainty index; the multimodal observation data includes visual observation data, force and torque observation data, tactile observation data, and joint state observation data;

[0008] Safety constraints are calculated and feasible regions are constructed based on operational intent, task phase, and uncertainty index.

[0009] The expected curve is generated based on the operational intent and task stage, and the impedance parameters are adaptively updated in combination with the task stage and uncertainty index.

[0010] Based on the updated impedance parameters and the desired curve, control commands are generated, projected onto the feasible region, and the projected control commands are sent to the execution interface to drive the robotic arm of the collaborative robot. The actual response data is read back and compared with the desired curve to obtain the residual.

[0011] Drift detection is performed based on the residual. When it is determined to be in a drift state, online self-calibration of the compensation parameters to be calibrated is triggered. After calibration, the data is written back to the collaborative robot controller. The compensation parameters to be calibrated include at least one of the following: sensor zero bias, friction parameters, and TCP parameters.

[0012] Preferably, the uncertainty index is calculated based on the output of the intention inference model through at least one method:

[0013] The uncertainty index is calculated based on the normalized information entropy of the operational intent and the probability distribution of the task stage.

[0014] Uncertainty index is calculated based on the variance of multiple sub-model outputs;

[0015] Uncertainty indices are obtained based on the Monte Carlo Dropout method.

[0016] Preferably, the adaptive update of the impedance parameter is achieved by comparing the smooth uncertainty index with a first uncertainty threshold and a second uncertainty threshold, and matching different control strategies: the first uncertainty threshold is less than the second uncertainty threshold;

[0017] When the smoothing uncertainty index is less than or equal to the first uncertainty threshold, the normal control strategy is adopted.

[0018] When the smoothing uncertainty index is greater than the second uncertainty threshold, a conservative control strategy is adopted.

[0019] When the smoothing uncertainty index is less than or equal to the second uncertainty threshold and greater than the first uncertainty, the impedance parameter is updated by interpolation.

[0020] Preferably, the normal control strategy includes:

[0021] Stage reference parameters are determined based on the task stage; the stage reference parameters include stiffness reference parameters and damping reference parameters;

[0022] An uncertainty mapping function is constructed based on the smoothing uncertainty index, the first uncertainty threshold, and the second uncertainty threshold. The stiffness correction and damping correction are calculated by combining the stiffness adjustment coefficient and the damping adjustment coefficient, respectively.

[0023] Based on the stage baseline parameters, stiffness correction and damping correction, combined with the application of upper and lower limit constraints and the rate of change limit within adjacent control cycles, the updated stiffness and damping are obtained.

[0024] Preferably, the conservative control strategy includes:

[0025] Reduce the upper limit of stiffness to the preset safe stiffness upper limit or reduce the rate of change of stiffness to the preset safe threshold of stiffness change rate, while increasing the lower limit of damping to the preset safe damping threshold or increasing the rate of change of damping to the preset safe threshold of damping change rate.

[0026] The stiffness and damping are updated according to the normal control strategy to reduce the difference between stiffness and safety stiffness, and the difference between damping and safety damping; the safety stiffness and safety damping are preset values.

[0027] Preferably, the method for projecting the control command to the feasible region is as follows: under the constraints of the feasible region, obtain the control command with the smallest deviation from the control command.

[0028] Preferably, a gradual degradation strategy is triggered when the feasible region is empty or the constraint meets the saturation condition;

[0029] The feasible region is empty, meaning there is no control command vector that simultaneously satisfies all safety constraints; the control command vector is an end-effector velocity, pose increment, or joint velocity vector.

[0030] The constraint satisfies the saturation condition when the constraint utilization rate is greater than or equal to the preset saturation threshold, or when the minimum margin of the control command after projection is lower than the preset safety margin threshold.

[0031] Preferably, the gradual degradation strategy includes:

[0032] According to the preset tiered degradation rules, the speed limit and stiffness limit in the control command are reduced. When the feasible domain is empty or the duration of the constraint satisfying the saturation condition is greater than or equal to the preset degradation time threshold, the system switches to the degradation segment.

[0033] In the degradation section, safety stiffness and safety damping are used as impedance parameters. The feasible region is reconstructed based on the safety constraints corresponding to the degradation section, and the control command is projected onto the feasible region.

[0034] Preferably, the drift detection method is as follows:

[0035] When the absolute value of the residual is greater than a preset residual threshold and the duration exceeds a preset duration threshold, it is determined to be a drift candidate state;

[0036] After determining that it is a candidate state of drift, it is further determined whether the collaborative robot is in a preset safety calibration window. The safety calibration window includes a low contact force state, an unloaded state, and a safe posture state.

[0037] When the safety calibration window is in effect, it is determined that the system is in a drift state.

[0038] On the other hand, this application also discloses a control system for a collaborative robot arm, including:

[0039] The sensing unit is used to collect multimodal observation data of the collaborative robot's robotic arm;

[0040] The intent inference model is used for data preprocessing of multimodal observation data, and the operational intent, task stage, and uncertainty index are obtained through the intent inference model.

[0041] The trajectory generation module is used to generate a desired curve based on the operational intent and task stage.

[0042] The impedance control module is used to update the impedance parameters based on the task stage and uncertainty index, and to generate control commands based on the updated impedance parameters.

[0043] The safety constraint module is used to calculate safety constraints and construct a feasible region based on the operation intention, task stage and uncertainty index, and to perform projection processing on the control command;

[0044] The execution interface is used to send the projected control commands to the collaborative robot arm and read back the actual response data.

[0045] The residual calculation module is used to calculate the residual based on the actual response data and the expected curve;

[0046] The online calibration module is used to perform drift detection based on the residual and to perform online self-calibration when drift is detected, and to write the updated compensation parameters to be calibrated back to the collaborative robot controller.

[0047] Compared with the prior art, the advantages of the present invention are:

[0048] (1) This invention uses an uncertainty-driven adaptive adjustment of impedance parameters as the core design, and dynamically adjusts the control strategy based on the reliability inferred from the multimodal intent. In high uncertainty scenarios, it strengthens safety constraints and reduces impact risks, while in low uncertainty scenarios, it ensures control compliance and operational efficiency. This fundamentally solves the core pain point of existing technologies where compliance and safety are difficult to balance, and adapts to the dynamic working conditions of complex human-machine collaboration.

[0049] (2) The present invention uses an online self-calibration mechanism triggered by residuals to compensate for drift errors such as sensor zero bias and friction parameters in real time, avoiding problems such as control deviation and contact impact caused by parameter distortion. This not only reduces the manpower cost of offline calibration, but also ensures the accuracy of control commands, and indirectly improves the safety and reliability of human-machine collaboration.

[0050] (3) The present invention uses a step-by-step safety degradation mechanism to smoothly adjust control parameters and safety constraints when the system is abnormal, avoiding equipment damage and work interruption caused by violent shutdown, taking into account both safety protection and work continuity, and adapting to the flexible needs of human-machine collaboration. Attached Figure Description

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0052] Figure 1 This is a flowchart of a control method for a collaborative robot arm according to the present invention;

[0053] Figure 2 This is a structural diagram of the control system of a collaborative robot arm according to the present invention. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to specific embodiments:

[0055] In collaborative robotic arm operations involving human-machine interaction, the robotic arm must perform precision operations such as contact assembly and workpiece docking under conditions of close human collaboration, environmental disturbances, and noise and obstruction of sensor signals. This scenario requires ensuring the robotic arm's compliance during interactions with workpieces and humans to avoid rigid impacts, while simultaneously maintaining a safety baseline under uncertain operating conditions, parameter drift, and abnormal commands. Current technologies cannot dynamically adapt control strategies to operational uncertainties. Long-term operation is prone to reduced accuracy due to sensor and model parameter drift, and the lack of safe and controllable command constraints and smooth degradation mechanisms makes it highly susceptible to problems such as insufficient compliance leading to collisions, excessive safety reducing efficiency, or direct shutdown and interruption of operations under abnormal conditions.

[0056] This invention addresses the core issues in collaborative robot scenarios, such as the difficulty in balancing compliant interaction and safety protection, the easy degradation of accuracy during long-term operation, and the insufficient safety backup capability under abnormal working conditions. Through an overall technical solution involving uncertainty adaptive impedance adjustment, residual-triggered online self-calibration, control command feasible domain constraints, and progressive safety degradation, this invention achieves a synergistic balance between compliance, accuracy, and safety in human-robot collaborative scenarios.

[0057] like Figure 1 As shown, a control method for a collaborative robot arm includes:

[0058] S1. Collect multimodal observation data of the collaborative robot arm, process the multimodal observation data and input it into the intention inference model to obtain the operation intention, task stage and uncertainty index; among which, the multimodal observation data includes visual observation data, force and torque observation data, tactile observation data and joint state observation data.

[0059] Specifically, data processing refers to the preprocessing operations performed on multimodal observation data to eliminate invalid noise in the original observation data and to unify the data units and data formats.

[0060] Visual observation data includes RGB images or depth maps acquired by vision sensors, relative pose and distance of targets or obstacles, and alignment deviations between the end effector and the workpiece; force and torque observation data includes six-dimensional data from the end effector. The first three items are the three-axis force components along the tool coordinate system at the end of the robotic arm, and the last three items are the three-axis torque components around the coordinate system. This data is used to accurately characterize the spatial contact force and torque state between the end of the robotic arm and the workpiece, environment, or human body. Tactile observation data includes contact switch status, pressure distribution in the contact area, and contact slip indication. Joint status observation includes joint angle, joint speed, and joint torque.

[0061] The intent inference model is a reasoning model that integrates multimodal perception information, enabling the identification of operational intent and task stages during human-computer collaboration, as well as the quantification of the reliability of corresponding inference results. Operational intent characterizes the goal and interaction method of the current task, including: aligning with a target, inserting or removing, following along a surface (such as grinding / applying glue / polishing), transporting and placing, teaching and following, and exiting. Task stages include approach, initial contact, stable contact, operation, exit, and degradation stages. The degradation stage is used to trigger conservative control or retreat strategies when constraints exceed limits or uncertainties exceed thresholds.

[0062] Uncertainty index is a quantitative indicator that characterizes the reliability of the output results of the intention inference model. It is used to reflect the uncertainty level of the current working environment, sensor data and inference results.

[0063] In this embodiment, the intent inference model includes a feature extraction layer, a multimodal fusion layer, an inference layer, and an uncertainty quantification layer. The feature extraction layer employs multiple parallel feature extraction branches, including convolutional layers, LSTM layers, and multilayer perceptron layers, to extract features from input data of different data types. The multimodal fusion layer spatiotemporally aligns the features extracted from each branch and inputs them into a cross-attention fusion layer. This layer dynamically weights the importance of features from different modalities through an attention mechanism, uncovering correlation features between different modalities and outputting a multimodal fusion feature vector that integrates visual, force and torque, tactile, and joint state information, thus addressing the heterogeneity fusion problem of multi-source data. The inference layer uses a multilayer perceptron and a Softmax probability output layer as the core inference model. It inputs the multimodal fusion feature vector into an MLP for feature dimensionality reduction and nonlinear mapping, and then outputs the operation intent and the probability distribution of the task stage through the Softmax layer. The maximum probability is taken as the final inference result for the operation intent and task stage. The uncertainty quantification layer performs uncertainty quantification based on the probability distribution features output by the inference layer, processing in parallel with the inference layer and outputting synchronously.

[0064] By inputting multimodal observation data into the intent inference model, the model outputs the corresponding operational intent, task stage, and uncertainty index.

[0065] S2. Calculate safety constraints and construct feasible regions based on operational intent, task stage, and uncertainty index.

[0066] Specifically, safety constraints are multi-dimensional quantitative control restrictions adapted to human-machine collaborative operation scenarios. Their control dimensions and benchmark range are determined by the operational intent and task stage, and the quantitative boundary is dynamically adjusted with the uncertainty index. The core purpose is to avoid safety risks such as human-machine collisions, contact impacts, and extreme movements during the operation of the robotic arm.

[0067] The feasible region is a closed valid value space for the control command vector of the collaborative robot. It is constructed by combining the calculated safety constraints with the physical execution limits of the robotic arm itself. It is a quantitative constraint boundary for control commands such as end-effector velocity, pose increment, and joint motion, ensuring that the control commands projected into this space meet the safety requirements and the execution capabilities of the robotic arm. The construction process of both is strongly correlated with the operation intention, task stage, and uncertainty index, realizing dynamic adaptation to the current working conditions and providing a core basis for the safety verification of subsequent control commands.

[0068] In one implementation, based on the identified operational intent, the core control dimensions of safety constraints are determined to ensure that the constraint dimensions match the control objectives. Then, combined with the current task stage, basic constraint benchmarks are set for each control dimension. For example, in the approach stage, the motion speed benchmark is increased and the collision distance benchmark is decreased; in the stable contact stage, the end contact force benchmark is decreased and the pose fine-tuning benchmark is increased, so that the constraint benchmarks conform to the action characteristics of the operation stage. Finally, based on the quantified value of the uncertainty index, the basic constraint benchmarks of each dimension are dynamically scaled and adjusted. The higher the uncertainty index value, the more the constraint boundaries of each dimension are narrowed in the direction of safety, such as reducing the speed limit and reducing the allowable range of contact force. The lower the uncertainty index value, the more the constraint boundaries are relaxed, ultimately obtaining a full-dimensional safety constraint that adapts to the current working conditions. The safety constraints include speed constraints, kinetic energy constraints, contact force constraints, torque constraints, collision distance constraints, etc., and are uniformly represented as linear constraints or convex constraints.

[0069] The construction of the feasible region first determines the core representation form of the collaborative robot's control commands, and selects a six-dimensional control command vector that matches the task control as the modeling object of the feasible region.

[0070] The aforementioned full-dimensional safety constraint set is transformed into numerical constraints for each dimension of the control command vector. For example, velocity constraints are transformed into upper and lower limits for the velocity dimension of the command vector, and contact force constraints are transformed into the range of values ​​for the stress control dimension. Subsequently, the inherent physical execution limits of the robotic arm itself, including joint angle limits, maximum output torque, and maximum end-effector velocity, are incorporated into the numerical constraints for each dimension. Finally, through linear modeling in the spatial domain, all dimensional constraints are integrated to construct a closed and continuous six-dimensional quantized feasible domain. The feasibility domain is then validated in real time to ensure that it is a non-empty valid value space, providing a reliable quantized constraint space for subsequent projection operations of control commands.

[0071] In this embodiment, the feasible region is represented as Ω={u|A·u≤b}, where A is the constraint coefficient matrix obtained by discretizing the control command vector u at the current moment based on the safety constraints. The control command vector u is a standardized vector expression of the control commands of the collaborative robot arm. Depending on the actual situation, one of the end effector velocity, pose increment, or joint velocity is selected as the control command vector. It should be noted that the control command vector is an empty vector framework with physical meaning and fixed dimensions. It only defines which dimensions the control command should control the robot arm from, but does not have specific values. b is the corresponding constraint upper bound vector, a standardized vector expression that defines the upper limit of legal values ​​for each dimension of the control command vector u. It follows a strict matching principle with the control command vector. Whichever of the end effector velocity, pose increment, or joint velocity vector the control command vector is selected as, b is the vector of the corresponding dimension. Each element of b corresponds to the maximum legal value upper limit of the control command vector in the same position dimension. It is the core carrier for quantifying the upper bound constraints of each dimension when constructing the feasible region.

[0072] S3. Generate the desired curve based on the operational intent and task stage, and adaptively update the impedance parameters in combination with the task stage and uncertainty index.

[0073] Specifically, the expected curve is a continuous quantitative reference curve that adapts to the current operation intention and task stage. It includes the expected force curve and the expected pose curve, which are used to characterize the target state that the robotic arm end needs to track during the operation. Its curve type matches the operation intention, the curve segments correspond one-to-one with the task stage, and the dimension is consistent with the representation form of the control command vector. It is the core tracking target of impedance control.

[0074] Impedance parameters are the core quantitative parameters for impedance control of collaborative robots. They include at least stiffness and damping parameters. Adaptive updates are based on the operational characteristics of the task stage and use uncertainty indices as the basis for dynamic adjustment. The impedance parameters are quantitatively corrected in real time and continuously to achieve the goal of enhancing compliance under high uncertainty and ensuring operational accuracy under low uncertainty. This update logic is strongly correlated with the operational intent, task stage, and uncertainty indices.

[0075] In one implementation, the type of the desired curve and the target tracking dimension are determined based on the identified operational intent, such as a pose tracking desired curve for an alignment operation and a stress tracking desired curve for an insertion operation. In this implementation, the desired force curve and desired pose curve are derived from a pre-calibrated trajectory library, and the corresponding desired force curve and desired pose curve are searched in the trajectory library according to the operational intent and task stage.

[0076] For adaptive updating of impedance parameters, first determine the stiffness reference parameters and damping reference parameters corresponding to the current task stage, and then dynamically correct the reference parameters according to the quantized value of the uncertainty index. The correction process ensures that the impedance parameters are within the allowable range of the robotic arm and that the parameter change rate meets the control stability requirements, thus obtaining the updated impedance parameters.

[0077] S4. Generate control commands based on the updated impedance parameters and the desired curve, project the control commands onto the feasible region, obtain the projected control commands and send them to the execution interface to drive the robotic arm of the collaborative robot, read back the actual response data and compare it with the desired curve to obtain the residual.

[0078] Specifically, the control command is a quantized command generated based on the updated impedance parameters and the desired curve, used to drive the collaborative robot arm to perform operational actions; the feasible region is the effective value space of the control command constructed by safety constraints, used to ensure that the control command meets safety requirements and the robot arm's execution capability; projection is the operation of mapping the originally generated control command into the feasible region to obtain a safe control command that satisfies all safety constraints; the execution interface is an interactive module used to send control commands to the robot arm and read back the robot arm's actual operational response data; the residual is the difference between the robot arm's actual response data and the desired curve, used for subsequent drift detection and online self-calibration trigger judgment.

[0079] In one implementation, based on the updated impedance parameters and the desired curve, an original control command is generated using an impedance control algorithm. This original control command is then projected onto the constructed feasible region, and a quadratic programming optimization problem is solved to obtain a projected control command that minimizes the deviation from the original control command and satisfies all safety constraints. The projected control command is then sent to an execution interface, which in turn sends the command to the collaborative robot arm to drive it to perform the corresponding task. Simultaneously, the actual response data of the robot arm is read back through the execution interface, and the actual response data is compared with the desired curve dimension by dimension to calculate the residual between the two.

[0080] S5. Drift detection is performed based on the residual. When it is determined that the device is in a drift state, online self-calibration of the compensation parameters to be calibrated is triggered. After calibration, the data is written back to the collaborative robot controller. The compensation parameters to be calibrated include at least one of the following: sensor zero bias, friction parameters, and TCP parameters.

[0081] Specifically, residual is the quantified deviation between the actual operational response of the robotic arm and the expected tracking curve, and it is a core indicator reflecting the deviation of the system's sensing and model parameters from the baseline state. Drift detection is a verification process based on the statistical characteristics of this residual and preset judgment rules to identify whether the system parameters have shifted. The parameters to be calibrated and compensated are key parameters in collaborative robot control that are prone to shifting with runtime and environmental changes. These parameters include sensor zero bias, friction parameters, TCP parameters, etc. The accuracy of these parameters directly affects the control precision and operational stability of the robotic arm. Online self-calibration is a parameter correction operation that can be completed without stopping the system during operation. After calibration, the updated parameters are written to the collaborative robot controller to achieve real-time parameter compensation and ensure the consistency of the system's long-term operation.

[0082] In one implementation, a statistical judgment threshold and duration requirement for the residuals are first set. The residuals calculated in real time are statistically analyzed using a sliding window. When the absolute value of the residuals continuously exceeds the preset threshold and the duration reaches the judgment duration, it is initially judged as a candidate state of drift. Then, it is checked whether the system is in a safe calibration window. This window must meet the requirement that the robotic arm is in a low contact force, no load, or safe posture operation state to avoid affecting the operation safety during the calibration process. If the safe window conditions are met, the system is judged to be in a drift state, triggering online self-calibration: for the sensor zero bias, the recursive least squares method is used to refit and correct the zero-position data collected by the sensor; for the friction parameters, the torque and velocity characteristics of the joint motion are collected through micro-excitation actions to update the friction model parameters; for the TCP parameters, the contact pose data between the end effector and the calibration part are used to complete the recalibration of the TCP coordinates; after all parameters are calibrated, the updated parameters are written to the collaborative robot controller in batches to overwrite the original offset parameters and complete the parameter compensation.

[0083] In summary, this invention effectively avoids safety risks in human-machine collaboration and improves operational safety by constructing dynamic safety constraints and feasible regions based on operational intent, task stage, and uncertainty indicators. By adaptively updating impedance parameters, it enhances the compliance of the robotic arm in high-uncertainty scenarios to avoid rigid impacts and jamming, while ensuring operational accuracy in low-uncertainty scenarios, thus solving the problem that fixed impedance parameters cannot simultaneously guarantee compliance and accuracy. Furthermore, by projecting control commands into the feasible region, it ensures the safety and reliability of commands. Combined with an online self-calibration mechanism driven by residuals, parameter offsets can be corrected without stopping the machine, ensuring the consistency and stability of the robotic arm's long-term operation.

[0084] To elaborate on the technical details of this invention, this application will further explain the calculation method for the uncertainty index:

[0085] The uncertainty index is calculated based on the output of the intention inference model through at least one method:

[0086] The uncertainty index is calculated based on the normalized information entropy of the operational intent and the probability distribution of the task stage.

[0087] Uncertainty index is calculated based on the variance of multiple sub-model outputs;

[0088] Uncertainty indices are obtained based on the Monte Carlo Dropout method.

[0089] Specifically, the uncertainty index is a parameter used to quantify the reliability of the output results of the intention inference model. Its calculation is based on the model's output data. Different calculation methods reflect the uncertainty level of perceived information and inference results in the current working scenario from different dimensions, providing a quantitative basis for subsequent impedance parameter adjustment and safety constraint construction.

[0090] In one implementation, the uncertainty index is calculated based on the normalized information entropy of the operation intention and the probability distribution of the task stage. First, the probability distribution of the operation intention output by the intention inference model is obtained. and task phase probability distribution Calculate the information entropy of each. and The formula for calculating information entropy is:

[0091] ,

[0092] Normalize the information entropy by dividing it by the logarithm of the number of corresponding categories to obtain the uncertainty corresponding to the operational intent. Uncertainty corresponding to the task phase ,in The number of categories of operational intent, The number of categories in the task phase is used, and the maximum value of the two is taken as the current uncertainty index.

[0093] The adaptive update of the impedance parameter is achieved by comparing the smooth uncertainty index with the first uncertainty threshold and the second uncertainty threshold, and matching different control strategies: where the first uncertainty threshold is less than the second uncertainty threshold.

[0094] When the smoothing uncertainty index is less than or equal to the first uncertainty threshold, the normal control strategy is adopted.

[0095] When the smoothing uncertainty index is greater than the second uncertainty threshold, a conservative control strategy is adopted.

[0096] When the smoothing uncertainty index is less than or equal to the second uncertainty threshold and greater than the first uncertainty, the impedance parameter is updated by interpolation.

[0097] Normal control strategies include:

[0098] Stage reference parameters are determined based on the task stage; the stage reference parameters include stiffness reference parameters and damping reference parameters;

[0099] An uncertainty mapping function is constructed based on the smoothing uncertainty index, the first uncertainty threshold, and the second uncertainty threshold. The stiffness correction and damping correction are calculated by combining the stiffness adjustment coefficient and the damping adjustment coefficient, respectively.

[0100] Based on the stage baseline parameters, stiffness correction and damping correction, combined with the application of upper and lower limit constraints and the rate of change limit within adjacent control cycles, the updated stiffness and damping are obtained.

[0101] Conservative control strategies include:

[0102] Reduce the upper limit of stiffness to the preset safe stiffness upper limit or reduce the rate of change of stiffness to the preset safe threshold of stiffness change rate, while increasing the lower limit of damping to the preset safe damping threshold or increasing the rate of change of damping to the preset safe threshold of damping change rate.

[0103] The stiffness and damping are updated according to the normal control strategy to reduce the difference between stiffness and safety stiffness, and the difference between damping and safety damping; the safety stiffness and safety damping are preset values.

[0104] Specifically, the adaptive update of the impedance parameter is a process of dynamically switching the control strategy based on the matching result of the smoothed uncertainty index and the preset threshold. The smoothed uncertainty index is a stable value obtained after filtering the original uncertainty index. The first and second uncertainty thresholds are preset judgment boundaries based on the safety requirements of the operation scenario and the system characteristics. Different control strategies are adapted to low, medium and high uncertainty operation scenarios respectively, so as to ensure operation safety while taking into account operation accuracy and compliance.

[0105] In one implementation, the original uncertainty index is first smoothed using an exponential moving average to obtain a smoothed uncertainty index. :

[0106] ,

[0107] in, This is the smoothing coefficient.

[0108] Preset first uncertainty threshold Second uncertainty threshold ,and Less than ;when Less than or equal to At this time, the normal control strategy is adopted: first determine the stiffness reference parameters corresponding to the current task stage s. With damping reference parameters Construct the uncertainty mapping function:

[0109] ,

[0110] Combined with stiffness adjustment coefficient and damping adjustment coefficient The stiffness correction is calculated as follows Damping correction amount is To obtain the initial updated stiffness Damping Then, upper and lower bound constraints and rate of change constraints of adjacent control cycles are applied to both to obtain the final updated stiffness and damping.

[0111] when Greater than At that time, a conservative control strategy is adopted: the upper limit of stiffness value is reduced to the preset safe stiffness upper limit. The lower limit of the damping value is increased to the preset safety damping lower limit. Simultaneously, stiffness and damping are adjusted according to the update logic of the normal control strategy, so that the stiffness is adjusted towards... Approaching, damping direction Approaching.

[0112] when Less than or equal to And greater than When the impedance parameters are updated, an interpolation method is used. That is, the baseline parameters of the normal control strategy and the safety parameters of the conservative control strategy are interpolated through the uncertainty mapping function to obtain the current impedance parameters.

[0113] The method for projecting control commands onto the feasible region is as follows: under the constraints of the feasible region, obtain the control command with the smallest deviation from the control command.

[0114] Specifically, projecting control instructions to the feasible domain involves mapping the original, unverified control instructions to operations within a legal value space constructed by safety constraints. The core objective of this operation is to preserve the control intent of the original control instructions as much as possible while satisfying all safety constraints, avoiding excessive deviation of control instructions due to safety constraints, and ensuring the continuity and accuracy of operations.

[0115] In one implementation, the feasible region is first represented as a linear constraint set Ω = {u|A·u≤b}, where A is the constraint coefficient matrix, obtained by discretizing the control command vector u using safety constraints such as velocity constraints, kinetic energy constraints, contact force and torque constraints, and collision distance constraints, and b is the corresponding upper bound vector of the constraints. The original control command is... By solving the quadratic programming optimization problem Receive the original control command Minimum deviation post-projection control command This instruction satisfies all safety constraints and can be directly sent to the execution interface to drive the robotic arm to operate.

[0116] When the feasible region is empty or the constraint meets the saturation condition, the gradual degradation strategy is triggered.

[0117] When the feasible domain is empty, it means that there is no control command vector that simultaneously satisfies all safety constraints; the control command vector is the end-effector velocity, pose increment, or joint velocity vector.

[0118] The constraint satisfies the saturation condition when the constraint utilization rate is greater than or equal to the preset saturation threshold, or when the minimum margin of the control command after projection is lower than the preset safety margin threshold.

[0119] Gradual degradation strategies include:

[0120] According to the preset tiered degradation rules, the speed limit and stiffness limit in the control command are reduced. When the feasible domain is empty or the duration of the constraint satisfying the saturation condition is greater than or equal to the preset degradation time threshold, the system switches to the degradation segment.

[0121] In the degradation section, safety stiffness and safety damping are used as impedance parameters. The feasible region is reconstructed based on the safety constraints corresponding to the degradation section, and the control command is projected onto the feasible region.

[0122] Specifically, the progressive degradation strategy is a safety fallback strategy triggered when the feasible region is empty or the constraints are saturated. An empty feasible region means that there are currently no control instructions that meet all safety constraints, while constraint saturation means that the current control instructions are close to the limits of the safety constraints. This strategy reduces control parameters in stages and switches to a conservative control mode to maintain the system's operational capabilities as much as possible while ensuring operational safety and avoiding operational interruptions caused by direct shutdown.

[0123] In one implementation, a saturation threshold is first preset. Safety margin threshold Degradation time threshold And the reduction in the upper limits of speed and stiffness for each stage of degradation. The real-time calculation constraint utilization rate is:

[0124] ,

[0125] The minimum margin after projection is:

[0126] ,

[0127] when Greater than or equal to Sometimes, or Less than or equal to If the constraint is deemed saturated, or if the feasible region is empty, progressive degradation is triggered. Progressive degradation specifically involves: first, reducing the upper limits of the speed and stiffness of the control commands according to preset tiered rules, reconstructing the feasible region, and projecting the control commands; if the feasible region remains empty or the constraint saturation state persists for a certain period... Then switch to the degradation stage and switch the impedance parameter to the preset safety stiffness. With safety damping At the same time, the feasible domain is reconstructed using stricter downgraded segment safety constraints, and control commands are projected onto this feasible domain. At this time, the system enters a conservative control mode to prioritize operational safety.

[0128] The method for drift detection is as follows:

[0129] When the absolute value of the residual is greater than the preset residual threshold and the duration exceeds the preset duration threshold, it is determined to be a drift candidate state;

[0130] After determining that it is a candidate state for drift, it is further determined whether the collaborative robot is in a preset safety calibration window. The safety calibration window includes low contact force state, no-load state and safe posture state.

[0131] When in the safety calibration window, it is determined to be in a drift state.

[0132] Specifically, drift detection is a verification process used to identify whether there is a shift in the parameters of sensors, models, etc. in a collaborative robot system. It judges whether the system parameters deviate from the baseline state by the change characteristics of the residual. The verification of the safety calibration window is to ensure that the calibration operation is triggered in a safe operating state, so as to avoid the calibration process from affecting the safety of the operation or causing danger.

[0133] In one implementation, a residual threshold is first preset. Duration threshold Real-time calculation of the actual response of the robotic arm With the expected curve residual ,when Greater than And the duration reached When the system is in a state of drift, it is determined to be a candidate state. Then, the current working state of the robotic arm is detected. If the robotic arm is in a state of low contact force, no load (no load at the end) or safe posture (joint angle within the preset safe range), it is determined to be in a safe calibration window. At this time, the system is determined to be in a drift state and online self-calibration operation is triggered.

[0134] For sensor zero-bias calibration, the recursive least squares method is used to refit and correct the zero-point data acquired by the sensor, eliminating the zero-point offset error of the sensor. After correction, the new zero-bias parameters are written into the parameter storage module of the robot controller.

[0135] For friction parameter calibration, the robotic arm joints are controlled to perform small-amplitude micro-excitation movements, and the torque and velocity characteristics data during the joint movement are collected. The Coulomb friction coefficient and viscous friction coefficient of the friction model are then updated to make the friction model more consistent with the actual joint friction characteristics.

[0136] For TCP parameter calibration: using the multi-point contact pose data between the end effector and the preset calibration component, the new TCP (tool center point) coordinates are obtained through coordinate transformation, and the TCP parameters are recalibrated to ensure the pose control accuracy of the end effector.

[0137] After all parameters are calibrated, the updated parameters are written to the collaborative robot controller in batches, overwriting the original offset parameters, thus completing this online self-calibration. The system then returns to normal operation.

[0138] The present invention also provides a control system for a collaborative robot arm, such as... Figure 2 As shown, it includes:

[0139] The sensing unit is used to collect multimodal observation data of the collaborative robot's robotic arm;

[0140] The intent inference model is used for data preprocessing of multimodal observation data, and the operational intent, task stage, and uncertainty index are obtained through the intent inference model.

[0141] The trajectory generation module is used to generate the desired curve based on the operation intention and task stage;

[0142] The impedance control module is used to update impedance parameters based on the mission phase and uncertainty index, and to generate control commands based on the updated impedance parameters.

[0143] The safety constraint module is used to calculate safety constraints and construct feasible regions based on operational intent, task stage, and uncertainty index, and to perform projection processing on control commands.

[0144] The execution interface is used to send the projected control commands to the collaborative robot arm and read back the actual response data.

[0145] The residual calculation module is used to calculate the residual based on the actual response data and the expected curve;

[0146] The online calibration module is used to perform drift detection based on residuals and perform online self-calibration when drift is detected, writing the updated compensation parameters to be calibrated back to the collaborative robot controller.

[0147] The above embodiments are merely illustrative of the technical concept and features of the present invention, intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and should not be construed as limiting the scope of protection of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and thus all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

Claims

1. A control method for a collaborative robot arm, characterized in that, include: Multimodal observation data of the collaborative robot arm is collected, the multimodal observation data is processed, and the data is input into the intention inference model to obtain the operation intention, task stage, and uncertainty index; the multimodal observation data includes visual observation data, force and torque observation data, tactile observation data, and joint state observation data; Safety constraints are calculated and feasible regions are constructed based on operational intent, task phase, and uncertainty index. The expected curve is generated based on the operational intent and task stage, and the impedance parameters are adaptively updated in combination with the task stage and uncertainty index. Based on the updated impedance parameters and the desired curve, control commands are generated, projected onto the feasible region, and the projected control commands are sent to the execution interface to drive the robotic arm of the collaborative robot. The actual response data is read back and compared with the desired curve to obtain the residual. Drift detection is performed based on the residual. When a drift state is detected, online self-calibration of the compensation parameters to be calibrated is triggered. After calibration, the data is written back to the collaborative robot controller. The compensation parameters to be calibrated include at least one of the following: sensor bias, friction parameters, and TCP parameters. The adaptive update of the impedance parameter is achieved by comparing the smooth uncertainty index with a first uncertainty threshold and a second uncertainty threshold, and matching different control strategies: the first uncertainty threshold is less than the second uncertainty threshold; When the smoothing uncertainty index is less than or equal to the first uncertainty threshold, the normal control strategy is adopted. When the smoothing uncertainty index is greater than the second uncertainty threshold, a conservative control strategy is adopted. When the smoothing uncertainty index is less than or equal to the second uncertainty threshold and greater than the first uncertainty, the impedance parameter is updated by interpolation. The normal control strategy includes: Stage reference parameters are determined based on the task stage; the stage reference parameters include stiffness reference parameters and damping reference parameters; An uncertainty mapping function is constructed based on the smoothing uncertainty index, the first uncertainty threshold, and the second uncertainty threshold. The stiffness correction and damping correction are calculated by combining the stiffness adjustment coefficient and the damping adjustment coefficient, respectively. Based on the stage reference parameters, stiffness correction and damping correction, combined with the application of upper and lower limit constraints and the rate of change limit in adjacent control cycles, the updated stiffness and damping are obtained. The conservative control strategy includes: Reduce the upper limit of stiffness to the preset safe stiffness upper limit or reduce the rate of change of stiffness to the preset safe threshold of stiffness change rate, while increasing the lower limit of damping to the preset safe damping threshold or increasing the rate of change of damping to the preset safe threshold of damping change rate. The stiffness and damping are updated according to the normal control strategy to reduce the difference between stiffness and safety stiffness, and the difference between damping and safety damping; the safety stiffness and safety damping are preset values.

2. The control method for a collaborative robot arm according to claim 1, characterized in that, The uncertainty index is calculated based on the output of the intention inference model through at least one method: The uncertainty index is calculated based on the normalized information entropy of the operational intent and the probability distribution of the task stage. Uncertainty index is calculated based on the variance of multiple sub-model outputs; Uncertainty indices are obtained based on the Monte Carlo Dropout method.

3. The control method for a collaborative robot arm according to claim 1, characterized in that, The method for projecting control commands onto the feasible region is as follows: under the constraints of the feasible region, obtain the control command with the smallest deviation from the control command.

4. The control method for a collaborative robot arm according to claim 1, characterized in that, When the feasible region is empty or the constraint meets the saturation condition, the gradual degradation strategy is triggered. The feasible region is empty, meaning there is no control command vector that simultaneously satisfies all safety constraints; the control command vector is an end-effector velocity, pose increment, or joint velocity vector. The constraint satisfies the saturation condition when the constraint utilization rate is greater than or equal to the preset saturation threshold, or when the minimum margin of the control command after projection is lower than the preset safety margin threshold.

5. The control method for a collaborative robot arm according to claim 4, characterized in that, The gradual degradation strategy includes: According to the preset tiered degradation rules, the speed limit and stiffness limit in the control command are reduced. When the feasible domain is empty or the duration of the constraint satisfying the saturation condition is greater than or equal to the preset degradation time threshold, the system switches to the degradation segment. In the degradation section, safety stiffness and safety damping are used as impedance parameters. The feasible region is reconstructed based on the safety constraints corresponding to the degradation section, and the control command is projected onto the feasible region.

6. The control method for a collaborative robot arm according to claim 1, characterized in that, The drift detection method is as follows: When the absolute value of the residual is greater than a preset residual threshold and the duration exceeds a preset duration threshold, it is determined to be a drift candidate state; After determining that it is a candidate state of drift, it is further determined whether the collaborative robot is in a preset safety calibration window. The safety calibration window includes a low contact force state, an unloaded state, and a safe posture state. When the safety calibration window is in effect, it is determined that the system is in a drift state.

7. A control system for a collaborative robot arm, used to implement the control method for a collaborative robot arm according to any one of claims 1-6, characterized in that, include: The sensing unit is used to collect multimodal observation data of the collaborative robot's robotic arm; The intent inference model is used for data preprocessing of multimodal observation data, and the operational intent, task stage, and uncertainty index are obtained through the intent inference model. The trajectory generation module is used to generate a desired curve based on the operational intent and task stage. The impedance control module is used to update the impedance parameters based on the task stage and uncertainty index, and to generate control commands based on the updated impedance parameters and the desired curve. The safety constraint module is used to calculate safety constraints and construct a feasible region based on the operation intention, task stage and uncertainty index, and to perform projection processing on the control command; The execution interface is used to send the projected control commands to the collaborative robot arm and read back the actual response data. The residual calculation module is used to calculate the residual based on the actual response data and the expected curve; The online calibration module is used to perform drift detection based on the residual and to perform online self-calibration when drift is detected, and to write the updated compensation parameters to be calibrated back to the collaborative robot controller.

Images