Hybrid robot adaptive motion control method and system

Abstract

The application discloses a kind of hybrid robot adaptive motion control method and system, it is related to robot motion control technical field, including: obtaining robot system initial state;Establish nominal dynamics prediction model, define safety set and define control barrier function set;Obtain the statistical characteristics of model prediction residual by online learning;Correct dynamics model and generate feedforward compensation, while according to the safety set and barrier function of adaptive adjustment of uncertainty;Rolling optimization problem of integrated closed chain constraint and barrier function hard constraint is constructed;Solving optimization problem obtains desired control quantity, adjusts safety parameter to obtain fallback control quantity when solving fails;Synthesize feedforward compensation and control quantity to generate drive command.The application overcomes model mismatching problem by online learning and residual compensation, realizes safety adaptation using uncertainty, considers performance and safety through unified optimization framework, significantly improves the adaptability, accuracy and robustness of system under dynamic disturbance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot motion control, in particular to a hybrid robot adaptive motion control method and system. BACKGROUND

[0002] With the rapid development of industrial automation and intelligent manufacturing technology, hybrid robots have been widely used in precision assembly fields such as aerospace and automobile manufacturing due to their high stiffness, large load weight ratio and high dynamic performance. In recent years, model predictive control (MPC) based motion control methods have become a research hotspot in this field, which realizes high-performance trajectory tracking by optimizing the system behavior in the future time domain. At the same time, to cope with safety constraints in complex working environments, formal methods such as control barrier function (CBF) are introduced to ensure system safety. In addition, online learning technologies such as Gaussian process regression and deep neural networks are combined to compensate for unmodeled dynamics, further improving the robustness of the control system in the case of model mismatch.

[0003] However, the existing technology still has obvious limitations in practical application. First, the traditional MPC method relies heavily on an accurate dynamic model, while model mismatch problems caused by unmodeled dynamics such as joint friction, transmission hysteresis, link flexibility, and time-varying load are common in hybrid robots, resulting in decreased prediction accuracy and increased tracking error. Although some individual studies have attempted to introduce learning mechanisms to compensate for model errors, most lack quantitative evaluation of the uncertainty of the learning results, making it difficult to combine the confidence of model residuals with system safety control. Second, the existing safety constraint processing method is rigid, usually presetting fixed safety boundaries, which cannot dynamically adjust the safety margin according to the model reliability, resulting in system performance being limited by being too conservative or collision or overrun risks due to insufficient safety margin when the model is highly uncertain. Third, existing methods mostly focus on a single technical level, either improving model accuracy or only processing safety constraints, lacking a control framework that unifies online learning, safety barriers, and the unique closed-loop constraints of hybrid mechanisms, leading to control performance deterioration or even instability under near-singular configurations or dynamic disturbances. In addition, when the optimization problem is infeasible due to constraint conflicts or model mismatch, existing methods often lack effective online adaptive reconstruction mechanisms, making it difficult to ensure basic safety and controllability under extreme working conditions. SUMMARY

[0004] In view of the above problems, the present application is proposed.

[0005] Therefore, the present application solves the technical problem of the existing hybrid robot motion control method, which has the problems of insufficient control accuracy caused by model mismatch, inability of fixed safety constraints to adapt to model uncertainty, lack of learning uncertainty and safety control coordination mechanism, and how to realize high-precision adaptive motion control of hybrid robots under the premise of guaranteeing hard safety constraints.

[0006] To solve the above technical problems, the present application provides the following technical solutions.

[0007] In a first aspect, the present application provides a hybrid robot adaptive motion control method, comprising: obtaining a system initial state of the robot; the system initial state comprises initial state data obtained by sensor collection and preprocessing, and mechanism parameters retrieved from a robot configuration library;

[0008] Based on the initial state data and the mechanism parameters, a nominal dynamics prediction model is established, and a safety set and a control barrier function set are defined;

[0009] Real-time collection and preprocessing of current cycle state data, and receiving a task reference trajectory from a task planning system, inputting the current cycle state data into the nominal dynamics prediction model, and learning the statistical characteristics of the model prediction residual online; the statistical characteristics include residual mean and uncertainty;

[0010] The residual mean in the statistical characteristics is used to correct the nominal dynamics prediction model to obtain a corrected dynamics prediction model and generate a feedforward compensation, and the uncertainty is used to adaptively adjust the safety set and the control barrier function set;

[0011] Based on the corrected dynamics prediction model, the adaptively adjusted safety set and control barrier function set, the current cycle state data, and the task reference trajectory, a rolling optimization problem integrating closed-chain constraints and control barrier function hard constraints is constructed;

[0012] The rolling optimization problem is solved to obtain a desired control amount; when the solution fails, the safety parameters are adjusted based on the uncertainty to obtain a fallback control amount;

[0013] The feedforward compensation and the desired control amount or fallback control amount are synthesized to generate driving instructions and are executed, and the execution results are collected to update the next cycle state data.

[0014] As a preferred scheme of the hybrid robot adaptive motion control method of the present application, the initial state data includes the positions and velocities of each joint, the current and torque observation values of the driver, the spatial pose of the end effector, and the external forces and torques acting on the end and joints.

[0015] As a preferred scheme of the adaptive motion control method of the hybrid robot, the preprocessing comprises time synchronization, filtering and denoising, and calibration of the collected data, unification to the same coordinate system, and closed-chain constraint consistency verification and correction.

[0016] As a preferred scheme of the adaptive motion control method of the hybrid robot, the establishment of the nominal dynamics prediction model comprises constructing a nominal dynamics equation of the robot based on the mechanism parameters, and introducing a closed-chain constraint inherent to the hybrid mechanism, wherein the closed-chain constraint refers to a geometric relationship that must be met by a closed kinematic chain composed of parallel branches, and the closed-chain constraint is taken as an algebraic condition that must be followed to form the nominal dynamics prediction model.

[0017] The definition of the safety set comprises defining a joint position safety set, a joint velocity safety set, a joint torque safety set, an end-effector contact force safety set, and a cooperative safety distance set according to mechanism physical limits and task safety requirements, respectively.

[0018] The definition of the control barrier function set comprises constructing a corresponding control barrier function for each safety set, so that the control barrier function is non-negative within the safety set, and the discrete-time evolution of the barrier function is constrained in the rolling optimization.

[0019] As a preferred scheme of the adaptive motion control method of the hybrid robot, the online learning of the statistical characteristics of the model prediction residual comprises calculating a model prediction value of the system dynamic force based on the nominal dynamics prediction model and the current period state data, wherein the system dynamic force comprises a joint driving torque and an end-effector dynamic force.

[0020] The model prediction value of the system dynamic force is compared with the corresponding actual system dynamic force measured and observed by the sensor to obtain a model prediction residual.

[0021] Based on the model prediction residual and the corresponding system state collected in a plurality of continuous control periods, an online parameter learning algorithm is trained to output the statistical characteristics of the residual under the current system state, including a residual mean value for feeding forward compensation of the model, and an uncertainty for representing the reliability degree of the residual mean value.

[0022] As a preferred scheme of the adaptive motion control method of the hybrid robot according to the application, wherein: the correction of the nominal dynamics prediction model by using the residual mean in the statistical characteristics comprises: taking the residual mean as a real-time estimation of unmodeled dynamics and external disturbances, directly compensating the output of the nominal dynamics prediction model in additive form to form the corrected dynamics prediction model; and converting the residual mean into a feedforward force and torque instruction in the joint space and the task space as the feedforward compensation amount.

[0023] As a preferred scheme of the adaptive motion control method of the hybrid robot according to the application, wherein: the construction of the rolling optimization problem integrating the closed-chain constraint and the control barrier function hard constraint comprises: constructing an optimization objective function about the control sequence in a future limited time domain with the minimization of the task space trajectory tracking error and the system control energy consumption as the main target;

[0024] the corrected dynamics prediction model is taken as an equality constraint for predicting the future state evolution of the system; and the joint position, velocity and torque boundary in the adaptively adjusted safety set is taken as a hard constraint for the system state and control;

[0025] the closed-chain constraint of the hybrid mechanism is taken as an equality constraint in the optimization problem; and the safety condition corresponding to the adaptively adjusted control barrier function set is converted into a linear or nonlinear inequality constraint about the system state and the control input, and is taken as a hard constraint in the optimization problem;

[0026] The prediction time domain and the control time domain of the rolling optimization are set, the weight coefficients of each term in the optimization objective function are configured, and necessary slack variables and their penalty weights are configured for all constraints to ensure the feasibility of the optimization problem.

[0027] As a preferred scheme of the adaptive motion control method of the hybrid robot according to the application, wherein: the solving of the rolling optimization problem to obtain the expected control amount comprises: using a numerical optimization algorithm to solve the rolling optimization problem online, calculating the optimal control sequence in the future control time domain, and taking the first control amount of the sequence as the expected control amount in the current period;

[0028] When the numerical optimization algorithm fails to converge within a preset maximum number of iterations, or the slack variable introduced to meet the constraint exceeds its maximum allowable threshold, it is determined that the solving fails;

[0029] When the solving fails, a simplified optimization problem with only the guarantee of system transient safety as the primary goal is constructed based on the increase of the safety radius of the control barrier function set according to the uncertainty, and the fallback control quantity is obtained by solving the simplified optimization problem; wherein the simplified optimization problem at least contains the closed-chain constraint and the control barrier function constraint with enhanced safety;

[0030] Finally, the expected control quantity obtained by the successful solving process is output, or the fallback control quantity is output when the solving fails, as the effective control instruction of the current control period.

[0031] In a second aspect, an embodiment of the present application provides a hybrid robot adaptive motion control system, comprising:

[0032] An initial state acquisition module: acquiring the system initial state of the robot, wherein the system initial state includes: initial state data obtained by collecting and preprocessing through a sensor, and mechanism parameters retrieved from a robot configuration library;

[0033] A nominal modeling and safety definition module: establishing a nominal dynamics prediction model based on the initial state data and the mechanism parameters, and defining a safety set and a control barrier function set;

[0034] A real-time acquisition and online learning module: real-time acquisition and preprocessing of current period state data, and receiving a task reference trajectory from a task planning system, inputting the current period state data into the nominal dynamics prediction model, and online learning to obtain statistical characteristics of model prediction residuals, wherein the statistical characteristics include residual mean and uncertainty;

[0035] A model correction and feedforward generation module: correcting the nominal dynamics prediction model using the residual mean in the statistical characteristics to obtain a corrected dynamics prediction model and generate a feedforward compensation, and simultaneously adaptively adjusting the safety set and the control barrier function set according to the uncertainty;

[0036] A rolling optimization construction module: constructing a rolling optimization problem integrating closed-chain constraints and control barrier function hard constraints based on the corrected dynamics prediction model, the adaptively adjusted safety set and control barrier function set, the current period state data, and the task reference trajectory;

[0037] An online solving and guardian module: solving the rolling optimization problem to obtain an expected control quantity; when the solving fails, adjusting a safety parameter based on the uncertainty to obtain a fallback control quantity;

[0038] An instruction execution and closed-loop update module: synthesizing the feedforward compensation and the expected control quantity or fallback control quantity to generate a driving instruction and execute it, and collecting the execution result to update the next period state data.

[0039] The beneficial effects of the present application: through online learning and model prediction residual compensation, the model mismatch problem is effectively overcome, and the trajectory tracking accuracy is significantly improved; at the same time, the learning uncertainty is used to adaptively adjust the safety barrier function, the intelligent conservativeness switching of the system when the model reliability changes is realized, and the hard safety constraint under all working conditions is ensured; finally, by unifying learning, safety and mixed-linkage mechanism closed-loop constraints in the rolling optimization framework, the control performance and safety are considered, and the adaptability and robustness of the system under dynamic disturbance and complex tasks are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0041] Figure 1 The overall flowchart of a mixed-linkage robot adaptive motion control method provided for the first embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0043] Embodiment 1, refer to Figure 1 For an embodiment of the present application, a mixed-linkage robot adaptive motion control method is provided, comprising:

[0044] S1: obtaining the initial state of the system of the robot; the initial state of the system includes: initial state data obtained by collecting and preprocessing the sensor, mechanism parameters retrieved from the robot configuration library.

[0045] The initial state data includes the position and speed of each joint, the current and torque observation value of the driver, the spatial pose of the end effector, and the external force and torque received at the end and the joint.

[0046] It should be noted that the multi-source raw data of the present application constitutes a comprehensive state perception system. Among them, the joint position and speed are the basis of kinematics, and the driver current and torque observation value jointly constitute the direct or indirect observation of the dynamic state (the current can be converted into the output torque through the motor torque constant, and the torque value obtained by the observer is verified with each other), the spatial pose of the end effector is obtained through the kinematics forward solution, which is the reference for task space control, and the external force and torque at the end and joint are obtained through the six-dimensional force sensor or joint torque sensor, which is used to perceive the interactive dynamics of the robot and the environment. The synchronous collection of such multi-source heterogeneous data provides a comprehensive and complementary information foundation for subsequent construction of high-fidelity prediction models and safety evaluation, especially the dual perception of torque and force, which can effectively distinguish between internal dynamics and external environmental disturbances.

[0047] The preprocessing includes time synchronization, filtering and denoising, and calibration of the collected data, unification to the same coordinate system, and closed chain constraint consistency verification and correction.

[0048] The preprocessing procedure is a key data preparation work for the characteristics of the hybrid robot system. Time synchronization ensures that the data of different sensing units (such as encoders, force sensors, and vision systems) are correlated under a unified time reference, avoiding state estimation errors caused by transmission delays. Filtering and denoising uses a Kalman filter or a low-pass filter to suppress measurement noise and high-frequency interference, improving data quality. Calibration includes determination of sensor zero position, scale factor, and coordinate transformation parameters to ensure the accuracy of the physical meaning of the data. Unification to the same coordinate system converts all sensor data to the robot base coordinate system or task coordinate system, providing a consistent input reference for subsequent model-based control algorithms. Closed chain constraint consistency verification and correction is a key step specific to hybrid mechanisms, which uses mechanism principles (such as closed-loop vector equations) to reasonably test and project the measured state data, ensuring that the data always satisfy the geometric constraints of the parallel branch chain, and avoiding control command divergence or internal force accumulation caused by data violation of mechanism constraints from the source. This series of preprocessing operations collectively ensures the reliability, consistency, and compatibility of the input data with the physical characteristics of the mechanism.

[0049] It should also be noted that the mechanism parameters retrieved from the robot configuration library should include but not limited to: the mass, center of mass position, inertia tensor of each moving part, the transmission ratio, friction coefficient model parameters of each joint, the DH parameters or POE parameters of each link, and the key geometric dimensions (such as hinge point position, branch length, etc.) that define the topology and kinematics constraints of the hybrid mechanism. These parameters are the physical basis for accurately constructing the nominal dynamics model and defining the safety boundary. They are pre-calibrated and stored in the configuration library, ensuring the accuracy and repeatability of model construction, avoiding repeated identification after each power-on, and improving system deployment efficiency.

[0050] S2: Establish a nominal dynamics prediction model based on the initial state data and the mechanism parameters, and define a safety set and a control barrier function set.

[0051] Based on the mechanism parameters, a nominal dynamics equation of the robot is constructed, and a closed-chain constraint inherent to the hybrid mechanism is introduced, the closed-chain constraint being a geometric relationship that must be satisfied by a closed kinematic chain composed of parallel branches, and the closed-chain constraint being incorporated into the nominal dynamics equation as an algebraic condition that must be followed, thereby forming the nominal dynamics prediction model.

[0052] The definition of the safety set includes defining a joint position safety set, a joint velocity safety set, a joint torque safety set, an end contact force safety set, and a cooperative safety distance set according to the physical limits of the mechanism and the safety requirements of the task.

[0053] The definition of the control barrier function set includes constructing a corresponding control barrier function for each safety set, so that it is non-negative within the safety set, and imposing a constraint on the discrete-time evolution of the barrier function in the rolling optimization.

[0054] It should be noted that the establishment of the nominal dynamics prediction model in this step is a physical modeling process based on first principles. Specifically, based on the mechanism parameters (such as mass, inertia tensor, link geometry, etc.) retrieved from the configuration library, a complete differential equation describing the motion of the robot is constructed through Newton-Euler equations or Lagrange equations, which covers the effects of inertial force, Coriolis force, centrifugal force, and gravity. What is particularly critical is that for hybrid mechanisms, the closed-chain constraint is an inherent kinematic property, which manifests as a set of algebraic equations that limit the degrees of freedom of the system, composed of parallel branches. In constructing the dynamics model, the invention uses the framework of differential-algebraic equations to couple the geometric constraint equations describing the closed-loop relationship of the branches with the aforementioned dynamics differential equations, forming a complete constrained dynamics system. This makes the nominal model not only able to predict the dynamics behavior of the system, but also inherently satisfies the physical construction constraints of the mechanism, laying a solid foundation for generating physically feasible control commands. Compared with the traditional method of possibly ignoring or simplifying the closed-chain constraint, this modeling method avoids the risk of internal force accumulation, energy dissipation, and even structural damage caused by control commands violating the mechanism topology.

[0055] Further, the safety set defined by the invention constitutes a hierarchical and multi-dimensional safety boundary system.

[0056] The joint position safety set is directly determined by the mechanical structure limit of the robot, preventing hardware collision or damage caused by excessive movement.

[0057] Joint velocity safety set is based on the maximum allowable speed setting of motor and transmission components, to avoid speed too high caused by loss of step, jitter or overheating.

[0058] The boundary of joint torque safety set is determined according to the peak torque of the motor, the rated torque of the reducer and the structural strength of the connecting rod, which is the core guarantee to prevent the driver from being overloaded and the mechanical stress from being excessive.

[0059] The end contact force safety set is set for specific work tasks, for example, in assembly or polishing tasks, to avoid damaging the workpiece or tool, the contact force needs to be limited within a safe range.

[0060] The cooperative safety distance set is set to ensure the safety of personnel in human-robot cooperation scenarios, and the minimum distance between the robot end or connecting rod and the operator must be maintained.

[0061] These safety sets collectively constitute the "hard boundary" of the safe operation of the robot, which defines the close combination of the physical limits of the robot and the safety requirements of specific tasks, ensuring the comprehensiveness of safety evaluation.

[0062] It should also be noted that the control barrier function is a formalized method for converting the boundaries of the above safety sets into mathematical constraints that can be directly processed and strictly executed by the control system. For each safety set defined in the invention, a corresponding scalar function (i.e. CBF) is constructed, which is designed to be greater than or equal to zero only when the system state is inside the safety set. The core idea of control is to actively ensure that these CBF function values remain non-negative in the future by online control, thereby equivalently ensuring that the system state never crosses the safety boundary. In the rolling optimization, this idea is implemented as an inequality constraint on the discrete-time evolution of the CBF function. This constraint directly embeds safety requirements into the solution process of the optimization problem, so that any control sequence obtained by solving must ensure the safety of the system in the entire prediction time domain in a mathematically provable way while pursuing the optimal performance index. This method upgrades the traditional post-response safety protection strategy (such as emergency stop) to a proactive safety mechanism that prevents and ensures online, which is a key technical means to realize high safety and autonomous operation of robots.

[0063] S3: Real-time acquisition and preprocessing of current cycle state data, and receiving task reference trajectory from task planning system, inputting current cycle state data into nominal dynamics prediction model, and learning online to obtain statistical characteristics of model prediction residual; the statistical characteristics include residual mean and uncertainty.

[0064] Based on the nominal dynamics prediction model and the current cycle state data, a model predicted value of system dynamic force is calculated, the system dynamic force including joint driving torque and end effector dynamic force;

[0065] The model predicted value of the system dynamic force is compared with the corresponding actual system dynamic force measured and observed directly by sensors to obtain a model prediction residual;

[0066] Based on the model prediction residual and its corresponding system state collected in a plurality of continuous control cycles, an online parameter learning algorithm is trained to output statistical characteristics of the residual under the current system state, including residual mean value for feedforward compensation of the model, and uncertainty for representing the reliability degree of the residual mean value.

[0067] It should be noted that the model predicted value of the system dynamic force is a theoretical expected value obtained by numerical calculation after inputting the current cycle state data (mainly joint position, velocity and other kinematic quantities) into the nominal dynamics prediction model established by S2, which represents the joint driving torque or end effector dynamic force required in theory to maintain the motion under the current motion state. The actual system dynamic force is directly measured by the current sensor installed in the joint driver (converted by torque constant) or the joint torque sensor, and for the end dynamic force, it can be measured by the six-dimensional force sensor installed at the end, or it can be obtained by mapping the joint torque through the dynamics relationship. The difference between the predicted value and the actual measured value is the model prediction residual. This residual accurately represents the part of dynamics that the nominal model fails to describe, including but not limited to joint friction, transmission elasticity, unmodeled inertia coupling effect and external disturbance. This direct comparison based on physical quantities can more quickly and accurately locate and quantify the source of model mismatch compared with indirect learning based on state error.

[0068] Further, the application adopts an online parameter learning algorithm (such as recursive least squares, Bayesian regression or adaptive filter) to train the residual-state data pairs collected within a continuous time window. The learning process has short-term memory characteristics, and through a sliding window or forgetting factor mechanism, it can quickly track the time-varying characteristics of system dynamics, such as friction changes due to temperature rise or dynamic parameter drifts due to load changes. The residual mean output by the learning algorithm is the optimal unbiased estimate of unmodeled dynamics, which is injected into the control system as a feedforward compensation, effectively offsetting the tracking error caused by model mismatch, equivalent to installing an "adaptive calibrator" for the nominal model. The uncertainty output synchronously (usually represented in the form of variance or confidence interval) quantifies the statistical reliability of the current learning result. High uncertainty means that the system is in a state region where the dynamics are changing rapidly or not fully learned. The introduction of this statistical characteristic upgrades the control system from traditional "deterministic compensation" to "probability-aware robust compensation", providing a crucial decision basis for the intelligent adjustment of safety strategies in the subsequent S4 step.

[0069] S4: Modify the nominal dynamics prediction model using the residual mean in the statistical characteristics to obtain a modified dynamics prediction model and generate a feedforward compensation, and adaptively adjust the safety set and the control barrier function set according to the uncertainty.

[0070] The residual mean is used as a real-time estimate of unmodeled dynamics and external disturbances, and is directly compensated in additive form to the output of the nominal dynamics prediction model to form the modified dynamics prediction model; at the same time, the residual mean is converted into joint space and task space feedforward force and torque commands as the feedforward compensation.

[0071] It should be noted that the direct compensation in additive form is an efficient and physically meaningful modification strategy: directly superimposing the learned residual mean on the theoretical force / torque value calculated by the nominal dynamics model. This operation is equivalent to constructing a dynamically updated disturbance observer, whose output is a real-time estimate of unmodeled dynamics (such as nonlinear friction, parameter drift) and external disturbances (such as unknown load, environmental interaction). Through this compensation, the prediction accuracy of the modified dynamics prediction model is significantly improved, making it more closely match the true dynamic characteristics of the system. At the same time, the residual mean can be mapped to the joint space and task space respectively through the Jacobian matrix transpose and other dynamic relationships to generate corresponding feedforward force or torque commands. The feedforward compensation is injected in the control loop in advance, which can actively offset the expected model error, significantly reducing the burden on the feedback controller, thereby reducing the trajectory tracking error at the source and improving the response speed and control bandwidth of the system.

[0072] It should also be noted that the adaptive adjustment of the safety set and the control barrier function set according to the uncertainty is implemented as follows:

[0073] Dynamic scaling of the safety set: The system monitors the uncertainty representing the reliability of the mean of the residual in real time. When the uncertainty increases (indicating that the current learning result is less reliable, and the system is in an insufficient exploration or dynamic mutation state), the system will shrink the boundaries of the safety set of joint speed, joint torque and end contact force according to a pre-set proportional coefficient. For example, the maximum allowed joint speed is reduced from 100% to 70%. This "unreliable and conservative" strategy essentially replaces part of the performance potential with safety margin, effectively preventing over-limit danger caused by aggressive control when the model is not well understood.

[0074] Strengthening of the control barrier function constraint: For the control barrier function set, the system adjusts the parameters in the constraint inequality to change the minimum control effort required to ensure safety. When the uncertainty is high, the system will strengthen the constraint, that is, the control action must be able to more effectively pull the system state away from the safety boundary. In the optimization problem, this represents a more stringent feasibility condition, which directly leads to the generated controller being more cautious and conservative when approaching the safety boundary.

[0075] This intelligent dynamic adjustment of the safety level enables the system to fully exploit performance when the model is well understood, and automatically ensure safety when the model is not well understood, fundamentally solving the dilemma of fixed safety boundaries in dealing with model uncertainty, and achieving the optimal trade-off between performance and safety under different working conditions.

[0076] S5: Based on the corrected dynamic prediction model, the adaptively adjusted safety set and control barrier function set, the current period state data and the task reference trajectory, a rolling optimization problem is constructed, which integrates closed-chain constraints and control barrier function hard constraints.

[0077] An optimization objective function for the control sequence in the future finite time domain is constructed, with the main objectives being to minimize the task space trajectory tracking error and the system control energy consumption;

[0078] The corrected dynamic prediction model is used as an equality constraint for predicting the future state evolution of the system; the joint position, speed and torque boundaries in the adaptively adjusted safety set are used as hard constraints for the system state and control;

[0079] The closed-chain constraints of the hybrid mechanism are included in the optimization problem as equality constraints; the safety conditions corresponding to the adaptively adjusted control barrier function set are converted into linear or nonlinear inequality constraints related to the system state and control input, and are included in the optimization problem as hard constraints;

[0080] The prediction horizon and control horizon of the rolling optimization are set, the weight coefficients of each term in the optimization objective function are configured, and the necessary slack variables and their penalty weights are configured for all constraints to ensure the feasibility of the optimization problem.

[0081] It should be noted that the design of the optimization objective function in this step reflects a trade-off between performance and efficiency. The minimization of the task space trajectory tracking error specifically refers to minimizing the norm of the deviation of the end effector's expected position and attitude from the task reference trajectory within the prediction horizon, which directly determines the execution accuracy of the task. Minimizing system control energy consumption is usually embodied in minimizing the quadratic form of joint torque or current, which helps to reduce heat generation, improve energy efficiency, and prolong the life of the driver. By configuring the weight coefficients of these two terms in the objective function, the degree of emphasis on "accuracy" and "energy saving" of the system can be flexibly adjusted according to specific task requirements, for example, giving higher weight to tracking error in the finishing stage, and focusing more on energy consumption in the empty moving stage.

[0082] Furthermore, regarding the multi-constraint integration and rolling optimization mechanism, various types of physical and safety constraints are unified in an optimization framework for collaborative processing:

[0083] The dynamics equation constraint ensures that the predicted system behavior optimized conforms to the revised dynamics prediction model, ensuring the physical accuracy of the prediction.

[0084] The hard constraint of the safety set directly uses the physical boundaries adjusted by S4 (such as joint position, velocity, torque) as the upper and lower limits that the optimization variables must comply with, providing the most basic operation protection.

[0085] The closed-chain constraint as an equality constraint again emphasizes the special geometric structure of the hybrid mechanism, ensuring that all predicted states and control commands are always kinematically feasible, fundamentally avoiding mechanism jamming or internal force accumulation.

[0086] The control barrier function (CBF) as an inequality constraint is the core of active safety. It converts the requirement of "keeping the system state safe" into a mathematical condition about future states and controls that needs to be continuously satisfied. By taking it as a hard constraint, the optimization solver is forced to generate only those control sequences that can guarantee the non-negativity (i.e., safety) of the CBF within the prediction horizon.

[0087] It should also be noted that, to ensure that this complex optimization problem is always solvable in real-time control, the present invention introduces slack variables and configures corresponding penalty weights. When strict hard constraints cannot be satisfied simultaneously (e.g. instantaneous torque limit may need to be exceeded momentarily to avoid an unexpected obstacle), slack variables allow constraints to be violated to the minimum extent, but will be severely penalized in the objective function, thereby guiding the solver to find the most reasonable solution with the least violation. The length of the prediction horizon and the control horizon is a key trade-off: longer horizons can improve control foresight and stability, but will significantly increase the computational burden. By setting these parameters reasonably, and using an efficient numerical solver, the present invention realizes real-time rolling solution of complex optimization problems under limited computing resources, i.e. solving once per control period, and only implementing the first control amount of the optimal sequence, thereby forming a closed-loop feedback.

[0088] S6: solving the rolling optimization problem to obtain a desired control amount; when the solving fails, adjusting a safety parameter based on the uncertainty to obtain a fallback control amount.

[0089] Solving the rolling optimization problem online using a numerical optimization algorithm to calculate the optimal control sequence within the future control horizon, and taking the first control amount of the sequence as the desired control amount of the current period;

[0090] When the numerical optimization algorithm fails to converge within a preset maximum number of iterations, or the slack variable introduced to satisfy the constraints exceeds its maximum allowed threshold, it is determined that the solving has failed;

[0091] When the solving fails, increasing the safety radius of the control barrier function set based on the uncertainty to construct a simplified optimization problem that prioritizes ensuring the instantaneous safety of the system, and solving the simplified optimization problem to obtain the fallback control amount; wherein the simplified optimization problem at least includes the closed-chain constraint and the control barrier function constraint that has been strengthened in safety;

[0092] Finally outputting the desired control amount obtained by a successful solving process, or the fallback control amount when the solving fails, as the effective control instruction of the current control period.

[0093] It should be noted that, regarding online solving and failure determination mechanism:

[0094] The numerical optimization algorithm is used for online solution in this step, usually referring to the use of quadratic programming or nonlinear programming solver. In each control cycle, the system instantiates the optimization problem constructed in S5, and uses the hot start technique (i.e. using the optimal solution of the last cycle as the initial guess of this time) to accelerate the convergence process. The optimal control sequence output by the solver reflects the optimal decision of the system in the short term in the future, taking into account tracking performance, energy consumption and multiple safety constraints, and the first control quantity is implemented, which meets the rolling optimization principle of model predictive control. Double standards are set for the determination of failed solution: the maximum number of iterations without convergence indicates that the problem may be too complex or there may be numerical ill-conditioning; the relaxation variable exceeds the threshold, indicating that the constraint conflict is serious, and even if a solution is obtained, it needs to pay an unacceptable performance or safety price. This strict determination mechanism ensures that only high-quality optimization solutions are adopted.

[0095] Further, regarding the adaptive design of the fallback control mechanism:

[0096] This step is based on the uncertainty-based safety strategy. When the main optimization problem fails to solve, the system does not simply stop or follow the old instructions, but starts a degraded control mode. At this time, the system first increases the safety radius of the control barrier function set based on the uncertainty, which is manifested as scaling the safety margin parameter in the CBF constraint by the uncertainty to form a safety-enhanced control barrier function constraint. The simplified optimization problem constructed later, its objective function usually only retains the most basic control smoothness term, or even directly takes zero as the target, and the constraint set is simplified to the closed constraint that must be absolutely guaranteed and the strengthened CBF constraint. This design ensures that the fallback control quantity can guarantee the physical feasibility and instantaneous safety of the mechanism movement with the highest priority, especially when the model cognitive uncertainty is high, by sacrificing part of the performance to obtain higher safety guarantee. This hierarchical control architecture greatly enhances the survival ability of the system in extreme working conditions, avoids the system crash caused by the failure of a single optimization problem, and provides the last solid guarantee for the reliable autonomous operation of the robot in an unknown dynamic environment.

[0097] S7: Synthesize the feedforward compensation quantity and the desired control quantity or fallback control quantity to generate a driving instruction and execute it, and collect the execution result to update the state data for the next cycle.

[0098] It should be noted that regarding the synthesis and execution of the driving instruction:

[0099] This step completes the final conversion from control calculation to physical execution. The synthesis process is a linear superposition of the feedforward compensation generated by S4 step and the desired control (or fallback control) output by S6 step. The feedforward compensation is mainly responsible for offsetting the predictable model error and steady-state disturbance in the system, while the feedback control (desired or fallback control) is used to handle unmodeled dynamic error and random disturbance. This "feedforward-feedback" composite control structure takes full advantage of both: feedforward control improves the response speed and trajectory tracking accuracy of the system, and feedback control ensures the stability and robustness of the system. After the synthesized driving instructions are converted to analog signals and amplified in power, they drive the joint actuators (usually motors) to produce corresponding movements and forces. The low-delay implementation of this process is crucial to ensure the real-time performance of the control system.

[0100] It should also be noted that, regarding the closed-loop update mechanism:

[0101] After the instructions are executed, the system immediately collects the execution results through sensors, including new joint position, velocity, torque, etc. After these new data are preprocessed through a similar process as S1 step, they are updated as the state data for the next period. This operation completes the last link of the control loop and realizes the closed-loop feedback mechanism of rolling optimization. The updated state data will be used as the input for all steps in the next control period, such as online learning and optimization problem construction, so that the system can continuously adjust and optimize itself based on the latest actual state. This periodic "execution-sensing-update" cycle makes the method described in the present application a truly adaptive closed-loop control system, which can effectively deal with system dynamics changes, external disturbances, and environmental uncertainties, and continuously maintain high-precision motion control performance and reliable safety guarantees.

[0102] Embodiment 2, which is different from the previous embodiment, is a second embodiment of the present application:

[0103] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the prior art that make essential contributions or parts of the current technical solutions can be embodied in the form of software products, and the current computer software product is stored in a storage medium, including a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes.

[0104] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions for implementing logic functions, and can be embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable medium" can be any means that can contain, store, communicate, propagate or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be a machine-readable storage device (e.g., magnetic, optical or other storage device) merely an article of manufacture that launches the program into an instruction execution system, apparatus, or device.

[0105] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and stored in a computer memory.

[0106] Embodiment 3, as an embodiment of the present application, provides a hybrid robot adaptive motion control system, comprising an initial state acquisition module, a nominal modeling and safety definition module, a real-time acquisition and online learning module, a model correction and feedforward generation module, a rolling optimization construction module, an online solving and daemon module, and an instruction execution and closed-loop update module.

[0107] The initial state acquisition module acquires the system initial state of the robot, wherein the system initial state comprises: initial state data obtained by sensor acquisition and preprocessing, and mechanism parameters retrieved from a robot configuration library.

[0108] The nominal modeling and safety definition module establishes a nominal dynamics prediction model based on the initial state data and the mechanism parameters, and defines a safety set and a control barrier function set.

[0109] The real-time acquisition and online learning module acquires and pre-processes the current period state data in real time, receives a task reference trajectory from a task planning system, inputs the current period state data into the nominal dynamics prediction model, and learns online to obtain statistical characteristics of model prediction residuals, wherein the statistical characteristics comprise residual mean and uncertainty.

[0110] Model correction and feedforward generation module: the nominal dynamics prediction model is corrected by using the residual mean in the statistical characteristics, a corrected dynamics prediction model is obtained, a feedforward compensation amount is generated, and the safety set and the control barrier function set are adaptively adjusted according to the uncertainty;

[0111] Rolling optimization construction module: based on the corrected dynamics prediction model, the adaptively adjusted safety set and control barrier function set, the current period state data and the task reference trajectory, a rolling optimization problem integrating closed-chain constraints and control barrier function hard constraints is constructed;

[0112] Online solution and guardian module: the rolling optimization problem is solved to obtain the expected control amount; when the solution fails, the safety parameter is adjusted based on the uncertainty to obtain the fallback control amount;

[0113] Instruction execution and closed-loop update module: the feedforward compensation amount and the expected control amount or fallback control amount are synthesized to generate driving instructions and execute them, and the execution results are collected to update the next period state data.

[0114] Embodiment 4, which is an embodiment of the present application, provides a hybrid robot adaptive motion control method. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation / contrast experiments are used for scientific demonstration.

[0115] The test object is a hybrid robot platform, which is composed of a six-degree-of-freedom serial arm and a three-degree-of-freedom parallel positioning table. The end of the robot is equipped with a clamping assembly tool and a six-dimensional force sensor. The test scenario is "dynamic assembly-sliding fitting": the workpiece moves at a constant speed of about 0.42 m / s on the conveying line, and the robot needs to complete the tasks of grabbing, positioning and sliding insertion within a limited time. The system first performs S1: collects initial state data and loads mechanism parameters from the configuration library. The data channels cover joint position and velocity, drive current and torque observation, end pose, external force and torque at the end and joint. After time synchronization, filtering, calibration, coordinate unification, closed loop consistency verification and correction, the standardized initial data set is formed. Then perform S2: based on the data and mechanism parameters, construct the nominal dynamics prediction model, and define the safety set of joint position / velocity / torque, end contact force and human-robot collaboration distance, and the corresponding control barrier function. Enter S3: under the premise that the task planning system gives the "sliding fitting" trajectory, real-time acquisition of current cycle state data is performed, and the data is input into the nominal model to form a short-time prediction of joint driving torque and end generalized force; the prediction value and the measured value are compared to generate residual samples, and the residual mean and uncertainty are given through the online learning module. Perform S4: the residual mean is used as the equivalent estimation of unmodeled dynamics and external disturbance, the nominal model is corrected, and joint / task space feedforward compensation is generated; the uncertainty is used to adaptively tighten the safety set boundary and barrier parameter. Perform S5: according to the corrected model and adaptive safety framework, construct a rolling optimization problem integrating closed loop equations and barrier hard constraints, and the optimization objective considers both task space tracking error and energy consumption. Perform S6: use an online solver with warm start to output the first item of the control sequence as the expected control quantity within a time budget of 1-2 ms; if there is a timeout or the relaxation amount exceeds the threshold, trigger the rollback sub-problem based on the uncertainty, and only ensure safety and geometric feasibility. Finally, perform S7: the feedforward compensation and the expected (or rollback) control quantity are combined and sent to the drive side for execution, and the execution data is collected for the next cycle. To verify the creative effect, five groups of controls are set: no online learning, no CBF hard constraint, fixed safety boundary, no feedforward compensation, and only impedance control. All schemes are tested under the same load, same task, and same speed, and the indicators are counted in a 10-minute window. Table 1 can be referred to.

[0116] Table 1: Data reference table

[0117]

[0118] From the table information, it can be seen that after adopting the method, the comprehensive advantages of precision, safety and efficiency are obtained in the dynamic assembly task. First, the trajectory RMS error is significantly reduced from 1.78 mm without online learning and 1.21 mm without feedforward compensation to 0.62 mm, with a reduction of 65% and 49% respectively, which shows that the residual mean for model correction and feedforward injection can effectively offset the systematic deviation caused by unmodeled dynamics and external disturbances, reduce the dependence on high gain feedback, thereby improving the tracking accuracy and avoiding the chattering caused by high gain. Second, in terms of safety, there are 7 contact overruns and 9 barrier violations without CBF hard constraints, and the number of instability in the near-singularity region is reduced from 3, 4 and 2 of A, B and D to 0, which shows that the method of integrating closed-chain constraints and CBF hard constraints into rolling optimization avoids numerical ill-conditioning and boundary crossing, and automatically tightens the safety set when the uncertainty increases, so that the system is far away from the risk domain with a more conservative control strategy. Third, the feasible rate of optimization solution is 99.4%, which is higher than that of fixed safety boundary and no feedforward compensation (98.1%, 97.0%), and is obviously better than that of no CBF hard constraint (93.1%). This shows that the linkage of uncertainty and barrier parameters and the guarantee of minimum feasibility through backtracking subproblem when infeasible signs appear is the key mechanism to improve the stability of online optimization.

[0119] In terms of energy efficiency, the single-kiloprocess cycle energy consumption is reduced to 41.7 Wh, which is 13% and 16% lower than 47.9 Wh (without online learning) and 49.8 Wh (only impedance control), respectively. This result is attributed to two points: one is that the feedforward compensation makes the actuator not have to counteract the modeling error with high feedback torque for a long time, and the other is that the CBF constraint suppresses unnecessary impact and repeated correction, reducing the peak power. Correspondingly, the peak joint torque and the end-effector force peak are reduced from 79.1 N·m and 154.7 N (without online learning) to 62.4 N·m and 118.3 N, respectively, showing substantial relief of mechanical load, which helps to prolong the life of transmission and joint components. In terms of quality and yield, the workpiece surface defect rate is reduced from 1.4% to 3.2% in the control group to 0.3%, which is consistent with the zeroing of contact overrun events, reflecting the direct gain of process quality by active safety constraint and uncertainty-driven shrinking boundary. In terms of real-time performance, the average solving time delay is 1.8 ms, which is slightly higher than some controls, but still within the real-time control budget; at the same time, the rollback trigger rate is only 1.1%, indicating that the optimization problem maintains a benign solvable state most of the time, and can smoothly retreat when the disturbance is severe, without losing control. Comprehensive comparison can show that the innovations of this method - mean correction of online residual learning, safety self-adaptation based on uncertainty, closed-chain consistent rolling optimization and rollback guardian - form a quantifiable performance-safety-efficiency overall advantage in the same scene. This group of data proves the necessity and significance of the creative technical features, and also reflects the substantial improvement compared with traditional control strategies.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application rather than limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and all modifications and equivalent replacements should be covered in the scope of the claims of the present application.

Claims

1. A hybrid robot adaptive motion control method, characterized in that, The method comprises the following steps: acquiring the initial state of the robot system; the initial state of the system includes: initial state data collected and pre-processed by sensors, and mechanism parameters retrieved from a robot configuration library; based on the initial state data and the mechanism parameters, a nominal dynamics prediction model is established, and a safety set and a control barrier function set are defined; real-time acquisition and pre-processing of the current cycle state data, and receiving the task reference trajectory from the task planning system, inputting the current cycle state data into the nominal dynamics prediction model, and learning the statistical characteristics of the model prediction residual online; the statistical characteristics include residual mean and uncertainty; using the residual mean in the statistical characteristics to correct the nominal dynamics prediction model to obtain a corrected dynamics prediction model and generate a feedforward compensation, and simultaneously adjusting the safety set and the control barrier function set according to the uncertainty; based on the corrected dynamics prediction model, the adaptively adjusted safety set and control barrier function set, the current cycle state data, and the task reference trajectory, a rolling optimization problem integrating closed-chain constraints and control barrier function hard constraints is constructed; wherein the construction of the rolling optimization problem integrating closed-chain constraints and control barrier function hard constraints includes: constructing an optimization objective function about the control sequence in the future finite time domain, with the main goal of minimizing the task space trajectory tracking error and the system control energy consumption; taking the corrected dynamics prediction model as the equality constraint of predicting the future state evolution of the system; taking the joint position, velocity, and torque boundaries in the adaptively adjusted safety set as the hard constraints of the system state and control; taking the closed-chain constraints of the hybrid mechanism as equality constraints into the optimization problem; converting the safety conditions corresponding to the adaptively adjusted control barrier function set into linear or nonlinear inequality constraints about the system state and control input, and taking them as hard constraints into the optimization problem; setting the prediction time domain and control time domain of the rolling optimization, configuring the weight coefficients of each term in the optimization objective function, and configuring the slack variables and their penalty weights to ensure the feasibility of the optimization problem; solving the rolling optimization problem to obtain the expected control amount; when the solution fails, adjusting the safety parameters based on the uncertainty to obtain the fallback control amount; synthesizing the feedforward compensation and the expected control amount or fallback control amount to generate driving instructions and execute them, and collecting the execution results to update the next cycle state data. 2.The hybrid robot adaptive motion control method of claim 1, wherein, The initial state data includes the position and velocity of each joint, the current and torque observation value of the driver, the spatial pose of the end effector, and the external force and torque received at the end and the joint. 3.The hybrid robot adaptive motion control method of claim 2, wherein, The preprocessing includes time synchronization, filtering and denoising, and calibration of the collected data, unification to the same coordinate system, and consistency verification and correction of the closed-chain constraints. 4.The hybrid robot adaptive motion control method of claim 3, wherein, The establishing of the nominal dynamics prediction model comprises: constructing a nominal dynamics equation of the robot based on the mechanism parameters; and introducing inherent closed-chain constraints of the hybrid mechanism, wherein the closed-chain constraints refer to geometric relations that must be satisfied by a closed kinematic chain composed of parallel branches, and the closed-chain constraints are taken as algebraic conditions that must be followed to form the nominal dynamics prediction model; The defining of the safety set comprises: defining a joint position safety set, a joint velocity safety set, a joint torque safety set, an end effector contact force safety set, and a cooperative safety distance set according to mechanism physical limits and task safety requirements respectively; The defining of the control barrier function set comprises: constructing a corresponding control barrier function for each safety set, so that the control barrier function is non-negative within the safety set, and a discrete-time evolution of the barrier function is constrained in the rolling optimization. 5.The hybrid robot adaptive motion control method of claim 4, wherein, The online learning of the statistical characteristics of the model prediction residual comprises: calculating a model prediction value of a system dynamic force based on the nominal dynamics prediction model and the current period state data, wherein the system dynamic force comprises a joint driving torque and an end effector dynamic force; Comparing the model prediction value of the system dynamic force with a corresponding actual system dynamic force measured and observed by a sensor to obtain a model prediction residual; Based on the model prediction residuals collected in a plurality of continuous control periods and corresponding system states, an online parameter learning algorithm is used for training, so as to output statistical characteristics of the residual under the current system state, including a residual mean value for feeding forward compensation of the model, and an uncertainty for representing a reliability degree of the residual mean value. 6.The hybrid robot adaptive motion control method of claim 5, wherein, The modifying of the nominal dynamics prediction model by using the residual mean value in the statistical characteristics comprises: taking the residual mean value as a real-time estimation of unmodeled dynamics and external disturbances, directly compensating an output of the nominal dynamics prediction model in an additive form to form a modified dynamics prediction model; and converting the residual mean value into a joint space and a task space to obtain a feedforward force and torque instruction as the feedforward compensation amount.

7. The hybrid robot adaptive motion control method of claim 1, wherein, The solving of the rolling optimization problem to obtain the expected control amount comprises: using a numerical optimization algorithm to solve the rolling optimization problem online, calculating an optimal control sequence in a future control time domain, and taking a first control amount of the sequence as the expected control amount of the current period; When the numerical optimization algorithm fails to converge within a preset maximum number of iterations, or a relaxation variable introduced to satisfy a constraint exceeds a maximum allowed threshold, it is determined that the solving fails; When the solving fails, the safety radius of the control barrier function set is increased based on the uncertainty, a simplified optimization problem that only takes guaranteeing instantaneous safety of the system as a primary goal is constructed, and the fallback control amount is obtained by solving the simplified optimization problem; wherein the simplified optimization problem at least contains the closed-chain constraint and a safety-enhanced control barrier function constraint; Finally, the expected control amount obtained in a successful solving process is output, or the fallback control amount is output as an effective control instruction of the current control period when the solving fails.

8. A hybrid robot adaptive motion control system for implementing the hybrid robot adaptive motion control method according to any one of claims 1 to 7, characterized in that, The method comprises the following steps: An initial state acquisition module: acquire the system initial state of the robot, including: initial state data obtained by collecting and preprocessing through sensors, and mechanism parameters retrieved from a robot configuration library; Nominal modeling and safety definition module: based on the initial state data and the mechanism parameters, establish a nominal dynamics prediction model, and define a safety set and a control barrier function set; Real-time acquisition and online learning module: real-time acquisition and preprocessing of current period state data, and receiving of a task reference trajectory from a task planning system, input of the current period state data into the nominal dynamics prediction model, online learning of statistical characteristics of model prediction residuals, including residual mean and uncertainty; Model correction and feedforward generation module: correction of the nominal dynamics prediction model using the residual mean in the statistical characteristics, obtaining of a corrected dynamics prediction model and generation of a feedforward compensation, and adaptive adjustment of the safety set and the control barrier function set according to the uncertainty; Rolling optimization construction module: construction of a rolling optimization problem integrating closed-chain constraints and control barrier function hard constraints based on the corrected dynamics prediction model, the adaptively adjusted safety set and control barrier function set, the current period state data, and the task reference trajectory; Online solving and guardian module: solving of the rolling optimization problem to obtain a desired control amount; when the solving fails, adjusting safety parameters based on the uncertainty to obtain a fallback control amount; Instruction execution and closed-loop update module: synthesis of the feedforward compensation and the desired control amount or fallback control amount, generation and execution of driving instructions, and acquisition of execution results to update next period state data.

Images