A redundant robot arm teleoperation method and system with adaptive learning and null-space regularization

Abstract

The application relates to the technical field of redundant mechanical arms, in particular to a redundant mechanical arm teleoperation method and system based on adaptive learning and null space regularization. The method comprises reading a current joint configuration and an end position, and determining an end error; a first joint increment is calculated according to the end error; a target reference joint configuration is selected in a nearest neighbor selection manner in a dynamically updated reference joint configuration set; a null space projection matrix of the current joint configuration is constructed to determine a second joint increment that meets preset end task requirements; and a joint configuration of a next operation round is updated based on the first joint increment, the second joint increment and the current joint configuration. The application can form a complete closed loop of "input collection-end solution-preference observation-set update-null space adjustment-execution feedback", and improve the control accuracy and stability of the redundant mechanical arm.

Description

Technical Field

[0001] This application relates to the field of redundant robotic arm technology, and in particular to a method and system for teleoperating a redundant robotic arm with adaptive learning and zero-space regularization. Background Technology

[0002] Teleoperation technology is a technology for robotic arms, and its demand is growing rapidly in two types of scenarios: one is direct remote control (surgical robots, nuclear power plant maintenance, etc.); the other is robot imitation learning data collection, which requires a lot of manual teleoperation demonstrations.

[0003] However, despite the broad application prospects of teleoperation technology, it still has certain shortcomings in terms of accuracy and stability, especially when handling complex tasks with redundant robotic arms. Operators often face problems such as joint motion oscillations and abrupt changes. These problems are particularly evident in the operation of redundant robotic arms. Furthermore, operators cannot guarantee that the manipulated robotic arm will perform the optimal pose during task completion. The limitations of existing technologies become even more prominent, especially when high-precision operations are required. The main drawbacks of existing technologies are the static and fixed joint configurations, which cannot adapt to individual operator differences, and the discontinuous joint configuration switching, resulting in a poor user experience. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a method and system for teleoperating a redundant robotic arm with adaptive learning and zero-space regularization, which can solve at least one of the problems in the prior art.

[0005] According to one aspect of the embodiments of this application, a method for teleoperating a redundant robotic arm with adaptive learning and null space regularization is proposed, the method comprising: Read the current joint configuration of the redundant robotic arm and the end effector pose under the current joint configuration, and determine the end effector error based on the end effector pose and the preset desired end effector pose; The first joint increment that meets the preset end-task requirements is calculated based on the end-task error. In the dynamically updated set of reference joint configurations, the target reference joint configuration is selected by nearest neighbor selection. Construct the null space projection matrix of the current joint configuration, and determine the second joint increment that meets the preset end-effector task requirements based on the preset null space correction gain, the null space projection matrix, the target reference joint configuration, and the current joint configuration. The joint configuration for the next operation round is updated based on the first step joint increment, the second step joint increment, and the current joint configuration. The reference joint configuration set is dynamically updated in the following manner: The preference signal for the redundant robotic arm under the target joint configuration is collected, and the preference signal includes the number of joint velocity direction reversals and the magnitude of the correction behavior. The decision to include the target joint configuration in the candidate joint configurations is based on the number of joint velocity direction reversals and the magnitude of the correction behavior. The feasibility of the candidate joint configurations is verified, and the candidate joint configurations that pass the feasibility verification are added to the reference joint configuration set. The number of reference joint configurations in the reference joint configuration set does not exceed a preset number threshold. For any reference joint configuration in the set of reference joint configurations, if the number of corrections to the reference joint configuration within a preset time window is greater than a preset correction threshold, then the reference joint configuration is removed from the set of reference joint configurations.

[0006] In the above scheme, determining the end effector error based on the end effector pose and the preset desired end effector pose includes: If the redundant robotic arm is currently in a three-dimensional input mode, the difference between the desired end-effector pose and the end-effector pose is taken as the end-effector error. If the redundant robotic arm is currently in master-slave mode, the product of the read master joint increment of the redundant robotic arm and the preset mapping matrix is ​​used as the end effector error.

[0007] In the above scheme, the first step of calculating the joint increment that meets the preset end-task requirements based on the end-task error includes: If the redundant robotic arm is currently in the three-dimensional input mode, the first step joint increment is obtained by performing a pseudo-inverse solution using damped least squares and the end-effector error. If the redundant robotic arm is currently in the master-slave mode, the first step joint increment is determined by multiplying the master arm joint increment with the preset mapping matrix.

[0008] In the above scheme, the target reference joint configuration is selected from the dynamically updated set of reference joint configurations using a nearest neighbor selection method, including: Among the set of reference joint configurations, the target reference joint configuration with the smallest distance from all other reference joint configurations except itself is selected.

[0009] In the above scheme, updating the joint configuration for the next operation round based on the first-step joint increment, the second-step joint increment, and the current joint configuration includes: The final joint increment for the current joint configuration is determined based on the first step joint increment and the second step joint increment. The final joint increment is superimposed on the current joint configuration to obtain the joint configuration for the next operation round.

[0010] In the above scheme, the method further includes: Identify the task type of the redundant robotic arm in the current joint configuration; Determine the standard joint configuration based on the task type; The normalized joint configuration is calculated based on the current joint configuration and the standard joint configuration. The standardized joint configuration, the end-effector pose, the task type, and the motion commands for the redundant robotic arm are compiled into sample data and stored.

[0011] In the above scheme, the step of calculating the normalized joint configuration based on the current joint configuration and the standard joint configuration includes: Determine the null space projection matrix of the current joint configuration; Based on the standard joint configuration, the current joint configuration, and the null projection matrix, a direction vector is determined to point the current joint configuration toward the standard joint configuration in order to extract the components of the undisturbed end-effector pose. The components of the undisturbed end pose are superimposed on the current joint configuration to obtain the normalized joint configuration.

[0012] In the above scheme, the feasibility verification of the candidate joint configuration includes: Determine the minimum singular value corresponding to the candidate joint configuration; Determine the joint angle vectors of each joint in the candidate joint configuration. The feasibility of the candidate joint configuration is verified by comparing each of the joint angle vectors with a preset limit margin and by comparing the minimum singular value with a preset singular threshold. If each of the joint angle vectors is less than the preset limit margin, and the minimum singular value is greater than the preset singular threshold, then the feasibility verification of the candidate joint configuration is deemed successful.

[0013] According to one aspect of the embodiments of this application, a redundant robotic arm teleoperation system with adaptive learning and null space regularization is proposed. The system is used to implement the redundant robotic arm teleoperation method with adaptive learning and null space regularization as described above. The system includes a master operating terminal, a communication module, a control module, a storage module, a slave operating terminal robotic arm, and a feedback module. The control module includes a behavior observation unit, an online learning unit, a null space optimization unit, a regularization control unit, and an instruction generation unit.

[0014] The beneficial effects of this application are as follows: This application calculates the first joint increment that meets the preset end-effector task requirements by measuring the end-effector error, and constructs the null-space projection matrix of the current joint configuration. Combined with the target reference joint configuration selected from the reference joint configuration set, it determines the second joint increment that meets the preset end-effector task requirements. This updates the joint configuration for the next operation round, and the process is iteratively repeated. Simultaneously, this application utilizes preference signals to dynamically update the reference joint configuration set, solving the problem of statically fixed joint configurations. Therefore, this application can form a complete closed loop of "input acquisition—end-effector solution—preference observation—set update—null-space adjustment—execution feedback," improving the control accuracy and stability of redundant robotic arms. Attached Figure Description

[0015] Figure 1 This is an architecture diagram of the redundant robotic arm teleoperation system with adaptive learning and zero-space regularization provided in the embodiments of this application; Figure 2 A flowchart illustrating the adaptive learning and null space regularization method for teleoperating a redundant robotic arm provided in an embodiment of this application; Figure 3 This is an overall logic diagram of the redundant robotic arm teleoperation method with adaptive learning and null space regularization provided in the embodiments of this application; Figure 4 This is an update logic diagram of the set of reference joint configurations proposed in this application; Figure 5 This is a flowchart of the zero-space regularization data acquisition proposed in this application; Figure 6 These are schematic diagrams of two different joint configurations proposed in this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the solutions of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] It should be noted that while some processes described in the specification, claims, and accompanying drawings include multiple steps appearing in a specific order, it should be clearly understood that these steps may not be performed in the order they appear herein, or may be performed in parallel. The step numbers are merely used to distinguish different steps and do not themselves represent any execution order. Furthermore, descriptions such as "first," "second," or "objective" in this document are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. "Multiple" in this document refers to at least two.

[0018] It is worth noting that, in the specific embodiments of this application, data related to the joint configuration, end effector pose, and joint increments of the robotic arm are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the target object is required, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. For example, when an embodiment of this application needs to obtain data related to the joint configuration, end effector pose, and joint increments of the robotic arm, separate permission or consent from the target object can be obtained through pop-up windows or redirection to a confirmation page. After obtaining separate permission or consent from the target object, the data related to the joint configuration, end effector pose, and joint increments of the robotic arm used to enable the embodiments of this application to operate normally can then be obtained.

[0019] Teleoperation is rapidly gaining demand in two types of scenarios: one is direct remote control (surgical robots, nuclear power plant maintenance, etc.); the other is robot imitation learning data collection, where learning methods such as ACT and Diffusion Policy require a large number of manual teleoperation demonstrations.

[0020] However, despite the broad application prospects of teleoperation technology, it still has certain shortcomings in terms of accuracy and stability, especially when handling complex tasks with redundant robotic arms. Operators often face problems such as joint motion oscillations and abrupt changes. These problems are particularly evident in the operation of redundant robotic arms. Furthermore, operators cannot guarantee that the manipulated robotic arm will perform the optimal pose during task completion. The limitations of existing technology become even more pronounced, especially when high-precision operations are required. Specifically, the following shortcomings exist: Redundant robotic arms, due to their greater degrees of freedom than the minimum required to complete a task, possess excellent flexibility and adaptability in complex operational tasks (such as telemedicine, nuclear power plant maintenance, and deep-sea exploration). However, traditional and current technologies lack optimized utilization of redundant degrees of freedom, leading to the following problems: When a robotic arm is performing a task, its internal posture is unstable, which may cause sudden jumps and shaking. The system does not proactively consider requirements such as joint limitation, obstacle avoidance, and tool posture maintenance, which poses a safety risk. Switching between different tasks relies on manual parameter tuning, resulting in poor portability and high deployment costs.

[0021] In summary, the current shortcomings mainly stem from the calculation of end-effector velocity increments to joint velocity increments in the teleoperation process, i.e., the inverse kinematics (IK) module. For master-slave arm mode, although explicit IK calculation is not required, the operator needs to directly control each joint of the master arm. However, novice teleoperators often struggle to effectively control the movement of each joint. Therefore, it is necessary to propose a novel teleoperation control method that fully utilizes redundant degrees of freedom to optimize and stabilize the internal posture while ensuring the accuracy of the end-effector task.

[0022] There are three main types of teleoperated devices: (1) VR headset + controller (60~90 Hz hand pose tracking); (2) master-slave robotic arm (500~1000 Hz joint status reading); (3) three-dimensional input device (such as 3D mouse, 500-1000Hz joint reading). Seven-degree-of-freedom (7-DOF) redundant robotic arms are widely used due to their superior flexibility. While only six degrees of freedom are needed to achieve three-dimensional end-effector position and attitude control, the 7-DOF robotic arm provides one additional redundant degree of freedom, enabling better end-effector pose control. The internal joint configuration can still be changed even when fixed. ,like Figure 6 As shown. Traditional teleoperation uses the Jacobi pseudo-inverse to solve the problem, only providing the minimum norm solution, completely ignoring redundant degrees of freedom, and the joint configuration... It drifts randomly with the operation history and lacks proactive optimization.

[0023] The core challenge in engineering zero-space control is "how to dynamically maintain a suitable target joint configuration in a teleoperation scenario." Teleoperation and autonomous control have two fundamental differences: Challenge 1: Teleoperation involves a "human-in-the-loop" phenomenon—different operators have significantly different expectations for the "comfortable" joint configuration, and even the same operator's habits evolve as their skills improve. Fixed targets cannot perceive or adapt to this dynamic change.

[0024] Challenge 2: When teleoperation data is used for imitation learning, the null space components of different demonstrations may be drastically different, resulting in contradictory joint-level mapping signals in the training data, which impairs the imitation learning effect. This problem has not been clearly identified and resolved in existing literature and patents.

[0025] Main disadvantages of existing technology Disadvantage 1: The target joint configuration is statically fixed and cannot adapt to individual differences among operators. In existing zero-space control telesystems, the target joint configuration is pre-set by experts and remains fixed during operation. When different operators take over the system, the preset target does not match their natural habits, requiring operators to continuously make corrective actions to counteract zero-space guidance, increasing cognitive burden. Manual recalibration is required every time a new operator or task is introduced, resulting in high deployment costs and poor portability.

[0026] Disadvantage 2: Inconsistent null space components in the demonstration data impair the effectiveness of imitation learning. The joint configurations of different demonstrations differ significantly in the null space direction, resulting in contradictory joint-level mappings in the training data. This makes it difficult to train the policy network and leads to poor joint-level stability of the learned policy.

[0027] Disadvantage 3: Discontinuous joint configuration switching, resulting in a poor user experience. When approaching unusual joint configurations or when switching targets, small end-effector inputs by the operator may cause large jumps in the joint, resulting in poor teleoperation continuity.

[0028] The purpose of this application (1) To address the first and third drawbacks: The reference joint configuration set is dynamically updated online by collecting operator preference signals in real time. To enable zero-space control of the target It automatically adapts to the current operator's habits without requiring manual intervention.

[0029] (2) Regarding the second drawback: Before writing the demonstration data into the training samples, the normalized joint configuration is pre-calculated by null projection. = + ·( This eliminates the zero-space difference between different presentations.

[0030] The following section introduces the adaptive learning and null space normalization redundant robotic arm teleoperation system of this application: Overall technical architecture (see) Figure 1 ) Based on the traditional teleoperation framework, this application adds three functional modules: a behavior observation unit (32), an online learning unit (33), and a normalization control unit (35). Main operating terminal (1): It receives operator input and supports three modes: VR headset + controller (60~90 Hz), master-slave robotic arm (500~1000 Hz), and 3D input device.

[0031] Communication module (2): It supports low-latency protocols such as real-time Ethernet (EtherCAT) and is responsible for the timing management and buffering of high-frequency bidirectional data.

[0032] Control module (3): The core computing unit contains six sub-units: ① Task space control unit (31) – calculates Δq1; ② Behavior observation unit (32) – collects operator preference signals under conditional constraints; ③ Online learning unit (33) – dynamically maintains set Q based on preference signals; ④ Null space optimization unit (34) – selects q from Q. Calculate Δq2; ⑤ Normalization control unit (35) – Calculation under data acquisition mode (Storage only); ⑥ Instruction generation unit (36) — synthesize Δq = Δq1 + Δq2 and output it.

[0033] Storage module (4): Stores a reference joint configuration set Q (dynamically maintained) and a specification configuration library (indexed by task type). ) and control parameters.

[0034] From the operating end of the robotic arm (5): A seven-DOF redundant robotic arm executes joint increment commands Δq and returns the current joint configuration q.

[0035] Feedback module (6): The system acquires and transmits the current joint configuration q, joint torque τ, and end-effector pose x at high speed.

[0036] The following is a terminology explanation for some parameter symbols appearing in this application: The symbols used in this paper are uniformly defined according to the following rules (symbols with the prefix Δ represent increments per step; q family symbols without Δ represent absolute values ​​of joint angles): [Joint Configuration (Absolute Value)] The joint angle vector provided in real time by the feedback module for the current joint configuration is the fundamental state variable for all control calculations in the teleoperation system. The reference joint configuration (or target reference joint configuration) is selected from the set of reference joint configurations Q using a nearest neighbor strategy and is used for null space correction. To standardize joint configurations—standard configurations indexed by task type in the standard configuration library, used for null space normalization; To normalize joint configurations—joint configurations that are normalized in zero space are only used for storage and are not sent to the robotic arm for execution; [Joint increment (a small amount added per control step)] The first step is the joint increment—solved by damped least squares pseudo-inverse—to meet the requirements of the end-task. The second step increment—the null space correction increment—does not change the end-effector pose under a local linear approximation. The final joint increment—Δq = Δq1 + Δq2—is actually sent to the joint driver for execution. Δq m Master arm joint increment—Joint increment input measured by the master arm in master-slave mode. N The number of times the joint velocity direction reverses within a 30-second sliding window (a preset time window) – used to quantify the smoothness of the operation trajectory; |Δq c | The amplitude of the reverse joint correction movement performed by the operator (c = correction) [Terminal Space and Matrices] The current end-effector pose is calculated from q using positive kinematics; The desired end-effector pose is determined by motion commands input by the operator. The end-effector pose error — e = x; The Jacobian matrix J = J(q) describes the linear mapping between joint velocity and end-effector velocity. The null projection matrix is ​​P = I. J†J is used to project a vector onto the null space; T is the master-slave proportional mapping matrix – used to map the master arm joint increment to the slave arm first step joint increment. α is the null space correction gain (scalar, presettable) – used for the null space correction increment calculation in the second step; The preferred set of reference joint configurations is a dynamic set of joint configurations driven by operator preference signals.

[0037] Terminology Explanation 1. Redundant Robot Arm A robotic arm with more degrees of freedom than the minimum required to complete an end-effector task. Six degrees of freedom are needed to achieve three-dimensional spatial position and attitude control; a robotic arm with seven or more joints is considered redundant. This invention focuses on a seven-degree-of-freedom redundant robotic arm.

[0038] 2. Joint Configuration The vector composed of the angles of each joint of the robotic arm is denoted as q. The joint configuration completely describes the internal state of the robotic arm; the end-effector pose x is uniquely determined by q through forward kinematics, but the same x corresponds to infinitely many different q (redundancy). Throughout this text, all references to internal states are consistently represented by q and its variants (qi). , qc, ) indicates that it is not used in conjunction with the end pose x.

[0039] 3. Joint Increment The change in joint angle within a single control cycle (single operation round). Satisfying q(t) = q(t) 1) + Δq(t). The two-step update mechanism of this invention operates at the incremental level: Δq1 is the first-step joint increment, Δq2 is the second-step increment, and Δq = Δq1 + Δq2 is the final joint increment. Symbols with the Δ prefix are increments, while q-family symbols without Δ are absolute configurations; the two are strictly distinguished.

[0040] 4. [End-Effector Pose] The spatial state of the end effector is described by six quantities, including three-dimensional position coordinates and three-dimensional attitude angles, denoted as x. The desired end effector pose is denoted as xd, and the end effector error is denoted as e = xd. x.

[0041] 5. [Jacobian Matrix] The matrix describing the linear mapping relationship between joint velocity and end-effector velocity, denoted as J = J(q), satisfies... = J· Its pseudo-inverse is denoted as J†, and it is used to solve for the minimum norm joint velocity solution that satisfies the end-effector task.

[0042] 6. Damped Least Squares (DLS) A regularized solution method is introduced to introduce a damping coefficient λ to avoid numerical divergence near singular configurations, and the first step joint increment Δq1 is output.

[0043] 7. [Null Space Projector] Defined as P = I The vector J†J, after being projected onto P, lies in the null space and does not affect the end-effector pose under the condition of local linear approximation (J·P ≈ 0).

[0044] 8. [Null Space Component] The null space component of joint configuration q is defined as P·q, representing the part of q that is independent of the end-effector pose. It forms the basis for configuration consistency analysis and the establishment of a canonical configuration library.

[0045] 9. [Preferred Joint Configuration Set] The set of reference joint configurations stored in the storage module, denoted as Q = {q 1, q 2, …, q n The system is updated online in real time, driven by operator preference signals.

[0046] 10. [Operator Preference Signal] A quantifiable kinematic signal that can be observed in real time during teleoperation, reflecting the operator's preference for the current joint configuration q. Preference scoring must be performed under identical control conditions (same task stage, end-effector position error < ε_pos, no contact state switching) to eliminate interference from non-q factors.

[0047] 11. Null Space Canonicalization Processing steps specifically designed for demonstrating the data storage chain: through computation = q + P·(qc Align joint configuration q) with normalized joint configuration qc to obtain normalized joint configuration. It is only used for storage and is not sent to the robotic arm for execution.

[0048] 12. [Standardized Canonical Joint Configuration] The standard joint configurations predefined for specific task types in the specification configuration library are denoted as qc. They can be obtained through three methods: expert teaching, null component clustering of historical demonstration data, or multi-objective optimization.

[0049] 13. Imitation Learning This invention describes a method for robots to acquire skills by learning from human demonstration data. Teleoperation is currently the mainstream method for collecting and training data through imitation learning, and the zero-space normalization mechanism of this invention is specifically designed for this scenario.

[0050] The following provides a detailed description of the specific implementation methods of the embodiments of this application: Please see Figure 2 , Figure 2 This is a flowchart illustrating the redundant robotic arm teleoperation method with adaptive learning and null space regularization provided in the embodiments of this application. Figure 2 The illustrated method for teleoperating redundant robotic arms using adaptive learning and null space regularization includes: Step 210: Read the current joint configuration of the redundant robotic arm and the end-effector pose under the current joint configuration, and determine the end-effector error based on the end-effector pose and the preset desired end-effector pose. Step 220: Calculate the first joint increment that meets the preset end-of-line task requirements based on the end-of-line error; Step 230: Select the target reference joint configuration from the dynamically updated set of reference joint configurations using the nearest neighbor selection method; Step 240: Construct the null space projection matrix of the current joint configuration, and determine the second joint increment that meets the preset end-effector task requirements based on the preset null space correction gain, the null space projection matrix, the target reference joint configuration, and the current joint configuration. Step 250: Update the joint configuration for the next operation round based on the first step joint increment, the second step joint increment, and the current joint configuration.

[0051] First, the overall process of this application can be as follows: Figure 3 The update steps for the reference joint configuration set can be as follows: Figure 4 As shown, it includes the following steps: The method of this application includes two parallel threads: a main control thread (M1~M7) and a preference observation thread (M3). M3 and M4 are the update steps for the reference joint configuration set in this application. M1: Input acquisition and terminal error calculation; Read the current joint configuration from the feedback module. and end pose Calculate the terminal error: e = x (VR / 3D input mode, x is the terminal error). e = T · Δq m (Master-slave mode, Δq) m (Main arm joint increment); A low-pass filter (cutoff frequency 30 Hz) is applied to the error signal to suppress high-frequency noise.

[0052] M2: Step 1: Calculate the joint increment Δq1 (Task Space Control Unit 31) Calculate the first-step joint increment to meet the requirements of the end-point task. (Notice: (This represents the increment of joint angle per step, in absolute terms of non-joint configurations).

VR / 3D Input Mode

[0053] The damping coefficient λ is adaptively adjusted based on the minimum singular value σmin(J): when σmin ≥ ε, λ = 0; when σmin < ε, λ = λ0·(1 σ_min / ε)², where ε is the preset comparison threshold.

[0054] [Master-Slave Mode] Δq1 = T · Δq m Only the accuracy of the final task is guaranteed, and redundant degrees of freedom are not constrained. Subsequent steps complete zero-space optimization.

[0055] M3: Real-time acquisition of operator preference signals (behavioral observation unit 32) The behavior observation units operate in parallel at a 2000 Hz oversampling frequency. The preference signal will only apply to the current joint configuration if the following three conditions are met simultaneously. Accumulated scoring: (a) Current task label not switched; (b) End-effector position error ‖e_pos‖ < ε_pos (default 5 mm); (c) End-effector contact state not switched. This conditionalization mechanism excludes non-task difficulty, contact disturbances, etc. Factors that interfere with preference ratings.

[0056] The preference signal takes the following form: [Signal Type 1] Number of times the velocity direction reverses (N) (q): Defined as the number of times the joint velocity direction reverses within a 30-second sliding window (a preset time window), used to quantify the smoothness of the operation trajectory; N The lower the (q) value, the smoother the operator's control is near the current joint configuration, which is the main preference signal.

[0057] [Signal Form Two] Correction Behavior Amplitude | Δq c|: When a reversal of the joint velocity direction is detected, the amplitude of the operator's reverse joint correction action is recorded; the larger the amplitude, the less satisfied the operator is with the current q.

[0058] M4: Reference joint configuration set Q updated online (Online learning unit 33) The online learning unit maintains a reference joint configuration set Q, according to the following rules: [Add Rule] Under the same control conditions, when the system detects that the operator is continuously operating near a certain target joint configuration q, and the number of times the joint velocity direction reverses N within that time window, the rule applies. (q) Low, no significant correction action (|Δq) c (Smaller), here the number of times the joint velocity direction is reversed is N. (q) is compared with the preset reversal number threshold and the correction behavior amplitude is compared with the preset correction amplitude. When both are less than their corresponding thresholds, q is included in the candidate. The candidate configuration (i.e. the candidate joint configuration of this application) is officially added to Q after passing the feasibility verification (the limit margin of each joint distance is >10°, the minimum singular value σ_min is >0.01) and the Euclidean distance between it and the existing members in the set Q is greater than the merging threshold θ_merge. Here, 10° is the preset limit margin and 0.01 is the preset singular threshold.

[0059] [Removal Rule] When the working area corresponding to a member (a certain reference joint configuration) in Q is continuously avoided by the operator—frequent correction actions (i.e., the number of corrections within the preset time window) or bypass behavior are observed within a continuous time T_rem—then it is removed from Q.

[0060]

Merge Rule

[0061] The total number of members in set Q does not exceed the upper limit N_max (default 20, a preset threshold). If the number exceeds this limit, the member with the lowest priority is removed.

[0062] Initially, Q is filled with several reference joint configurations taught by experts. During operation, the system gradually replaces the no longer applicable expert configurations with the operator's preferred configuration, achieving a smooth transition from "expert knowledge" to "operator's personalized preferences".

[0063] M5: Nearest neighbor selection based on joint configuration (zero-space optimization unit 34) From the set The nearest neighbor strategy is used to select the reference joint configuration. : q = argmin‖q q i ‖2 (for all q) i ∈ Q) The merge threshold θ_merge ensures that there is a certain distance between each pair of set members before and after the switch. Differences are bounded, and the zero-space correction direction remains continuous, avoiding joint jumps caused by target switching.

[0064] M6: Step 2: Calculate the increment Δq2 in step 2 Construct the current joint configuration null projection matrix at the location = I J(q)†·J(q), calculate the second step increment. : Δq² = α · P · (q q) Where α is the null space correction gain (scalar, default 0.5). Under the local linear approximation of the current control cycle, J·P ≈ 0, therefore the first-order effect of Δq2 on the end pose is approximately zero. A constraint is imposed on ‖Δq2‖_∞ (≤0.05 rad / step) to ensure that the local linearity assumption holds validly.

[0065] M7: Final Joint Increment Synthesis and Execution: Synthetic final joint increment : Δq = Δq1 + Δq2 right A joint velocity safety limit (‖Δq‖_∞ ≤ 1.5 rad / step) is applied and smoothed by a low-pass filter before being sent to the joint actuator, and the robotic arm executes the command. ← + The updated and After receiving feedback from the feedback module, the next control cycle begins (1 ms).

[0066] In some embodiments, the method further includes: Identify the task type of the redundant robotic arm in the current joint configuration; Determine the standard joint configuration based on the task type; The normalized joint configuration is calculated based on the current joint configuration and the standard joint configuration. The standardized joint configuration, the end-effector pose, the task type, and the motion commands for the redundant robotic arm are compiled into sample data and stored.

[0067] Here, a null-space regularization data acquisition method is specifically provided (see...). Figure 5 ) The standardized data acquisition process (A1~A4) and the normal control loop (M1~M7) run independently and in parallel, without introducing additional control delay. Core purpose: This process is specifically designed to demonstrate the data storage link and standardize the joint configuration. Only the training sample buffer is written; it is not sent to the robotic arm driver—the robotic arm still actually executes the output of the M7. The real-time control quality and operator experience of remote operation are completely unaffected by this process.

[0068] A1: Task type identification: Real-time determination of the current task type can be achieved using at least one of the following methods: (a) workspace location partitioning; (b) contact force pattern recognition; (c) operator active labeling (highest priority).

[0069] A2: Standard Joint Configuration Lookup: Search for standard joint configurations from the standard configuration library based on the task tags. Establishment methods (choose one of three): Method 1 – Expert demonstration; Method 2 – Data-driven clustering (extracting the null space component P(q)·q from the demonstration, then clustering, and taking the cluster centers as...) Method 3 – Multi-objective optimization solution.

[0070] A3: Zero-space normalization projection (core - only applies to storage links): Input: Current joint configuration (From the feedback module) and standard joint configuration (From A2). Calculating the regularized joint configuration. : = q + P(q) · (qc q) Where P(q) = I J(q)†·J(q) Formula interpretation: P(q)·( ) will be from the current joint configuration Pointing to standard joint configuration The direction vector is projected onto the null space, and the components that do not affect the end effector pose are extracted and superimposed onto the null space. Above. Under the local linear approximation conditions of the current control cycle, and The corresponding end pose is the same (J·P·v ≈ 0).

[0071] Apply a constraint to the normalization step size (‖P(q)·( (||_∞ ≤ 0.1 rad) to ensure the local linearity assumption is valid. Output: Regularized joint configuration It is only transmitted to the data recording module (A4), and not to the robot arm driver.

[0072] To understand the practical effect of the above calculations, a simplified numerical explanation is given below. Let the current joint configuration q and the standard configuration q_c of the seven-DOF robotic arm in a certain control cycle (a certain operation round) be respectively: q = [0.2, 0.5, 1.2, 0.8, 0.1, 0.3, 0.4] , q_c = [0.3, 0.8, 1.0, 0.6, 0.3, [0.5, 0.2] Calculate the difference vector q_c q = [0.1, 0.3, 0.2, 0.2, 0.2, 0.2, 0.2] After null space projection, we get P(q)·(q_c) q) = [0.05, 0.12, 0.08, 0.03, 0.06, 0.04, 0.01] (Only the null space component is retained after projection; the task space component is eliminated.) Superposition yields a regularized configuration. = [0.25, 0.62, 1.12, 0.77, 0.16, [0.34, 0.39] . The offset relative to q falls entirely within the null space, and the change in end pose is negligible under the condition of local linear approximation.

[0073] Imposing a constraint on the normalization step size: when ||P(q)·(q_c)|| When q)‖_∞>0.1 rad, the step size is truncated along the projection direction to satisfy the constraint, so as to ensure that the local linearity assumption is valid.

[0074] Output: Regularized joint configuration . The data is only transmitted to the data recording module in step A4, not to the robotic arm driver. The robotic arm actually executes the joint increment command Δq output by the control loop (M1 to M7).

[0075] A4: Training Sample Storage Write the following data into the training sample buffer: sample(t) = (timestamp t, normalized joint configuration) , terminal pose , Task label, Operator action instruction a(t)).

[0076] Operator action command a(t): In VR mode, this is the operator's desired end-effector displacement Δ. In master-slave mode, the master arm joint increment Δq m Traditional storage solutions (Random accumulation of null space components); This scheme stores (The null component has been aligned to) (Direction), end-effector pose data Exactly the same.

[0077] In summary, this application has the following advantages: 1. Advantage 1: The zero-space objective is upgraded from static and fixed to operator preference-adaptive, and the preference score is conditionally controlled to ensure causal validity. Existing technologies rely on fixed target joint configurations, requiring manual recalibration for new operators or tasks. This application addresses this by acquiring operator preference signals based on joint kinematics under identical control conditions. This transforms the motion signals generated during teleoperation into a learning basis for the zero-space target, achieving automated, personalized adaptation without human intervention. The conditional preference scoring mechanism effectively eliminates interference from non-joint configuration factors, ensuring the validity of the causal chain between the preference signal and the joint configuration q. This is a unique technological innovation in teleoperation and cannot be directly derived from existing autonomous control technologies.

[0078] 2. Advantage Two: For the first time, this application systematically solves the problem of null component inconsistency in teleoperation data, which is beneficial to improving imitation learning efficiency: This application achieves this through null normalization ( = q + P·(qc q), without changing the end-effector pose x, aligns the null-space components of different joint configurations to the normalized direction, eliminating data inconsistencies and facilitating the policy network's learning of stable and consistent joint-level control behavior. Normalization only applies to the storage link ( (Not executing) does not affect the real-time control quality of remote operation at all—this is what distinguishes this invention from "chaosq". The essence of existing solutions such as "performing zero-space bias control" lies in this.

[0079] 3. Advantage three: The system has strong versatility and is compatible with multiple teleoperation input modes: The two-step update framework (Δq = Δq1 + Δq2) is decoupled from the teleoperated input device. The same control logic is transparently adapted to VR mode, master-slave arm mode, and 3D input device mode. Only the input mode of step M2 needs to be configured.

[0080] It should be understood that the various implementation methods provided in this application can be combined arbitrarily to achieve different technical effects.

[0081] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0082] The above is a detailed description of the embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for teleoperating a redundant robotic arm with adaptive learning and null space regularization, characterized in that, The method includes: Read the current joint configuration of the redundant robotic arm and the end effector pose under the current joint configuration, and determine the end effector error based on the end effector pose and the preset desired end effector pose; The first joint increment that meets the preset end-task requirements is calculated based on the end-task error. In the dynamically updated set of reference joint configurations, the target reference joint configuration is selected by nearest neighbor selection. Construct the null space projection matrix of the current joint configuration, and determine the second joint increment that meets the preset end-effector task requirements based on the preset null space correction gain, the null space projection matrix, the target reference joint configuration, and the current joint configuration. The joint configuration for the next operation round is updated based on the first step joint increment, the second step joint increment, and the current joint configuration. The reference joint configuration set is dynamically updated in the following manner: The preference signal for the redundant robotic arm under the target joint configuration is collected, and the preference signal includes the number of joint velocity direction reversals and the magnitude of the correction behavior. The decision to include the target joint configuration in the candidate joint configurations is based on the number of joint velocity direction reversals and the magnitude of the correction behavior. The feasibility of the candidate joint configurations is verified, and the candidate joint configurations that pass the feasibility verification are added to the reference joint configuration set. The number of reference joint configurations in the reference joint configuration set does not exceed a preset number threshold. For any reference joint configuration in the set of reference joint configurations, if the number of corrections to the reference joint configuration within a preset time window is greater than a preset correction threshold, then the reference joint configuration is removed from the set of reference joint configurations.

2. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 1, characterized in that, The step of determining the end effector error based on the end effector pose and the preset desired end effector pose includes: If the redundant robotic arm is currently in a three-dimensional input mode, the difference between the desired end-effector pose and the end-effector pose is taken as the end-effector error. If the redundant robotic arm is currently in master-slave mode, the product of the read master joint increment of the redundant robotic arm and the preset mapping matrix is ​​used as the end effector error.

3. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 2, characterized in that, The first step of calculating the joint increment that satisfies the preset end-task requirements based on the end-task error includes: If the redundant robotic arm is currently in the three-dimensional input mode, the first step joint increment is obtained by performing a pseudo-inverse solution using damped least squares and the end-effector error. If the redundant robotic arm is currently in the master-slave mode, the first step joint increment is determined by multiplying the master arm joint increment with the preset mapping matrix.

4. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 1, characterized in that, In the dynamically updated set of reference joint configurations, the target reference joint configuration is selected using a nearest neighbor selection method, including: Among the set of reference joint configurations, the target reference joint configuration with the smallest distance from all other reference joint configurations except itself is selected.

5. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 1, characterized in that, The step of updating the joint configuration for the next operation round based on the first-step joint increment, the second-step joint increment, and the current joint configuration includes: The final joint increment for the current joint configuration is determined based on the first step joint increment and the second step joint increment. The final joint increment is superimposed on the current joint configuration to obtain the joint configuration for the next operation round.

6. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 1, characterized in that, The method further includes: Identify the task type of the redundant robotic arm in the current joint configuration; Determine the standard joint configuration based on the task type; The normalized joint configuration is calculated based on the current joint configuration and the standard joint configuration. The standardized joint configuration, the end-effector pose, the task type, and the motion commands for the redundant robotic arm are compiled into sample data and stored.

7. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 6, characterized in that, The step of calculating the normalized joint configuration based on the current joint configuration and the standard joint configuration includes: Determine the null space projection matrix of the current joint configuration; Based on the standard joint configuration, the current joint configuration, and the null projection matrix, a direction vector is determined to point the current joint configuration toward the standard joint configuration in order to extract the components of the undisturbed end-effector pose. The components of the undisturbed end pose are superimposed on the current joint configuration to obtain the normalized joint configuration.

8. The method for teleoperating a redundant robotic arm with adaptive learning and null space regularization according to claim 1, characterized in that, The feasibility verification of the candidate joint configurations includes: Determine the minimum singular value corresponding to the candidate joint configuration; Determine the joint angle vectors of each joint in the candidate joint configuration. The feasibility of the candidate joint configuration is verified by comparing each of the joint angle vectors with a preset limit margin and by comparing the minimum singular value with a preset singular threshold. If each of the joint angle vectors is less than the preset limit margin, and the minimum singular value is greater than the preset singular threshold, then the feasibility verification of the candidate joint configuration is deemed successful.

9. A redundant robotic arm teleoperation system with adaptive learning and null space regularization, characterized in that, The system is used to implement the adaptive learning and null space normalization redundant robotic arm teleoperation method as described in any one of claims 1-8. The system includes a master operating terminal, a communication module, a control module, a storage module, a slave operating terminal robotic arm, and a feedback module. The control module includes a behavior observation unit, an online learning unit, a null space optimization unit, a normalization control unit, and an instruction generation unit.

Images