A mechanical arm stage action amplitude correction control method

Abstract

This invention provides a staged motion amplitude correction control method for a robotic arm, relating to the field of intelligent robotic arm control technology. The method involves acquiring input information for the current control cycle of the robotic arm, inputting this information into a basic motion generation strategy, and outputting a basic motion. The current task execution stage is determined based on the real-time task execution status of the robotic arm. Matching motion amplitude correction parameters are generated based on the current task execution stage. These parameters are then used to correct the position and attitude components of the basic motion to obtain the final executed motion. The final executed motion is then converted into low-level control commands for the robotic arm and sent to the robotic arm controller to complete the task execution. This invention adds a stage-aware correction module between the model and the robotic arm controller without altering the main structure of the basic VLA model. This module independently corrects the translation and rotation amplitudes of the basic motion in stages, balancing rapid long-distance movement with high-precision short-distance fine-tuning.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for robotic arms, and in particular to a method for correcting and controlling the amplitude of staged movements of a robotic arm. Background Technology

[0002] Robotic manipulation tasks typically require generating continuous motion outputs based on visual observations, task objectives, and the state of the robotic arm to drive the arm to perform operations such as approaching, grasping, placing, lifting, or posture adjustment. Existing technologies, such as kinematic modeling, trajectory planning, impedance control, or force-position hybrid control, offer good stability in structured scenarios. However, these methods usually require manual modeling, path planning, and parameter tuning tailored to specific robotic arm configurations, end effectors, target object positions, and task flows. This results in high adaptation costs when the operating environment, task objectives, or object states change.

[0003] To reduce manual modeling and rule design, existing technologies often employ imitation learning or reinforcement learning methods to train robot strategies. Imitation learning learns the mapping from observation to action through taught trajectories, while reinforcement learning improves task performance through environmental interaction and reward signals. However, imitation learning typically relies on the quality and coverage of the taught data, while reinforcement learning depends on the design of interaction samples and reward functions. When the target task data is limited in scale or the coverage of different action amplitude stages is uneven, it remains difficult to reliably ensure that the robotic arm has appropriate action scales at different execution stages.

[0004] With the development of vision-language-action (VLA) models, VLA models can combine visual observations, verbal commands, and robotic arm states to generate action outputs, thereby improving the adaptability of robot strategies to different task semantics and operational scenarios. Lightweight VLA models such as SmolVLA further reduce training, fine-tuning, and deployment costs, making them suitable for rapid adaptation in laboratory robotic arms, low-cost robot platforms, and specific robotic arm operation tasks. However, lightweight VLA models are limited by model structure and parameter size. When there are differences between the target robotic arm configuration, camera viewpoint, or task distribution and the pre-training data, their action outputs become more dependent on the distribution of the target fine-tuning data.

[0005] In specific downstream task adaptation, supervised fine-tuning is a commonly used method. It typically utilizes teaching data collected by the target robot to further train the model or action head, enabling the model to fit reference action labels in the target task data. This method is simple to implement and can quickly adapt to new robotic arm platforms, camera perspectives, and task distributions. However, its optimization objective is primarily to fit the action labels in the dataset. When the sample distribution of different execution stages in the target task data is unbalanced, or when certain action amplitude ranges account for a high proportion, the fine-tuned strategy tends to shrink towards the high-proportion action amplitude ranges, thereby weakening the differences in action scale between different execution stages.

[0006] In addition, efficient parameter fine-tuning, reinforcement learning-based adaptation methods, and techniques such as action normalization, sample resampling, loss function weight adjustment, or global action scaling are also commonly used to improve the training cost, task success rate, or action output stability of robot strategies. Among these, efficient parameter fine-tuning mainly reduces training costs, reinforcement learning methods can utilize online interactive data or offline trajectory data to improve task rewards, and global action processing methods can improve the overall action scale stability; however, these methods typically do not establish explicit staged analysis and correction mechanisms for the action amplitude requirements of different execution stages in multi-stage robotic arm operation tasks.

[0007] The existing methods such as global action scaling, fixed loss weights, and sample resampling are all global uniform corrections, which can only apply a single scaling factor to the entire task process and cannot distinguish the differentiated amplitude requirements of different stages such as approach, fine-tuning, and contact adjustment. At the same time, the existing residual compensation schemes lack stage switching logic and hard amplitude constraints. The total compensation amount will significantly deviate from the basic motion direction output by the VLA model, which will destroy the main trend of the action in the visual language planning. Furthermore, the lightweight VLA model suffers from amplitude collapse after fine-tuning. The lack of an explicit segmented amplitude correction architecture for multi-stage tasks in the existing technology is the core technical pain point that this invention needs to solve. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a staged motion amplitude correction control method for a robotic arm. Without altering the main structure of the basic VLA model, a stage perception and correction module is added between the model and the robotic arm controller to independently correct the translation and rotation amplitudes of the basic motion in stages, taking into account both long-distance rapid propulsion and short-distance high-precision fine adjustment.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A method for staged motion amplitude correction control of a robotic arm consists of three main processes: offline data processing, staged training of the correction module, and online closed-loop control. Specifically, it includes the following steps:

[0011] Step 1: Offline data processing, performing offline task data analysis and motion amplitude stage division for the target robotic arm;

[0012] First, multiple operation trajectories of the target robotic arm task are acquired. Each operation trajectory includes visual observation, robotic arm status, task language instructions, and reference motion labels. Then, the corresponding translational motion amplitude and rotational motion amplitude are calculated based on the position control components and attitude control components in the reference motion labels. Furthermore, the changes in motion amplitude in different operation trajectories are statistically analyzed, smoothed, and divided into stages to obtain the motion amplitude stages, stage boundaries, and corresponding motion amplitude statistics for each stage during the execution of the robotic arm task.

[0013] The offline time dimension stage boundaries are converted into online usable geometric switching rules. The horizontal plane distance between the end effector and the target position is selected as the online stage switching variable. The distance samples corresponding to each stage boundary are counted to generate stage switching thresholds for online real-time determination of task stages.

[0014] Step 2: During the training phase of the calibration module, construct the phase-aware motion amplitude calibration control module;

[0015] After determining the motion amplitude stage, the correction control module is set between the motion output end of the basic motion generation strategy and the robotic arm controller. Based on the robotic arm state, basic motion, basic motion amplitude, task stage information and target pose error, the correction control module outputs position scale gating coefficients and attitude scale gating coefficients for correcting the basic motion, and outputs restricted residual correction amount in the preset stage, thereby performing staged amplitude correction of the position increment and attitude increment of the basic motion.

[0016] The calibration module inputs include the robotic arm state, basic motion, basic motion position / attitude amplitude norm, target position error, target attitude error, distance between the end effector and the target plane, gating coefficient of the previous cycle, and weight of the current stage. After normalization, all input features are fed into a shared feature extraction network to obtain shared state features. An independent stage branch network is set for each motion amplitude stage to generate gating raw variables and residual raw variables for each stage.

[0017] The original gating variables are mapped to the corresponding pre-defined gating intervals via a Sigmoid function, and then weighted and fused using soft-stage weights in the 0-1 interval to obtain the global position-scale gating coefficients. Attitude scale gating coefficient Adjacent stages rely on soft weights for smooth transitions, avoiding gating abrupt changes.

[0018] The original residual variables are subject to dual constraints by combining the maximum residual amplitude of each stage with the tanh function. The residual terms are only enabled in the fine-tuning, attitude alignment and contact adjustment stages, while the residuals are set to zero in the long-distance approach stage.

[0019] Execution of action composition logic:

[0020] Non-residual activation phase: , ;

[0021] Residual activation phase: , ;

[0022] in, and These represent the actual end-effector position and end-effector attitude increment, respectively. and These represent the end-effector position and end-effector pose increments in the basic motion, respectively. and This is a restricted residual correction term;

[0023] The gripper control components directly use the basic strategy output;

[0024] Step 3: Freeze the base model and train the correction control module using reinforcement learning;

[0025] Keeping the main parameters of the basic action generation strategy unchanged, only the correction control module is trained; the training process uses the PPO proximal policy optimization reinforcement learning algorithm to optimize and model the correction module as a Markov decision process;

[0026] Two types of stage-differentiated reward functions are set up: coarse-grained reward is used in the non-fine-tuning coarse stage to reduce the target horizontal distance, promote task stage switching, and suppress invalid rotation; fine reward is used in the fine-tuning high-precision stage to reduce 3D pose error, reward task success, and at the same time increase rotation penalty, motion smoothing penalty, and residual amplitude penalty.

[0027] The total optimization loss includes PPO policy loss, value loss, policy entropy regularization term, stage identification auxiliary loss, action amplitude consistency loss, gating prior loss, and residual smoothing constraint loss; the action amplitude consistency loss penalizes actions whose amplitudes exceed the amplitude range of the corresponding stage dataset after correction; the gating prior loss constrains the gating coefficients of each stage to fit the offline statistical stage gating prior.

[0028] The training adopts a course-based step-by-step training process: the first step is to close the residual correction branch and train only the scale gating branch; the second step is to open the high-precision stage residual branch and train the residual network separately; the third step is to reduce the learning rate and jointly fine-tune all branches of gating and residual until the model converges.

[0029] Step 4: Use the trained calibration control module to perform online closed-loop real-time control of the robotic arm;

[0030] After training is completed, in each online control cycle, the current visual observation, robotic arm status, and task language instructions are first acquired, and basic actions are generated by the basic action generation strategy. Then, based on the current stage switching threshold of the distance between the robotic arm end effector and the target plane, the soft weights of each stage are calculated, and the amplitude of the basic actions is corrected by the trained correction control module to obtain the final execution action. Finally, the final execution action is converted into end effector pose control instructions, joint control instructions, servo control target or gripper control instructions, and sent to the robotic arm controller for execution. The above process is repeated until the task is completed.

[0031] The basic motion generation strategy is the lightweight VLA model SmolVLA, or other visual language motion models or robot strategy models that can receive visual observations, language commands, and robotic arm states and output continuous robot actions.

[0032] The beneficial effects of adopting the above technical solution are as follows: The robotic arm staged motion amplitude correction control method provided by this invention transforms the amplitude adaptation problem of the output motion of the lightweight visual language motion model into the motion amplitude correction problem in the closed-loop control link of the robotic arm. By setting a correction control module between the basic motion generation strategy and the robotic arm controller, the final executed motion can better adapt to the control requirements of the current task stage of the robotic arm, thereby improving the stability of the robotic arm closed-loop control process; This invention determines the motion amplitude stage based on the position motion amplitude and posture motion amplitude distribution in the target robotic arm task data, and performs online stage judgment in combination with the current state of the robotic arm, task progress, or target pose error, so that the motion amplitude correction process can match different execution stages such as robotic arm approach, fine-tuning, posture alignment, grasping, placement, or contact adjustment, improving the motion scale adaptation capability in multi-stage operation tasks; This invention, while keeping the main parameters of the basic motion generation strategy unchanged, only trains the lightweight correction control module, which can retain the basic The existing visual language understanding and basic motion direction prediction capabilities of the strategy avoid retraining or significant fine-tuning of the entire basic model, thereby reducing the costs of model adaptation, training, and deployment. This invention corrects the position and posture increments of the basic motion through position-scale gating coefficients and attitude-scale gating coefficients, respectively, and introduces a constrained residual correction amount in the preset stage. At the same time, combined with motion amplitude consistency constraints, gating constraints, and residual amplitude constraints, the corrected execution motion is kept within a reasonable motion distribution range, reducing the risk of excessive motion amplification, abrupt motion changes, or deviation from the basic motion direction, and improving the continuity, safety, and reliability of the robotic arm control process. Compared with motion normalization, fixed loss weight adjustment, or global motion scaling, this invention can adopt different motion amplitude compensation intensities and residual correction methods for different robotic arm task stages. This not only alleviates the motion amplitude collapse problem that occurs after the lightweight visual language motion model is adapted to the target task, but also improves the closed-loop execution efficiency, task completion rate, and task execution reliability of the robotic arm. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall environment for closed-loop control of robotic arm task execution provided in an embodiment of the present invention;

[0034] Figure 2 The complete flowchart of the staged amplitude correction control method provided in the embodiments of the present invention is shown. Detailed Implementation

[0035] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0036] like Figure 1As shown, this embodiment takes the closed-loop grasping control process of a robotic arm as the application scenario. The basic motion generation strategy adopts the SmolVLA lightweight VLA model. The focus is on how, after the basic motion generation strategy outputs the basic motion, the calibration control module generates the final execution motion adapted to the current task stage and converts it into control instructions that can be executed by the robotic arm controller.

[0037] The method in this embodiment is generally divided into an offline processing stage, a correction module training stage, and an online closed-loop control stage, such as... Figure 2 As shown, it specifically includes:

[0038] Offline processing stage: The system acquires the target robotic arm task data, performs amplitude statistics on the position control components and attitude control components in the reference motion labels, and divides the motion amplitude stages according to the change law of motion amplitude with the task progress, obtains the stage boundaries and the corresponding motion amplitude statistics of each stage, and generates online geometric switching thresholds.

[0039] Correction module training phase: Based on the motion amplitude phase, the system constructs a correction control module between the motion output end of the basic motion generation strategy and the robotic arm controller, freezes the basic VLA model parameters, and trains the correction module using course-based PPO reinforcement learning, so that it can output matching position scale gating coefficients, attitude scale gating coefficients and restricted residual corrections according to different task phases.

[0040] In the online closed-loop control phase: the system acquires the current visual observation, robotic arm status and task language instructions in each control cycle, and outputs basic actions by the SmolVLA model; then the trained correction control module performs amplitude correction on the basic actions to obtain the final execution action; subsequently, the final execution action is converted into end-effector pose control instructions, joint control instructions, servo control target or gripper control instructions, and sent to the robotic arm controller for execution, and the above process is repeated until the task is completed or the preset termination condition is met.

[0041] The method in this embodiment can introduce the phased motion amplitude patterns in the target task data into the online control process of the robotic arm without updating the main parameters of the basic motion generation strategy. This allows the robotic arm to obtain more matched motion scales in different execution stages such as approach, fine-tuning, posture alignment, grasping, placement, or contact adjustment, thereby improving the stability, execution efficiency, and task completion reliability of the robotic arm's closed-loop control process. The specific steps are explained below.

[0042] Step 1: Offline phase task data processing, namely target robotic arm task data analysis and motion amplitude phase division.

[0043] This section analyzes the variation patterns of motion amplitude in the target robotic arm's task data and obtains the stage division results required for subsequent motion amplitude correction. Specifically, a motion amplitude sequence and a global motion scale curve are first constructed based on the target robotic arm's task data. Then, the motion amplitude stage boundaries are determined according to the global motion scale curve, and the offline stage boundaries are converted into stage switching rules usable in the online control process.

[0044] Step 1.1: Construction of motion amplitude sequence and global scale curve analysis.

[0045] In this embodiment, target robotic arm task data is first constructed or acquired. This target robotic arm task data can be a single semantic task dataset, meaning that each trajectory corresponds to the same or essentially the same task language instructions and operational objectives, differing only in the initial pose of the robotic arm, target position, object posture, camera observation, or environmental perturbations. Optionally, the task data can be used to generate or assist in generating a basic reference trajectory using inverse kinematics methods, and controlled perturbations can be applied to the pose step length, end-effector displacement increment, posture increment, or joint change amplitude to obtain robotic arm task trajectories with different motion amplitude distributions. Through this method, a data foundation reflecting the changes in motion scale during the robotic arm task execution can be obtained while maintaining relative consistency in the task objectives.

[0046] Each trajectory in the target robotic arm task data includes multiple consecutive time steps. Each time step includes at least visual observations, robotic arm status, task language commands, and reference action labels. The visual observations may include first-person view images and wrist camera images. The robotic arm status describes the current end effector and gripper status. The reference action labels describe the end effector control actions and gripper control actions for the next control step.

[0047] After obtaining the target fine-tuning dataset, the motion amplitude is calculated for the reference motion label in each trajectory. Let the... The trajectory in the first The action vector for each time step is:

[0048] ;

[0049] in, Indicates the end position control component. Indicates the end-effector attitude control components. This represents the gripper control component. Since the method in this embodiment primarily focuses on the amplitude changes of continuous position and attitude movements, the amplitudes of translational and rotational movements are calculated separately:

[0050] ;

[0051] .

[0052] Since different trajectories may have different lengths, to facilitate the establishment of a unified stage analysis template within the dataset, the motion amplitude sequence of each trajectory is resampled to a uniform length and mapped onto a normalized task progress axis. Optionally, after resampling, the motion amplitude sequence is smoothed to reduce the impact of single-step noise or abnormal spikes on subsequent stage detection.

[0053] Furthermore, the amplitudes of translational and rotational movements are merged into a unified motion scale score. For example, for the first... The trajectory in the first The action scale score at a normalized progress point can be expressed as:

[0054] ;

[0055] in, and These represent the normalized or standardized amplitudes of translational and rotational movements, respectively. and These are the weighting parameters for translational and rotational motion amplitudes, respectively. Subsequently, the motion scale scores of all trajectories in the dataset on the normalized task progress axis are statistically summarized to obtain the global motion scale curve:

[0056] ;

[0057] in, Indicates the number of trajectories in the dataset. Indicates the first Global motion scale score at a normalized progress point.

[0058] The global motion scale curve is used to characterize the overall trend of motion amplitude changes in the target task dataset as the task progresses, and serves as the basis for subsequent motion amplitude stage division.

[0059] It should be noted that the above-described method for constructing motion scale scores is only one possible implementation. In other implementations, motion scale curves may be constructed based solely on translational motion amplitude, solely on rotational motion amplitude, or a combination of translational motion amplitude, rotational motion amplitude, and other control quantities. This invention does not limit this approach.

[0060] Step 1.2: Boundary detection of motion amplitude stage and construction of online switching rules.

[0061] After obtaining the global motion scale curve, the locations where significant changes occur in the statistical characteristics of motion amplitude during task execution are determined based on the curve, and motion amplitude stages are divided accordingly. Specifically, methods such as change point detection, piecewise fitting, threshold detection, cluster analysis, or manual empirical rules can be used to determine one or more candidate stage boundaries from the global motion scale curve.

[0062] In this embodiment, the global motion scale curve can be divided into several continuous segments according to the task progress, and the mean, variance, or other statistics of the motion scale score within each segment can be calculated. When the difference in motion scale between adjacent segments exceeds a preset threshold, it is considered that there is a change in motion amplitude at the corresponding position, and this position is used as a candidate stage boundary. When the difference in motion scale between adjacent segments is small, or the boundary is mainly caused by local noise, abnormal spikes, or short-term fluctuations, the adjacent segments are merged. A minimum stage length constraint can also be set to avoid generating motion amplitude stages that are too short or have no practical control significance.

[0063] Let the final number of motion amplitude stages be . The corresponding set of stages can be represented as:

[0064] ;

[0065] in, Let i represent the i-th motion amplitude stage, i=1,2,…,N. For each motion amplitude stage, further statistical analysis is performed on the translational motion amplitude range, rotational motion amplitude range, amplitude mean, amplitude quantile, or other amplitude statistics within that stage. This data serves as the basis for determining the subsequent stage-based scale gating range, data consistency constraints, and online motion amplitude correction.

[0066] After obtaining the offline motion amplitude stage, the offline stage boundary is further converted into stage switching rules in the online control process. Since the actual task progress during online robotic arm execution may not perfectly match the time progress of the taught trajectory, if stage switching is based solely on the normalized time progress, situations may arise where the time progress has entered the later stages but the end effector has not yet approached the target. Therefore, the method in this embodiment preferably uses geometric quantities that reflect the actual task state as the basis for online stage switching.

[0067] For proximity-based tasks, the distance between the current position of the end effector and the target position can be selected as the stage switching variable. Let the current position of the end effector be:

[0068] ;

[0069] The target location is:

[0070] ;

[0071] The horizontal distance between the end effector and the target position at the current moment can be expressed as:

[0072] ;

[0073] In this embodiment, the offline stage boundary is mapped back to the corresponding time step of each trajectory in the dataset, and the horizontal plane distance sample at that time step is statistically analyzed; subsequently, the online geometric switching threshold is determined based on the median, mean, or preset quantile of the distance sample. Let the first... The first stage to the second The online geometric switching threshold for each stage is When executing online, the current distance can be used as a reference. With threshold The relationship is used to determine the current action amplitude stage.

[0074] It should be noted that the above method of using horizontal plane distance as the online switching variable is mainly applicable to proximity-type robotic arm operations. In other tasks, three-dimensional position error, attitude error, distance between the gripper and the target object, contact state, force feedback signal, task progress estimate, or a combination of the above information can also be used as the basis for online stage switching, depending on the characteristics of the task.

[0075] By constructing the above-mentioned action amplitude stage boundary detection and online switching rules, the stage-based action amplitude change patterns implicit in the target fine-tuning data can be transformed into online executable stage judgment criteria, providing a foundation for the construction and training of the subsequent stage-aware lightweight gating action amplitude correction module.

[0076] Step 2: Method for constructing the calibration control module.

[0077] This section constructs a correction control module based on the aforementioned motion amplitude stage division results of the target robotic arm task. This module is located between the motion output end of the basic motion generation strategy and the robotic arm controller. After the basic motion is generated, it performs staged correction on the position increment and attitude increment of the basic motion according to the current robotic arm state, the basic motion amplitude, and the motion amplitude stage. It also introduces constrained residual correction in a preset stage to obtain the final execution motion adapted to the current robotic arm task stage.

[0078] Step 2.1: Overall structure and motion synthesis method of the correction module.

[0079] In this embodiment, the basic motion generation strategy outputs basic motions based on the current visual observation, the state of the robotic arm, and the task language instructions. Let the first... The basic actions at each control moment are:

[0080] ;

[0081] in, This represents the increment of the end position in the basic motion. This represents the end-effector pose increment in the basic motion. This indicates the gripper control component.

[0082] The correction control module takes the basic motion and the current robotic arm state characteristics as input, and outputs position-scale gating coefficients and attitude-scale gating coefficients, denoted as follows: and This is used to adjust the position increment and attitude increment amplitudes in the basic motion. For stages where residual correction is not enabled, the correction module only uses scale gating to generate the execution motion:

[0083] ;

[0084] ;

[0085] For preset fine-tuning stages, pre-contact alignment stages, or other stages requiring local compensation, the correction module further outputs constrained residual correction terms. and And superimpose it onto the scale-corrected motion:

[0086] ;

[0087] ;

[0088] Therefore, the final execution action can be uniformly represented as:

[0089] ;

[0090] Among them, the gripper control component The output can be directly based on the basic strategy, or it can be generated by an independent gripper-controlled branch in other implementations. During the non-residual activation phase, the residual correction term can be set to zero, at which point the correction module degenerates into a pure scale-gated motion amplitude correction module.

[0091] Step 2.2: Input feature construction and stage-aware representation.

[0092] To enable the correction module to correct motion amplitude based on the current execution state of the robotic arm, the correction control module uses state features related to motion amplitude correction as input. These state features include at least one of the following: the robot's current end effector state, basic motion, the position norm and attitude norm of the basic motion, target relative position error, attitude error, distance from the end effector to the target position, historical gating coefficients, the amplitude of the motion executed at the previous moment, and the current task progress or stage progress.

[0093] In this embodiment, the input features of the correction module are represented as follows:

[0094] ;

[0095] in, Indicates the status of the robotic arm. Indicates the basic strategy output action. and These represent the basic positional motion amplitude and the basic rotational motion amplitude, respectively. and These represent the target's relative position error and attitude error, respectively. This indicates the horizontal distance between the end effector and the target position. and This represents the position-scale gating factor and rotation-scale gating factor at the previous moment. This indicates the current task progress or stage progress.

[0096] To reduce the dimensional differences between different physical quantities, the input features can be normalized:

[0097] .

[0098] The normalized input features are fed into the shared feature extraction network to obtain the shared state features:

[0099] ;

[0100] in, Indicates a shared feature extraction network. Indicates its network parameters, This indicates a shared state characteristic.

[0101] Suppose the task execution process is divided into: The motion amplitude stage. To enable the correction module to learn the motion amplitude adjustment rules at different stages, a stage-aware branch can be constructed based on shared state features. For the first... The characteristics of each action amplitude stage are represented as follows:

[0102] ;

[0103] in, Indicates the first Branch mapping corresponding to each action amplitude stage Indicates the first Each stage corresponds to a specific stage characteristic. Through this structure, the correction module can generate stage-related motion amplitude correction parameters based on the robotic arm's state and stage information, ensuring that the motion amplitude correction process matches the actual execution stage of the robotic arm.

[0104] Step 2.3: Generation of staged gating range and scale gating.

[0105] After obtaining the motion amplitude stage, the gating range for the corresponding stage is determined based on the motion amplitude statistics of each stage. The gating range is used to constrain the degree to which the correction module amplifies or reduces the basic motion, avoiding an excessively large gating coefficient that leads to unstable motion execution, or an excessively small gating coefficient that leads to further contraction of the motion amplitude.

[0106] For the For each motion amplitude stage, the amplitude distribution of translational and rotational motions in the target robotic arm task data for that stage is statistically analyzed. Based on the mean, median, quantiles, statistical upper and lower bounds, or amplitude ratios, the corresponding position-scale gating range and attitude-scale gating range are determined. Preferably, the gating range is not directly determined by extreme maximum or minimum values, but rather by the median, preset quantiles, or truncated statistical ranges to reduce the impact of abnormal motion values ​​on the gating range.

[0107] In this embodiment, the gating range can be determined based on the mean, median, quantiles, statistical upper and lower bounds, or amplitude ratio of the action amplitude distribution at each stage. For example, when the overall action amplitude output by the base strategy at a certain stage is less than the reference amplitude in the target fine-tuning data for that stage, the scale gating range for that stage can be increased accordingly; when the action amplitude output by the base strategy is close to or exceeds the reference amplitude range, the scale gating range for that stage can be decreased or limited accordingly. Thus, different action amplitude stages can obtain gating ranges that match their action scale statistical characteristics.

[0108] Furthermore, the correction module determines the stage weight corresponding to the current state based on the online stage switching rules. Let the stage weight corresponding to the current moment be:

[0109] ;

[0110] in, This indicates that the current state belongs to the first... The weights of each action amplitude stage, satisfying:

[0111] ;

[0112] The stage weights can be either hard stage labels or soft stage weights. When using soft stage weights, a smooth transition can be made between adjacent stages, reducing abrupt changes in gating parameters near stage boundaries.

[0113] For the For each motion amplitude stage, the correction module adjusts according to the stage characteristics. Generate the position-gated primitive variables and the rotation-gated primitive variables respectively:

[0114] ;

[0115] ;

[0116] in, , , and These are the mapping parameters for the corresponding stage branches. To avoid the network directly outputting unbounded action amplification factors, the original variables are mapped to the gating range of the corresponding stage through bounded functions, resulting in the... Position-scale gating coefficients and attitude-scale gating coefficients for each stage:

[0117] ;

[0118] ;

[0119] in, This represents the Sigmoid function or other mapping functions with limited range of values; , These are the maximum and minimum values ​​of displacement and rotation gating at each stage, determined by the distribution of the dataset.

[0120] Finally, the gating outputs of each stage are weighted and fused according to the current stage weight to obtain the final position-scale gating coefficients and rotation-scale gating coefficients:

[0121] ;

[0122] .

[0123] Therefore, the correction control module can adjust the motion amplitude correction parameters related to the current robotic arm task stage and the robotic arm state generation stage, so that different motion amplitude stages correspond to different correction intensities, thereby avoiding the problem that uniform global scaling is difficult to adapt to multi-stage robotic arm operation tasks.

[0124] Step 2.4: Preset Residual Correction in the Fine-Tuning Stage.

[0125] Scale gating is primarily used to amplify or reduce the amplitude along the direction of the basic strategy action, effectively preserving the main direction of the basic action. However, in close-range fine-tuning, pre-contact alignment, contact adjustment, or other high-precision control stages, scale scaling alone may still be insufficient to compensate for local position or attitude errors. Therefore, the correction control module can further introduce a constrained residual correction term in a preset stage.

[0126] Let the first The original variables of position residual and attitude residual corresponding to each motion amplitude stage are respectively and It is generated from the characteristics of the corresponding stage. To prevent the residual term from becoming too large and covering the basic action direction, the residual term is limited by a bounded function and a maximum residual magnitude:

[0127] ;

[0128] ;

[0129] in, and These represent the values ​​of the first action determined by the action amplitude during the dataset fine-tuning phase. The maximum allowable position residual and rotation residual values ​​for each motion amplitude stage. Used to restrict the original variables of the residuals to a bounded range.

[0130] Set a residual enable flag or residual enable weight for each motion amplitude stage, denoted as . .in, Indicates the first Constrained residual correction is enabled at each stage. Indicates the first In some implementations, restricted residual correction is disabled. It can also be set to Continuous weights within the interval indicate the degree to which residual correction is enabled.

[0131] Based on the stage weights and residual activation flags, the final position residuals and rotation residuals are obtained through fusion:

[0132] ;

[0133] .

[0134] For the long-distance approach or coarse adjustment phase, the residual correction term is set to zero to avoid excessive changes to the basic motion direction; for the close-range fine adjustment phase, pre-contact alignment phase, or high-precision compensation phase, restricted residual correction is enabled to perform small-scale local compensation based on scale gating.

[0135] In this manner, the constrained residual correction is only activated during the preset robotic arm control phase and is subject to the combined constraints of the maximum residual amplitude and the phase activation weight. This design can improve the local compensation capability of the robotic arm during the fine-tuning, alignment, and contact adjustment phases while maintaining the basic motion orientation, thus avoiding sudden motion changes or control instability caused by unconstrained residual output.

[0136] Step 3: Reinforcement learning training of the calibration control module.

[0137] This section is used to train the aforementioned correction control module. During training, the main parameters of the basic motion generation strategy remain frozen. The correction control module, as the module to be trained, outputs correction parameters based on the basic motion, the current state of the robotic arm, the target relative error, and the motion amplitude stage information. Through reinforcement learning training, the correction control module can output matching position scale gating coefficients, attitude scale gating coefficients, and optional restricted residual correction terms under different robotic arm task stages, thereby improving the adaptability of the final executed motion to the current control stage while maintaining the main direction of the basic motion.

[0138] Step 3.1: Training and modeling of the motion amplitude correction control module.

[0139] In this embodiment, the training process of the correction control module is modeled as a Markov decision process:

[0140] ;

[0141] in, Indicates the state or observation space. Indicates the motion correction space. This represents the state transition process jointly determined by robot dynamics, environmental state transitions, and the output of the basic policy. Represents the reward function, This represents the discount factor.

[0142] At each control moment The basic motion generation strategy outputs basic motions based on the current visual observation, the state of the robotic arm, and the task language instructions. During the training phase, the basic policy parameters remain frozen, and the correction policy constructs training observations based on the current state information and basic actions:

[0143] ;

[0144] in, Indicates the status of the robotic arm. Indicates the basic strategy output action. and These represent the positional and rotational amplitudes in the basic motion, respectively. and These represent the current position error and attitude error, respectively. This indicates the horizontal distance between the end effector and the target position. This indicates the current task progress or stage progress. This represents the stage weight obtained based on the aforementioned online geometric switching threshold.

[0145] The action correction space corresponds to the correction parameters output by the correction module. In one implementation, the policy action can be represented as:

[0146] ;

[0147] in, and Used to generate position-scale gating coefficients and rotation-scale gating coefficients. and Used to generate optional constrained position residuals and constrained rotation residuals. For stages where residual correction is not enabled, the residual terms are set to zero through the stage mask and do not participate in the final execution action synthesis.

[0148] Let the motion amplitude correction strategy be ,in, This represents the parameters of the correction strategy. The value function is expressed as... ,in, This represents the value function parameters. After the correction strategy outputs the action correction parameters, the final action is obtained according to the aforementioned action synthesis method. The execution action is then applied to a simulation environment or a real robotic arm environment to obtain the next state, reward signal, and termination flag. This forms the training samples. ;in, Indicates the reward at the current moment. Indicates whether the current episode has terminated. This represents the current stage weight or stage label. Through the above training model, the reinforcement learning process does not directly retrain the basic motion generation strategy, but is used to obtain staged motion amplitude correction parameters that can serve the closed-loop control of the robotic arm.

[0149] Step 3.2: Staged reward function and data consistency constraints.

[0150] To enable the calibration control module to adapt to the control requirements of different stages of the robotic arm's task, this embodiment constructs a staged reward function for the current motion amplitude stage. Specifically, the motion amplitude stage is divided into two categories: a non-fine-tuning stage and a fine-tuning stage. The non-fine-tuning stage mainly corresponds to long-distance approach or mid-distance alignment processes, and its training objective is to enable the end effector to stably approach the target area while maintaining sufficient motion propulsion capability. The fine-tuning stage mainly corresponds to close-range adjustment, attitude alignment, pre-contact adjustment, or other high-precision control processes, and its training objective is to reduce position and attitude errors and constrain motion smoothness and residual amplitude.

[0151] At the current stage When the process is not in the fine-tuning stage, a coarse-grained stage reward function is used:

[0152] ;

[0153] in, The horizontal distance guide item indicates the non-fine-tuning stage and is used to evaluate the proximity based on the horizontal plane distance between the end effector and the target position; This represents a progress auxiliary item for the non-fine-tuning phase, used to describe the change in target distance between adjacent time steps; This indicates a stage entry or stage advancement reward, used to encourage strategies to meet the geometric conditions for entering the next stage; This indicates a rotation penalty term during the non-fine-tuning phase, used to suppress excessive attitude adjustments or unnecessary rotations during the long-distance approach phase. , , and These are the weight coefficients for the corresponding items.

[0154] In this embodiment, the horizontal distance guidance item and the progress assistance item can be represented as:

[0155] ;

[0156] ;

[0157] in, This represents the horizontal distance between the end effector and the target position at the current moment. This represents the horizontal distance at the previous moment. is the scaling factor for the horizontal distance guide term.

[0158] At the current stage When in the fine-tuning phase, the reward function for the fine-tuning phase is used:

[0159] ;

[0160] in, This represents the position accuracy bonus during the fine-tuning phase, used to evaluate the three-dimensional position error between the end effector and the target position. This represents a local progress auxiliary term during the fine-tuning phase, used to describe the reduction in positional error between adjacent time steps; This indicates a mission success reward, used to provide additional rewards when conditions such as target arrival, attitude error, or operation completion are met. This represents the rotation penalty term during the fine-tuning phase, used to suppress rotations that deviate too much from the target value during the fine-tuning phase; This represents a motion smoothing penalty term, used to reduce abrupt changes in the execution of actions or gating coefficients between adjacent time steps; This represents the residual magnitude penalty term, used to limit the magnitude of the restricted residual correction term; , , , , and These are the weight coefficients for the corresponding items.

[0161] Position accuracy bonus and local progress assistance are represented as follows:

[0162] ;

[0163] ;

[0164] in, This represents the three-dimensional positional distance between the end effector and the target position at the current moment. This represents the distance to the three-dimensional position corresponding to the previous moment. The scale factor is used to reward positional accuracy.

[0165] During training, the corresponding stage-specific reward function is selected based on the current stage label. If the current stage is a non-fine-tuning stage, the non-fine-tuning stage reward is used; if the current stage is a fine-tuning stage, the fine-tuning stage reward is used.

[0166] ;

[0167] in, This represents the set of non-fine-tuning stages. This represents the set of fine-tuning stages. Through the aforementioned staged reward selection method, the correction control module can learn motion amplitude compensation rules primarily based on target area proximity during the non-fine-tuning stage, and learn motion correction rules primarily based on position accuracy, attitude alignment, and local stability control during the fine-tuning stage.

[0168] In addition, in order to keep the corrected final execution action within the reasonable motion amplitude range obtained from the target robotic arm task data statistics, the method of this embodiment introduces motion amplitude consistency constraints and gating prior constraints to limit the amplitude, gating coefficient and residual correction amount of the final execution action, thereby reducing the risk of control abrupt changes when the motion is excessively amplified, excessively contracted or switched between stages.

[0169] Dataset consistency loss mainly includes action amplitude consistency loss. and gated prior loss Loss of consistency in motion amplitude This is used to constrain the amplitude of the final executed action to be within the statistical range of the action amplitude of the dataset for the corresponding action amplitude stage. Let the first... The reference range for the positional motion amplitude value corresponding to each motion amplitude stage is as follows: The reference range for the rotational motion amplitude is: The corrected position and rotation amplitudes of the executed action at the current moment are as follows:

[0170] ;

[0171] The loss of motion amplitude consistency is then expressed as:

[0172] ;

[0173] in, Indicates the current time corresponding to the first The stage weight of each action amplitude stage, This represents the weighting coefficient between position amplitude constraints and rotation amplitude constraints. Using the above-described out-of-range penalty method, this item does not impose additional penalty when the final executed action amplitude is within the corresponding stage reference range; when the executed action amplitude is below the lower reference limit or above the upper reference limit, this item penalizes the excess portion, thereby suppressing excessive contraction or amplification of the action amplitude.

[0174] Gated prior loss The scale gating coefficients used to constrain the output of the correction control module are close to the stage gating priors obtained from the statistical analysis of the target robotic arm's task data. Let the first... The position gating prior center and rotation gating prior center for each motion amplitude stage are respectively and Then the definition is:

[0175] ;

[0176] in, and They represent the current time and the i-th time respectively. The position-scale gating coefficients and rotation-scale gating coefficients output by each stage branch. This represents the weighting coefficients between position-gated priors and rotation-gated priors. If the gated priors are given in interval form, the above squared error form can be replaced with an out-of-interval penalty form, applying a penalty only when the gated coefficients exceed the prior range.

[0177] Therefore, the action amplitude consistency loss and gating prior loss can be used as additional constraints in the subsequent optimization objectives, so that the correction control module can improve the task execution effect while keeping the final execution action within a reasonable control range.

[0178] Step 3.3: Update PPO optimization objectives and joint parameters.

[0179] After completing the design of the phased reward function and action amplitude constraints, this embodiment uses a proximal policy optimization algorithm to train the correction control module. The policy network is based on the current observations. Output action correction decision Value networks are used to estimate the expected return in the current state.

[0180] During interactive training, the current strategy is used to collect the control trajectory of the robotic arm, and the time difference error is calculated based on the reward sequence and value estimation.

[0181] ;

[0182] in, This is the discount factor. Subsequently, the advantage function is calculated using generalized advantage estimation:

[0183] ;

[0184] in, To estimate the smoothing coefficient for the advantage, Indicates the trajectory length. Advantage function. It is used to measure the quality of the current action correction decision relative to the value function estimate.

[0185] Let the probability ratio of the new strategy to the old strategy be:

[0186] ;

[0187] The loss of the PPO strategy is:

[0188] ;

[0189] in, Represents policy entropy. This indicates the stage recognition or stage perception auxiliary loss. This indicates the loss of consistency in motion amplitude. Indicates the prior loss of gating. This represents the residual magnitude and the loss due to smoothness constraints. to These are the corresponding weighting coefficients.

[0190] During the parameter update process, the collected online interactive trajectory of the robotic arm is divided into small batches of data, and the old strategy parameters are fixed. Under these conditions, the policy network parameters and value network parameters are updated based on the aforementioned joint loss. The policy network parameters are mainly updated based on the policy loss, entropy regularization, stage auxiliary loss, action amplitude consistency loss, gating prior loss, and residual constraint loss; the value network parameters are mainly updated based on the value network loss.

[0191] After the update is complete, assign the current policy parameters to the old policy parameters:

[0192] ;

[0193] Then, new online interaction trajectories are collected, and the advantage estimation, loss calculation, and parameter update process are repeated until the preset number of training rounds, task success rate threshold, or convergence condition is met.

[0194] By using the above-mentioned PPO optimization objectives and joint parameter update methods, the action amplitude correction strategy can be optimized under the task reward drive. At the same time, the action amplitude distribution that deviates from the original dataset can be limited by the prior constraints of the dataset and the stage-related constraints, thereby reducing the problems of abnormal action amplitude, excessive residual compensation and execution instability.

[0195] Step 3.4: Course-based training and phased adjustments.

[0196] To improve training stability, this embodiment adopts a course-based training method, gradually training the scale-gated branch and the restricted residual correction branch according to the stages of the robotic arm task.

[0197] First, the scale gating capability is trained during the non-fine-tuning phase. In this phase, the restricted residual correction branch is disabled, and only the scale gating branch is trained. This allows the correction control module to learn appropriate position and attitude scale gating coefficients to propel the robotic arm stably from the current motion amplitude phase to the next motion amplitude phase. The rewards for this phase primarily employ horizontal distance guidance, phase entry rewards, and coarse-grained attitude constraints.

[0198] Secondly, the scale gating capability of the subsequent fine-tuning stage is trained. After the non-fine-tuning stage can be entered relatively stably, the policy parameters obtained from the previous stage training are used as initialization. The scale gating branch and value network corresponding to the target motion amplitude stage are mainly trained, so that the robotic arm can further improve its position fine-tuning and attitude alignment capabilities near the target.

[0199] Next, the restricted residual correction branch is trained. Once the scale-gated strategy reaches a preset stability condition within the corresponding stage, the restricted residual correction branch is activated. This residual correction branch is only activated during preset fine-tuning, pre-contact alignment, contact adjustment, or local compensation stages, and is subject to residual amplitude and smoothness constraints. Its training objective is not to relearn the complete motion direction, but rather to compensate for local position, attitude, or execution errors based on the scale-gated strategy.

[0200] Finally, joint fine-tuning of the scale-gated branch and the residual correction branch is performed. During the joint fine-tuning stage, both are optimized simultaneously with a small learning rate, while retaining stage constraints, motion amplitude consistency constraints, gating prior constraints, residual amplitude constraints, and smoothness constraints, so that the correction control module achieves a balance between task performance and the safe range of motion amplitude.

[0201] Each learning stage has corresponding success criteria and training termination criteria. During training, the success of the current training round is determined based on whether the robotic arm reaches the preset position error threshold, posture error threshold, stage entry condition, number of consecutive stable steps, and / or task success status for that stage. Each training checkpoint is evaluated based on one or more indicators, including stage success rate, number of consecutive successes, position and posture errors, stage holding time, and cumulative reward. The checkpoint with the best evaluation result is selected as the optimal strategy parameter for that stage. When the stage success indicator reaches the preset stability condition within a certain number of consecutive training rounds or evaluation rounds, the current learning stage is terminated, and the corresponding optimal strategy parameter is used as the initialization parameter for the next learning stage. If the stability condition is not reached, training continues until the preset maximum number of training rounds, and the optimal checkpoint is selected from the saved checkpoints to enter the next stage.

[0202] Through the aforementioned course-based training and phased joint adjustment method, the scale gating branch is used to adjust the amplitude between different motion amplitude stages, while the constrained residual correction branch is used for small-range local compensation within the high-precision robotic arm control stage. Thus, the correction control module can improve the robotic arm's motion scale adaptability and execution stability during approach, fine-tuning, posture alignment, and contact adjustment stages while maintaining the basic motion's main orientation.

[0203] Step 4: Online inference and closed-loop control of the robotic arm in the post-training correction module.

[0204] After training the correction control module is completed, this embodiment no longer updates the PPO parameters during the online execution phase; the parameters of the basic motion generation strategy and the motion amplitude correction control module remain fixed. The motion amplitude correction control module is positioned between the basic motion generation strategy and the robotic arm controller, and is used to perform staged amplitude correction on the basic motion within each control cycle, generating the final execution motion adapted to the current execution phase.

[0205] Specifically, at each control moment, the current visual observation, robotic arm state, and task target information are acquired, and basic actions are output by a basic action generation strategy. The basic actions include end-effector position increment and end-effector posture increment, which are used to characterize the basic motion direction and basic motion amplitude of the robotic arm end effector.

[0206] Subsequently, the input information for the correction control module is constructed based on the current state of the robotic arm, basic movements, target relative position, posture error, historical gating information, and task progress; and based on the aforementioned stage boundaries, online geometric switching thresholds, and stage weight calculation methods, the current movement amplitude stage is determined, or the stage weights of each movement amplitude stage corresponding to the current moment are calculated.

[0207] Next, the input information is fed into the trained motion amplitude correction control module to obtain the current position scale gating coefficient and attitude scale gating coefficient. If the current stage belongs to a preset fine-tuning stage, contact adjustment stage, or local alignment stage, restricted residual correction is enabled, and restricted position residual and restricted attitude residual are output; if the current stage does not belong to the residual enabling stage, the residual correction amount is set to zero through a stage mask. Thus, restricted residual correction only takes effect in the control stage that requires high-precision local compensation, avoiding excessive changes to the basic motion direction during long-distance approach stages.

[0208] Then, following the aforementioned motion synthesis method, the basic motion, position scale gating coefficient, attitude scale gating coefficient, and restricted residual correction are synthesized into the final executed motion. The final executed motion is converted into the target pose of the robotic arm end effector, joint control commands, speed control commands, or control commands executable by the underlying controller, and sent to the robotic arm controller for execution.

[0209] After the robotic arm completes the action of the current control cycle, the system updates the robotic arm status, the gating coefficient of the previous moment, the amplitude of the action executed in the previous moment, and the necessary historical observation information, and then enters the next control cycle. The above process is repeated until the robotic arm completes the target task or meets the preset termination condition.

[0210] In the aforementioned online closed-loop control process, the basic motion generation strategy is used to generate motion outputs with task semantics and basic motion direction; the motion amplitude correction control module is used to gating and adjusting the position and attitude amplitudes of the basic motion according to the current motion amplitude stage, and to provide constrained residual compensation in the high-precision stage. Stage weights are used to achieve smooth switching between adjacent motion amplitude stages, and stage masks are used to limit the range of residual correction activation.

[0211] Through the above-mentioned online inference and closed-loop control method of the robotic arm after training, this embodiment can achieve phased correction from basic action to final execution action without updating the main structure of the basic action generation strategy. This enables the robotic arm to maintain sufficient motion propulsion capability in the long-distance approach stage and maintain good local control accuracy and execution stability in the close-range fine-tuning, posture alignment and contact adjustment stages.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for staged motion amplitude correction and control of a robotic arm, characterized in that: The overall process is divided into three main parts: offline data processing, phased training of the calibration module, and online closed-loop control. Specifically, it includes the following steps: Step 1: Offline data processing, performing offline task data analysis and motion amplitude stage division for the target robotic arm; First, multiple operation trajectories of the target robotic arm task are acquired. Each operation trajectory includes visual observation, robotic arm status, task language instructions, and reference motion labels. Then, the corresponding translational motion amplitude and rotational motion amplitude are calculated based on the position control components and attitude control components in the reference motion labels. Furthermore, the changes in motion amplitude in different operation trajectories are statistically analyzed, smoothed, and divided into stages to obtain the motion amplitude stages, stage boundaries, and corresponding motion amplitude statistics for each stage during the execution of the robotic arm task. The offline time dimension stage boundaries are converted into online usable geometric switching rules. The horizontal plane distance between the end effector and the target position is selected as the online stage switching variable. The distance samples corresponding to each stage boundary are statistically analyzed to generate stage switching thresholds for online closed-loop real-time control. Step 2: During the training phase of the calibration module, construct the phase-aware motion amplitude calibration control module; After determining the motion amplitude stage, the correction control module is set between the motion output end of the basic motion generation strategy and the robotic arm controller. The inputs to the correction control module include the robotic arm state, basic motion, basic motion position / attitude amplitude norm, target position error, target attitude error, distance between the end effector and the target plane, gating coefficient of the previous cycle, and weight of the current stage. After normalization, all input features are fed into a shared feature extraction network to obtain shared state features. An independent stage branch network is set for each motion amplitude stage to generate gating raw variables and residual raw variables for each stage, thereby performing staged amplitude correction on the position increment and attitude increment of the basic motion. The original gating variables are mapped to the corresponding pre-defined gating intervals via a Sigmoid function, and then weighted and fused using soft-stage weights in the 0-1 interval to obtain the global position-scale gating coefficients. Attitude scale gating coefficient Adjacent stages rely on soft weights for smooth transitions, avoiding gating abrupt changes. The original residual variables are subject to dual constraints by combining the maximum residual amplitude of each stage with the tanh function. The residual terms are only enabled in the fine-tuning, attitude alignment and contact adjustment stages, while the residuals are set to zero in the long-distance approach stage. Execution of action composition logic: Non-residual activation phase: , ; Residual activation phase: , ; in, and These represent the actual end-effector position and end-effector attitude increment, respectively. and These represent the end-effector position and end-effector pose increments in the basic motion, respectively. and This is a restricted residual correction term; The gripper control components directly use the basic strategy output; Step 3: Freeze the base model and train the correction control module using reinforcement learning; Keeping the main parameters of the basic action generation strategy unchanged, only the correction control module is trained; the training process uses the PPO proximal policy optimization reinforcement learning algorithm to optimize and model the correction control module as a Markov decision process; Two types of stage-differentiated reward functions are set up: coarse-grained reward is used in the non-fine-tuning coarse stage to reduce the target horizontal distance, promote task stage switching, and suppress invalid rotation; fine reward is used in the fine-tuning high-precision stage to reduce 3D pose error, reward task success, and at the same time increase rotation penalty, motion smoothing penalty, and residual amplitude penalty. The total optimization loss includes PPO policy loss, value loss, policy entropy regularization term, stage identification auxiliary loss, action amplitude consistency loss, gating prior loss, and residual smoothing constraint loss; the action amplitude consistency loss penalizes actions whose amplitudes exceed the amplitude range of the corresponding stage dataset after correction; the gating prior loss constrains the gating coefficients of each stage to fit the offline statistical stage gating prior. The training adopts a course-based step-by-step training process: the first step is to close the residual correction branch and train only the scale gating branch; the second step is to open the high-precision stage residual branch and train the residual network separately; the third step is to reduce the learning rate and jointly fine-tune all branches of gating and residual until the model converges. Step 4: Use the trained calibration control module to perform online closed-loop real-time control of the robotic arm; After training is completed, in each online control cycle, the current visual observation, robotic arm status, and task language instructions are first acquired, and basic actions are generated by the basic action generation strategy. Then, based on the current stage switching threshold for the distance between the robotic arm end effector and the target plane, the soft weights of each stage are calculated, and the amplitude of the basic actions is corrected by the trained correction control module to obtain the final execution action. Finally, the final execution action is converted into end effector pose control instructions, joint control instructions, servo control target or gripper control instructions, and sent to the robotic arm controller for execution. The above process is repeated until the task is completed. The basic motion generation strategy is the lightweight VLA model SmolVLA, or other visual language motion models or robot strategy models that can receive visual observations, language commands, and robotic arm states and output continuous robot actions.

2. The method for staged motion amplitude correction and control of a robotic arm according to claim 1, characterized in that: The specific method for step 1 is as follows: Step 1.1: Construction of motion amplitude sequence and global scale curve analysis; First, obtain multiple operation trajectories of the target robotic arm task. Each trajectory corresponds to the same or basically the same task language instructions and operation target. Each trajectory in the target robotic arm task includes multiple consecutive time steps. Each time step includes at least visual observation, robotic arm state, task language instructions, and reference action labels. The visual observation includes a main view image and a wrist camera image. The robotic arm state is used to describe the current end effector state and gripper state. The reference action labels are used to describe the end effector control action and gripper control action in the next control step. Calculate the motion amplitude for the reference motion label in each trajectory; let the first... The trajectory in the first The action vector for each time step is: ； in, Indicates the end position control component. Indicates the end-effector attitude control components. Indicates the gripper control component; Calculate the amplitude of translational motion and the amplitude of rotational motion separately: ； ； The motion amplitude sequence of each trajectory is resampled to a uniform length and mapped onto a normalized task progress axis; after resampling, the motion amplitude sequence is smoothed. The amplitude of translational movement and the amplitude of rotational movement are combined into a unified motion scale score; for the first The trajectory in the first The action scale score at each normalized progress point is represented as follows: ； in, and These represent the normalized or standardized amplitudes of translational and rotational movements, respectively. and These are the weighting parameters for the translational and rotational motion amplitudes, respectively. Subsequently, the motion scale scores of all trajectories on the normalized task progress axis are statistically summarized to obtain the global motion scale curve: ； in, Indicates the number of trajectories in the dataset. Indicates the first Global action scale score at a normalized progress point; The global motion scale curve is used to characterize the overall trend of motion amplitude changes in the target task dataset as the task progresses, and serves as the basis for subsequent motion amplitude stage division. Step 1.2: Boundary detection of motion amplitude stage and construction of online switching rules; After obtaining the global motion scale curve, the location where the statistical characteristics of motion amplitude change significantly during task execution is determined based on the curve, and the motion amplitude stages are divided accordingly. Specifically, one or more candidate stage boundaries are determined from the global motion scale curve by means of change point detection, piecewise fitting, threshold detection, cluster analysis or manual experience rules. The final number of motion amplitude stages is The corresponding set of stages is represented as: ； in, This represents the amplitude stage of the i-th action, where i = 1, 2, ..., N; For each motion amplitude stage, the range of translational motion amplitude, range of rotational motion amplitude, mean amplitude, amplitude quantile, or other amplitude statistics within that stage are further statistically analyzed to serve as the basis for determining the subsequent staged scale gating range, data consistency constraints, and online motion amplitude correction. After obtaining the offline action amplitude stage, the offline stage boundary is further converted into stage switching rules in the online control process, and geometric quantities that can reflect the actual task state are used as the basis for online stage switching. For proximity-type tasks, the distance between the current position of the end effector and the target position is selected as the stage switching variable; let the current position of the end effector be: ； The target location is: ； The horizontal distance between the end effector and the target position at the current moment is expressed as: ； The offline phase boundaries are mapped back to the corresponding time steps of each trajectory, and the horizontal plane distance samples at the corresponding time steps are statistically analyzed. Subsequently, the online geometric switching threshold is determined based on the median, mean, or preset quantile of the distance samples. Let the first... The first stage to the second The online geometric switching threshold for each stage is When executing online, the distance to the current horizontal plane will be used as a reference. With threshold The relationship is used to determine the current action amplitude stage.

3. The method for staged motion amplitude correction and control of a robotic arm according to claim 2, characterized in that: In step 1.1, multiple operation trajectories of the target robotic arm task are generated or assisted in generating basic reference trajectories using inverse kinematics methods, and controlled perturbations are applied to the pose step length, end displacement increment, attitude increment, or joint change amplitude to obtain robotic arm task trajectories with different motion amplitude distributions. Through the above methods, while maintaining the relative consistency of the task target, a data foundation that can reflect the motion scale change law during the execution of the robotic arm task is obtained.

4. The method for staged motion amplitude correction and control of a robotic arm according to claim 2, characterized in that: In step 1.2, the method for determining the candidate stage boundary is as follows: the global motion scale curve is divided into several continuous segments according to the task progress, and the mean, variance or other statistics of the motion scale score in each segment are calculated; when the difference in motion scale between adjacent segments exceeds a preset threshold, it is considered that there is a change in motion amplitude at the corresponding position, and this position is taken as the candidate stage boundary; when the difference in motion scale between adjacent segments does not exceed the preset threshold, or when the boundary is mainly caused by local noise, abnormal peaks or short-term fluctuations, the adjacent segments are merged; and a minimum stage length constraint is set to avoid generating motion amplitude stages that are too short or have no practical control significance.

5. The method for staged motion amplitude correction and control of a robotic arm according to claim 2, characterized in that: The specific method for step 2 is as follows: Step 2.1: Overall structure and motion synthesis method of the calibration module; The basic motion generation strategy outputs basic motions based on current visual observations, the state of the robotic arm, and task language instructions; let the first... The basic actions at each control moment are: ； in, This represents the increment of the end position in the basic motion. This represents the end-effector pose increment in the basic motion. Indicates the gripper control component; The correction control module takes the basic motion and the current state characteristics of the robotic arm as input, and outputs position-scale gating coefficients and attitude-scale gating coefficients, denoted as follows: and It is used to adjust the position increment amplitude and attitude increment amplitude in the basic motion; For the non-residual activation phase, only scale gating is used to generate execution actions: ； ； During the non-residual activation phase, the residual correction term is set to zero, at which point the correction module degenerates into a pure scale-gated motion amplitude correction module. For preset fine-tuning stages, pre-contact alignment stages, or other stages requiring local compensation, a constrained residual correction term is further output. and And superimpose it onto the scale-corrected motion: ； ； The final action is uniformly represented as: ； Among them, the gripper control component The output can be directly based on the basic strategy or generated by a branch controlled by an independent gripper. Step 2.2: Input feature construction and stage-aware representation; The correction control module uses state features related to motion amplitude correction as input, including at least one of the following: the robot's current end-effector state, basic motion, position norm and attitude norm of the basic motion, target relative position error, attitude error, distance from the end-effector to the target position, historical gating coefficients, the amplitude of the motion executed at the previous moment, and the current task progress or stage progress; the input features of the correction module are represented as follows: ； in, Indicates the status of the robotic arm. Indicates the basic strategy output action. and These represent the basic positional motion amplitude and the basic rotational motion amplitude, respectively. and These represent the target's relative position error and attitude error, respectively. This indicates the horizontal distance between the end effector and the target position. and This represents the position-scale gating factor and rotation-scale gating factor at the previous moment. Indicates the current task progress or stage progress; To reduce the dimensional differences between different physical quantities, the input features are normalized: ； The normalized input features are fed into the shared feature extraction network to obtain the shared state features: ； in, Indicates a shared feature extraction network. Indicates its network parameters, Indicates shared state characteristics; Suppose the task execution process is divided into: The first motion amplitude stage; in order for the correction module to learn the motion amplitude adjustment rules under different stages, a stage perception branch is constructed based on shared state features; for the first... The characteristics of each action amplitude stage are represented as follows: ； in, Indicates the first Branch mapping corresponding to each action amplitude stage Indicates the first The characteristics of each stage; Step 2.3: Generation of staged gating range and scale gating; After obtaining the motion amplitude stage, the gating range of the corresponding stage is determined based on the motion amplitude statistics of each motion amplitude stage. This is used to constrain the degree of amplification or reduction of the basic motion by the correction module, so as to avoid the execution of the motion being unstable due to the gating coefficient being too large, or the motion amplitude being further reduced due to the gating coefficient being too small. For the For each motion amplitude stage, the amplitude distribution of translational motion and rotational motion in the target robotic arm task data for that stage are statistically analyzed, and the corresponding position scale gating range and attitude scale gating range are determined based on the amplitude mean, median, quantile, statistical upper and lower bounds or amplitude ratio. The correction module determines the stage weight corresponding to the current state based on the online stage switching rules; the stage weight corresponding to the current moment is: ； in, This indicates that the current state belongs to the first... The weights of each action amplitude stage, satisfying: ； The stage weights can be either hard stage labels or soft stage weights. When using soft stage weights, a smooth transition is made between adjacent stages to reduce abrupt changes in gating parameters near stage boundaries. For the For each motion amplitude stage, the correction module adjusts according to the stage characteristics. Generate the position-gated primitive variables and the rotation-gated primitive variables respectively: ； ； in, , , and These are the mapping parameters for the corresponding stage branches; To avoid the network directly outputting unbounded action amplification factors, the original variables are mapped to the gating range of the corresponding stage using bounded functions, thus obtaining the first... Position-scale gating coefficients and attitude-scale gating coefficients for each stage: ； ； in, This represents the Sigmoid function or other mapping functions with limited range of values; , These are the maximum and minimum values ​​of displacement and rotational gating at each stage, determined by the dataset distribution. Finally, the gating outputs of each stage are weighted and fused according to the current stage weight to obtain the final position-scale gating coefficients and rotation-scale gating coefficients: ； ； Step 2.4: Preset restricted residual correction during the fine-tuning stage; Let the first The original variables of position residual and attitude residual corresponding to each motion amplitude stage are respectively and It is generated from the characteristics of the corresponding stage; to prevent the residual term from being too large and covering the basic action direction, the residual term is limited by a bounded function and a maximum residual magnitude: ； ； in, and These represent the values ​​of the first action determined by the action amplitude during the dataset fine-tuning phase. The maximum allowable position residual and rotation residual values ​​for each motion amplitude stage. Used to restrict the original variables of the residuals to a bounded range; Set a residual enable flag or residual enable weight for each motion amplitude stage, denoted as . ;in, Indicates the first Constrained residual correction is enabled at each stage. Indicates the first Disable restricted residual correction in each stage; Based on the stage weights and residual activation flags, the final position residuals and rotation residuals are obtained through fusion: ； ； For the long-distance approach or coarse adjustment phase, the residual correction term is set to zero to avoid excessive changes to the basic motion direction; for the close-range fine adjustment phase, pre-contact alignment phase, or high-precision compensation phase, restricted residual correction is enabled to perform small-scale local compensation based on scale gating.

6. The method for staged motion amplitude correction and control of a robotic arm according to claim 5, characterized in that: The specific method for step 3 is as follows: Step 3.1: Training and modeling of the motion amplitude correction control module; The training process of the correction control module is modeled as a Markov decision process: ； in, Indicates the state or observation space. Indicates the motion correction space. This represents the state transition process jointly determined by robot dynamics, environmental state transitions, and the output of the basic policy. Represents the reward function, Indicates the discount factor; At each control moment The basic motion generation strategy outputs basic motions based on the current visual observation, the state of the robotic arm, and the task language instructions. During the training phase, the basic policy parameters remain frozen, and the correction policy constructs training observations based on the current state information and basic actions. ； in, Indicates the status of the robotic arm. Indicates the basic strategy output action. and These represent the positional and rotational amplitudes in the basic motion, respectively. and These represent the current position error and attitude error, respectively. This indicates the horizontal distance between the end effector and the target position. This indicates the current task progress or stage progress. This represents the stage weight obtained based on the aforementioned online geometric switching threshold; The action correction space corresponds to the correction parameters output by the correction module; the strategy action is represented as: ； in, and Used to generate position-scale gating coefficients and rotation-scale gating coefficients. and Used to generate optional constrained position residuals and constrained rotation residuals; for stages where residual correction is not enabled, the residual terms are set to zero through the stage mask and do not participate in the final execution action synthesis; Let the motion amplitude correction strategy be ,in, The parameters represent the correction strategy; the value function is expressed as... ,in, The value function parameters are represented; after the correction strategy outputs the action correction parameters, the final execution action is obtained according to the action synthesis method in step 2. The execution action is then applied to a simulation environment or a real robotic arm environment to obtain the next state, reward signal, and termination flag; this forms the training samples. ,in, Indicates the reward at the current moment. Indicates whether the current episode has terminated. Indicates the current stage weight or stage label; Step 3.2: Staged reward function and data consistency constraints; The current motion amplitude stage constructs a staged reward function, dividing the motion amplitude stage into two categories: non-fine-tuning stage and fine-tuning stage. The non-fine-tuning stage corresponds to the long-distance approach or mid-distance alignment process, and its training objective is to enable the end effector to stably approach the target area and maintain sufficient motion propulsion capability. The fine-tuning stage corresponds to the close-range adjustment, attitude alignment, pre-contact adjustment or other high-precision control processes, and its training objective is to reduce position and attitude errors and constrain motion smoothness and residual amplitude. At the current stage When the process is not in the fine-tuning stage, a coarse-grained stage reward function is used: ； in, The horizontal distance guide item indicates the non-fine-tuning stage and is used to evaluate the proximity based on the horizontal plane distance between the end effector and the target position; This represents a progress auxiliary item for the non-fine-tuning phase, used to describe the change in target distance between adjacent time steps; This indicates a stage entry or stage advancement reward, used to encourage strategies to meet the geometric conditions for entering the next stage; This indicates a rotation penalty term during the non-fine-tuning phase, used to suppress excessive attitude adjustments or unnecessary rotations during the long-distance approach phase. , , and These are the weight coefficients for the corresponding items; The horizontal distance guide item and the progress aid item are represented as follows: ； ； in, This represents the horizontal distance between the end effector and the target position at the current moment. This represents the horizontal distance at the previous moment. The scaling factor for the horizontal distance guide term; At the current stage When in the fine-tuning phase, the reward function for the fine-tuning phase is used: ； in, This represents the position accuracy bonus during the fine-tuning phase, used to evaluate the three-dimensional position error between the end effector and the target position. This represents a local progress auxiliary term during the fine-tuning phase, used to describe the reduction in positional error between adjacent time steps; This indicates a mission success reward, used to provide additional rewards when conditions such as target arrival, attitude error, or operation completion are met. This represents the rotation penalty term during the fine-tuning phase, used to suppress rotations that deviate too much from the target value during the fine-tuning phase; This represents a motion smoothing penalty term, used to reduce abrupt changes in the execution of actions or gating coefficients between adjacent time steps; This represents the residual magnitude penalty term, used to limit the magnitude of the restricted residual correction term; , , , , and These are the weight coefficients for the corresponding items; Position accuracy bonus and local progress assistance are represented as follows: ； ； in, This represents the three-dimensional positional distance between the end effector and the target position at the current moment. This represents the distance to the three-dimensional position corresponding to the previous moment. The scale coefficient is used to reward positional accuracy. During training, the corresponding staged reward function is selected based on the current stage label; if the current stage is a non-fine-tuning stage, the non-fine-tuning stage reward is used; if the current stage is a fine-tuning stage, the fine-tuning stage reward is used. ； in, This represents the set of non-fine-tuning stages. This represents the set of fine-tuning stages; Introducing action amplitude consistency constraints and gating prior constraints to limit the amplitude, gating coefficient and residual correction amount of the final executed action, reducing the risk of control abrupt changes when actions are excessively amplified, excessively contracted or switched between stages; Dataset consistency loss includes action amplitude consistency loss. and gated prior loss Loss of consistency in motion amplitude This is used to constrain the amplitude of the final executed action to be within the statistical range of the action amplitude of the dataset for the corresponding action amplitude stage; let the first... The reference range for the positional motion amplitude value corresponding to each motion amplitude stage is as follows: The reference range for the rotational motion amplitude is: The corrected position and rotation amplitudes of the executed action at the current moment are as follows: ； The loss of motion amplitude consistency is then expressed as: ； in, Indicates the current time corresponding to the first The stage weight of each action amplitude stage, This represents the weighting coefficient between position amplitude constraints and rotation amplitude constraints. Through the above-mentioned out-of-range penalty method, when the final executed action amplitude is within the reference range of the corresponding stage, this item does not apply additional penalty; when the executed action amplitude is lower than the reference lower limit or higher than the reference upper limit, this item penalizes the excess part, thereby suppressing excessive contraction or excessive amplification of the action amplitude. Gated prior loss The scale gating coefficients used to constrain the output of the correction control module are close to the stage gating priors obtained from the statistical analysis of the target robotic arm task data; let the first... The position gating prior center and rotation gating prior center for each motion amplitude stage are respectively and Then the definition is: ； in, and They represent the current time and the i-th time respectively. The position-scale gating coefficients and rotation-scale gating coefficients output by each stage branch. This represents the weighting coefficients between position-gated priors and rotation-gated priors. If the gated priors are given in interval form, the squared error form is replaced with an out-of-interval penalty form, and a penalty is applied only when the gated coefficients exceed the prior range. Step 3.3: Update PPO optimization objective and joint parameters; The correction control module is trained using a near-end policy optimization algorithm; the policy network is based on current observations. Output action correction decision Value networks are used to estimate expected returns in the current state; During interactive training, the current strategy is used to collect the control trajectory of the robotic arm, and the time difference error is calculated based on the reward sequence and value estimation. ； in, Discount factor; Subsequently, the advantage function is calculated using generalized advantage estimation: ； in, To estimate the smoothing coefficient for the advantage, Represents trajectory length; dominance function Used to measure the quality of the current action correction decision relative to the value function estimate; Let the probability ratio of the new strategy to the old strategy be: ； The loss of the PPO strategy is: ； in, Represents policy entropy. This indicates the stage recognition or stage perception auxiliary loss. This indicates the loss of consistency in motion amplitude. Indicates the prior loss of gating. This represents the residual magnitude and smoothness constraint loss. to These are the corresponding weighting coefficients; During the parameter update process, the collected online interactive trajectory of the robotic arm is divided into small batches of data, and the old strategy parameters are fixed. Under these conditions, the policy network parameters and value network parameters are updated based on the above joint loss; the policy network parameters are updated based on the policy loss, entropy regularization, stage auxiliary loss, action amplitude consistency loss, gated prior loss, and residual constraint loss; the value network parameters are updated based on the value network loss. After the update is complete, assign the current policy parameters to the old policy parameters: ； Then, new online interaction trajectories are collected, and the advantage estimation, loss calculation, and parameter update process are repeated until the preset number of training rounds, task success rate threshold, or convergence condition is met.

7. The method for staged motion amplitude correction and control of a robotic arm according to claim 6, characterized in that: In step 3, to improve training stability, a course-based training method is adopted, and the scale-gated branch and the restricted residual correction branch are trained step by step according to the stage of the robotic arm task. First, the scale gating capability is trained in the non-fine-tuning stage; the restricted residual correction branch is closed, and only the scale gating branch is trained, so that the correction control module learns the appropriate position scale gating coefficient and attitude scale gating coefficient to drive the robot arm to stably enter the next motion amplitude stage from the current motion amplitude stage; the reward in this stage adopts horizontal distance guidance, stage entry reward and coarse-grained attitude constraint. Secondly, the scale gating capability of the subsequent fine-tuning stage is trained. After the non-fine-tuning stage can be entered relatively stably, the policy parameters obtained from the previous stage training are used as initialization to train the scale gating branch and value network corresponding to the target motion amplitude stage, so that the robotic arm can further improve its position fine-tuning and attitude alignment capabilities near the target. Next, training the constrained residual correction branch; Once the scale-gating strategy reaches the preset stability condition within the corresponding stage, the restricted residual correction branch is activated. The residual correction branch is only activated in the preset fine-tuning stage, pre-contact alignment stage, contact adjustment stage, or local compensation stage, and is subject to residual amplitude constraints and smoothness constraints. Its training objective is to compensate for local position errors, attitude errors, or execution errors based on scale gating. Finally, joint fine-tuning of the scale-gated branch and the residual correction branch is performed. During the joint fine-tuning stage, the scale-gated branch and the residual correction branch are optimized simultaneously with a small learning rate, while retaining stage constraints, motion amplitude consistency constraints, gating prior constraints, residual amplitude constraints and smoothness constraints, so that the correction control module achieves a balance between task execution effect and motion amplitude safety range. Each learning stage has corresponding success criteria and training termination criteria. During training, the success of the current training round is determined based on whether the robotic arm reaches the preset position error threshold, posture error threshold, stage entry condition, number of consecutive stable steps, and / or task success status for that stage. Each training checkpoint is evaluated based on one or more indicators, including stage success rate, number of consecutive successes, position and posture errors, stage holding time, and cumulative reward. The checkpoint with the best evaluation result is selected as the optimal strategy parameter for that stage. When the stage success indicator reaches the preset stability condition within a certain number of consecutive training rounds or evaluation rounds, the current course learning stage is terminated, and the corresponding optimal strategy parameter is used as the initialization parameter for the next course learning stage. If the stability condition is not met, training continues until the preset maximum number of training rounds, and the optimal checkpoint is selected from the saved checkpoints to enter the next stage.

Images