Neural network based robot trajectory optimization system

Abstract

This invention relates to the field of robotic arm control technology and discloses a neural network-based robotic arm motion trajectory optimization system. The system includes an industrial computer running a non-real-time operating system, calculating and outputting static parameter data to a shared memory data area; a real-time controller running an operating system kernel with real-time patches, comprising a phase space transformation module, a feedback compensation module, and an optimization solution module; multiple robotic arm servo drives connected to the real-time controller; the phase space transformation module generating a set of static inequality constraints; the feedback compensation module generating dynamic constraint boundary data based on measured torque and current data; and the optimization solution module solving a linear programming problem and issuing joint position control commands. This invention avoids computational power preemption by separating computational and control tasks, updates phase space constraints based on actual load conditions, extracts and generates a local rolling optimization window to replace the instruction data within the circular buffer, and compensates for the total system delay.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, specifically to a robotic arm motion trajectory optimization system based on neural networks. Background Technology

[0002] Existing robotic arm trajectory optimization technologies typically process trajectory planning and underlying control tasks within a single computing hardware unit. When running neural network models, the computationally intensive feature extraction task preempts the underlying operating system's computing resources, causing time delays in the generation and issuance of joint position control commands and velocity feedforward control commands, thus compromising the time determinism of the underlying control tasks.

[0003] Existing trajectory optimization systems rely on offline-set static parameters as phase space constraints, without incorporating measured torque and current data during robot arm operation for feedback compensation. Changes in the robot arm's operating conditions during operation can alter its actual load state. Fixed static constraint boundaries cannot be dynamically updated based on the actual load state, and the output control commands are prone to exceeding the rated torque limit of the robot arm's servo drive.

[0004] Trajectory optimization systems inherently involve a total latency in data transmission and computational inference. Existing systems lack a look-ahead motion state extrapolation mechanism to account for this total latency and lack local rolling optimization and timing command reconstruction capabilities. System latency can cause timing misalignments when control commands are sent to the robotic arm's servo driver, resulting in data gaps within the underlying execution buffer and disrupting the continuity of the robotic arm's motion trajectory and the smoothness of underlying control execution. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a robotic arm motion trajectory optimization system based on neural networks, which solves the problem of resource contention and disruption of time determinism caused by the centralized processing of trajectory planning and underlying control tasks in existing robotic arm trajectory optimization systems.

[0006] To achieve the above objectives, this invention provides a robotic arm motion trajectory optimization system based on neural networks, comprising: An industrial computer runs a non-real-time operating system and has a feature generation module to receive parameters from the host computer. The industrial computer calculates and infers through the neural network model inside the feature generation module and outputs static parameter data to the shared memory data area. The real-time controller connects to an industrial computer and runs an operating system kernel with real-time patches. It also includes a phase space transformation module, a feedback compensation module, and an optimization solution module. The real-time controller reads static parameter data through a shared memory data area. Multiple robotic arm servo drives are connected to a real-time controller via a field communication bus. The phase space transformation module is used to convert static parameter data into a set of static inequality constraints, and the feedback compensation module is used to calculate and generate dynamic constraint boundary data based on the measured torque and current data and static parameter data returned by multiple robotic arm servo drives. The optimization solution module receives the set of static inequality constraints and dynamic constraint boundary data, resolves the linear programming problem with time determinism, and generates joint position control commands and velocity feedforward control commands to send to multiple robotic arm servo drives.

[0007] Furthermore, the feature generation module includes a task parameter parsing and preprocessing unit. This unit receives upper-level task parameters, including starting point spatial coordinate parameters, ending point spatial coordinate parameters, and end load quality parameters, and performs engineering normalization on these parameters to generate normalized task parameters. The neural network model uses a multilayer perceptron structure and does not contain output neurons associated with time variables. The feature generation module extracts features based on the normalized task parameters and generates static parameter data.

[0008] Furthermore, before deploying the neural network model to the industrial computer, the feature generation module trains the neural network model offline in an offline computing environment. The feature generation module runs a pre-built robotic arm dynamics simulation model to perform iterative optimization calculations on multiple sets of different normalized task parameter samples, generating non-uniform rational B-spline curve parameters and prior pseudo-energy threshold sequences corresponding to the normalized task parameter samples. The feature generation module sets the normalized task parameter samples as the training input data of the neural network model, and merges the non-uniform rational B-spline curve parameters and prior pseudo-energy threshold sequences as the training label data of the neural network model. The feature generation module stops training and saves the neural network model when the mean square error value drops below a pre-set error threshold. The error threshold is set based on the steady-state background noise amplitude of the robotic arm joint position sensor.

[0009] Furthermore, the static parameter data includes non-uniform rational B-spline curve parameters and a priori pseudo-energy threshold sequence; the feature generation module constructs a continuous spatial pose equation based on the non-uniform rational B-spline curve parameters, and the feature generation module packages and integrates the continuous spatial pose equation and the priori pseudo-energy threshold sequence and writes them into the shared memory data area.

[0010] Furthermore, the phase space transformation module reads the continuous spatial pose equation and the prior pseudo-energy threshold sequence from the shared memory data area; the phase space transformation module calculates the first and second derivatives of the continuous spatial pose equation with respect to the time variable according to the chain rule, generating joint velocity conversion formulas and joint acceleration conversion formulas respectively; the phase space transformation module converts the joint velocity column vector and the joint acceleration column vector into mathematical expressions consisting only of path scalars, the first time derivative of the path scalars, and the second time derivative of the path scalars, constructing a phase space coordinate system; the phase space transformation module performs numerical intersection comparison between the pre-set rated maximum torque value and rated minimum torque value of the robotic arm servo driver and the prior pseudo-energy threshold sequence at the same path scalar nodes, generating a comprehensive boundary threshold sequence.

[0011] Furthermore, the phase space transformation module extracts the pre-stored multibody dynamics parameters of the robotic arm; the phase space transformation module combines the first-order geometric partial derivative vector of the continuous spatial pose equation with respect to the path scalar, the second-order geometric partial derivative vector of the continuous spatial pose equation with respect to the path scalar, and the multibody dynamics parameters of the robotic arm to perform matrix multiplication to generate a phase space state coefficient matrix; the phase space transformation module constructs a generalized phase space linear inequality constraint set by combining the phase space state coefficient matrix with the comprehensive boundary threshold sequence, and defines the generalized phase space linear inequality constraint set as a static inequality constraint set; the optimization solution module generates a continuous static reference safety envelope boundary line in the phase space state plane based on the static inequality constraint set, and defines the enclosed two-dimensional graphical region as the initial feasible region.

[0012] Furthermore, the feedback compensation module includes a resampling unit. This unit uses a pre-defined fixed time window to perform a moving average filter on the measured torque and current data, outputting smoothed measured torque and current data. The feedback compensation module extracts measured joint position data and performs numerical iterative approximation calculations using the continuous space pose equation to obtain the current path scalar value. The feedback compensation module aligns the smoothed measured torque and current data with the current path scalar value at the same time point to generate a phase space measured torque sequence. The feedback compensation module quantifies the total system delay time and performs motion state extrapolation calculations using the current path scalar value, the first-order time derivative value, and the second-order time derivative value to generate a look-ahead path scalar value. The feedback compensation module defines the look-ahead path scalar value as the look-ahead injection point parameter and sends it to the optimization solution module.

[0013] Furthermore, the feedback compensation module is equipped with a real-time evaluation unit for actual energy dissipation rate. The real-time evaluation unit for actual energy dissipation rate extracts values ​​from the measured torque sequence in phase space based on the current path scalar value to obtain the measured torque column vector corresponding to the current spatial pose node. The real-time evaluation unit for actual energy dissipation rate calculates the first-order partial derivative of the continuous spatial pose equation with respect to the path scalar value based on the current path scalar value to generate a geometric Jacobian column vector. The real-time evaluation unit for actual energy dissipation rate calculates the inner product of the measured torque column vector and the geometric Jacobian column vector to generate the actual energy dissipation rate value, and sets the actual energy dissipation rate value as the actual load state parameter characterizing the actual load state of the robotic arm body.

[0014] Furthermore, the feedback compensation module includes a boundary dynamic deformation calculation unit. This unit extracts data from the prior pseudo-energy threshold sequence based on the current path scalar value to obtain the node's prior pseudo-energy value. It then divides the actual energy dissipation rate value with the node's prior pseudo-energy value and a preset small positive number, calculating and comparing the result to generate a deformation factor value. Within the delay blind zone between the current path scalar value and the look-ahead path scalar value, the unit constructs a spline smoothing function. Finally, it combines the comprehensive boundary threshold sequence, deformation factor value, and spline smoothing function to perform smooth deformation multiplication and addition calculations, generating dynamic constraint boundary data.

[0015] Furthermore, the optimization solution module internally incorporates a timing instruction reconstruction and low-level execution overlay unit. The optimization solution module receives look-ahead injection point parameters, sets these parameters as the starting node for the current calculation cycle, and generates a local rolling optimization window. Within this window, the optimization solution module solves for the first-order time derivative of the path scalar using a time-deterministic linear programming approach, generating an optimal phase space state node sequence. The timing instruction reconstruction and low-level execution overlay unit divides the optimal phase space state node sequence into discrete time nodes, calculates and generates new joint position control commands and velocity feedforward control commands, and encapsulates the data bits to generate a data packet sequence. Based on the timestamp values ​​of the look-ahead injection point parameters, the timing instruction reconstruction and low-level execution overlay unit locates the target and writes it to memory address in the circular buffers within multiple robotic arm servo drives, and controls the data packet sequence to replace the existing old instruction data in the circular buffers in ascending time order, starting from the target's memory address.

[0016] This invention provides a robotic arm motion trajectory optimization system based on neural networks. It has the following beneficial effects: 1. This invention generates static parameter data by running a non-real-time operating system inside an industrial computer to perform neural network model inference, while simultaneously running an operating system kernel with real-time patches inside a real-time controller to read the static parameter data and perform optimization solutions. This separates the high-computing-power-consuming feature extraction task from the underlying control task, avoids time delays caused by computing power preemption, and ensures the time determinism of the generation and issuance of joint position control commands and speed feedforward control commands. 2. This invention calculates the actual energy dissipation rate based on the measured torque and current data and the current path scalar value through the feedback compensation module, generates the deformation factor value by combining the node prior pseudo-energy value, and then calculates the dynamic constraint boundary data by combining the comprehensive boundary threshold sequence. This data is provided to the optimization solution module to resolve the linear programming, thereby realizing the real-time update of the phase space constraint conditions according to the actual load state of the robotic arm body and preventing the output command from exceeding the rated torque limit of the robotic arm servo driver. 3. This invention receives the look-ahead injection point parameters through an optimization solution module and extracts them to generate a local scrolling optimization window. Within the local scrolling optimization window, a linear programming problem is solved to generate an optimal phase space state node sequence. A data packet sequence is generated through timing instruction reconstruction and the underlying execution overlay unit. The original old version instruction data inside the circular buffer is replaced in ascending order of time to compensate for the total system delay time and ensure the continuity of the robotic arm's motion trajectory and the smoothness of the underlying control execution. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a comparison diagram of the dynamic deformation of the constraint boundary in this invention; Figure 3 This is a comparison chart of trajectory tracking errors in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See attached document Figure 1 The present invention provides a robotic arm motion trajectory optimization system based on neural networks, including an industrial computer, a real-time controller, a field communication bus, and multiple robotic arm servo drives; the industrial computer is connected to the real-time controller, and the real-time controller is connected to multiple robotic arm servo drives through the field communication bus.

[0020] The industrial computer runs a non-real-time operating system and has an external communication interface to receive parameters from the host computer. The non-real-time operating system calculates neural network inference tasks based on these parameters and outputs static parameter data. The real-time controller runs an operating system kernel with real-time patches and is equipped with a hardware clock to provide a hardware clock reference signal. It receives static parameter data, re-solves time-deterministic linear programming problems, and performs instruction interpolation. The industrial computer and the real-time controller share a common memory data area, through which they exchange data across operating systems.

[0021] The field communication bus adopts the EtherCAT bus standard. The real-time controller is configured as the master device of the EtherCAT bus network, and multiple robotic arm servo drives are configured as slave devices of the EtherCAT bus network. A fixed synchronization communication cycle is set between the real-time controller and the multiple robotic arm servo drives, with the value of the synchronization communication cycle set to 1ms. Within each synchronization communication cycle, the real-time controller sends joint position control commands and speed feedforward control commands to the multiple robotic arm servo drives via the EtherCAT bus. At the same time, the multiple robotic arm servo drives return measured joint position data and measured torque and current data to the real-time controller via the EtherCAT bus. Each of the multiple robotic arm servo drives is equipped with a circular buffer to store the received control commands.

[0022] Industrial computers, operating systems that are not real-time, have a feature generation module. Real-time controllers, with their real-time patching operating system kernel, have a phase space transformation module, a feedback compensation module, and an optimization solution module.

[0023] The feature generation module is used to receive parameters from the host computer, calculate neural network inference tasks and generate static parameter data, and write the static parameter data into the shared memory data area allocated between the industrial computer and the real-time controller.

[0024] The phase space transformation module is used to read static parameter data from the shared memory data area, perform coordinate space transformation operations, and convert the static parameter data into a set of static inequality constraints in the phase space coordinate system.

[0025] The feedback compensation module is used to receive measured torque and current data returned by multiple robotic arm servo drives through the field communication bus, perform look-ahead calculations, and generate dynamic constraint boundary data based on the measured torque and current data and static parameter data.

[0026] The optimization solution module receives the set of static inequalities output by the phase space transformation module and the dynamic constraint boundary data output by the feedback compensation module. Based on the dynamic constraint boundary data, it solves a time-deterministic linear programming problem, generates joint position control commands and speed feedforward control commands, and sends them to multiple robotic arm servo drives via the field communication bus.

[0027] The feature generation module inside the industrial computer is equipped with a task parameter parsing and preprocessing unit. The task parameter parsing and preprocessing unit receives upper-level task parameters from the external communication interface. The upper-level task parameters include the starting spatial coordinate parameters, the ending spatial coordinate parameters, and the end load mass parameters. After receiving the upper-level task parameters, the unit performs engineering normalization processing on the upper-level task parameters, reads the pre-set maximum working space boundary value of the robotic arm, and linearly scales the starting spatial coordinate parameters and the ending spatial coordinate parameters according to the maximum working space boundary value of the robotic arm, and maps them to a fixed dimensionless numerical range.

[0028] In a preferred implementation, to prevent system crashes due to division by zero, a minimal positive real constant is pre-set within the task parameter parsing and preprocessing unit. The task parameter parsing and preprocessing unit synchronously reads the pre-set maximum rated load value of the robotic arm and adds this value to the minimal positive real constant to form a corrected denominator for the division operation. Based on this corrected denominator, the end-load quality parameter is divided, mapping it to a value range between 0 and 1, completing the data conversion to a uniform scale, thereby stabilizing the gradient propagation process within the neural network. After mapping, normalized task parameters are generated and used as input data for the feature generation module to calculate the neural network inference task.

[0029] The feature generation module incorporates a neural network model, employing a multilayer perceptron (MLP) architecture. The MLP consists of an input layer, multiple hidden layers, and an output layer. The input layer connects to the task parameter parsing and preprocessing unit, receiving the normalized task parameters output by this unit. Each hidden layer contains fully connected computation nodes and non-linear activation functions to receive the normalized task parameters and extract features. The output layer is configured as a time-independent node, containing no output neurons associated with time variables, and generates static parameter data based on the outputs of the hidden layers.

[0030] Before deploying the neural network model to an industrial computer, the feature generation module trains the model offline in an offline computing environment. During offline training, the feature generation module runs a pre-built robotic arm dynamics simulation model, inputting multiple sets of different normalized task parameter samples. The robotic arm dynamics simulation model performs iterative optimization calculations on these different sets of normalized task parameter samples, generating non-uniform rational B-spline curve parameters and a priori pseudo-energy threshold sequences corresponding to the normalized task parameter samples.

[0031] The feature generation module sets normalized task parameter samples as training input data for the neural network model, and merges non-uniform rational B-spline curve parameters and prior pseudo-energy threshold sequences as training label data for the neural network model. In this embodiment, the feature generation module internally configures a mean squared error function as a model convergence evaluation criterion. The feature generation module calculates the mean squared error value between the output result and the training label data, and updates the internal network weight values ​​of the neural network model through the backpropagation algorithm. The error threshold is set based on the steady-state background noise amplitude of the robotic arm joint position sensor. The feature generation module stops training when the mean squared error value drops below the set error threshold and saves the trained neural network model.

[0032] The neural network model generates static parameter data through the output layer. The static parameter data includes non-uniform rational B-spline curve parameters and a prior pseudo-energy threshold sequence.

[0033] The feature generation module extracts parameters from the non-uniform rational B-spline curve. These parameters include a control point sequence and its corresponding weight set. Based on these parameters, the feature generation module constructs a continuous spatial pose equation. The calculation formula for the continuous spatial pose equation is as follows: ; In the formula, Joint angle column vector; For path scalars; An integer representing the total number of control points minus one; The summation index is an integer; For a given degree of B-spline basis function; For the first The weight values ​​of each control point; For the first Control point vectors. Path scalar. The numerical range is limited to between 0 and 1, path scalar As independent variables without a time dimension, the pose equation in continuous space does not contain time variables.

[0034] The feature generation module simultaneously extracts a priori pseudo-energy threshold sequence and aligns this sequence with the path scalar on the numerical node sequence. The priori pseudo-energy threshold sequence represents the maximum allowable dynamic state variation limit of the robotic arm under multiple spatial pose nodes defined by the continuous spatial pose equation. The feature generation module packages and integrates the continuous spatial pose equation and the priori pseudo-energy threshold sequence and writes them into the shared memory data area.

[0035] The phase space transformation module reads the continuous spatial pose equation and the prior pseudo-energy threshold sequence from the shared memory data area, performs a coordinate space transformation on the continuous spatial pose equation, introduces a time variable, and calculates the first derivative of the equation with respect to the time variable using the chain rule to generate the joint velocity transformation formula. The joint velocity transformation formula is as follows: ; In the formula, This is the column vector of joint velocities; is the vector of first-order geometric partial derivatives of the pose equation in continuous space with respect to the path scalar; For path scalars; Let be the first time derivative of the path scalar.

[0036] After the first derivative calculation is completed, the phase space transformation module continues to calculate the second derivative of the joint velocity conversion formula with respect to the time variable, generating the joint acceleration conversion formula. The joint acceleration conversion formula is as follows: ; In the formula, This is the column vector of joint accelerations; is the vector of second-order geometric partial derivatives of the pose equation in continuous space with respect to the path scalar; For path scalars; Let be the square of the first time derivative of the path scalar; is the vector of first-order geometric partial derivatives of the pose equation in continuous space with respect to the path scalar; It is the second time derivative of the path scalar.

[0037] The phase space transformation module performs numerical transformation through joint velocity transformation formulas and joint acceleration transformation formulas, converting the joint velocity column vector and joint acceleration column vector into mathematical expressions consisting only of path scalars, the first-order time derivatives of the path scalars, and the second-order time derivatives of the path scalars. The path scalars, the first-order time derivatives of the path scalars, and the second-order time derivatives of the path scalars are set as independent coordinate axis variables to construct a phase space coordinate system, completing the mapping of kinematic state data to the phase space coordinate system and eliminating the direct coupling relationship between kinematic state data and time variables.

[0038] After mapping is completed, the phase space transformation module transforms the constraints into generalized inequalities, reads the pre-set rated maximum torque value and rated minimum torque value of the robotic arm servo driver, and synchronously reads the prior pseudo-energy threshold sequence from the shared memory data area. The rated maximum torque value, rated minimum torque value and prior pseudo-energy threshold sequence are compared with the numerical intersection at the same path scalar node to generate a comprehensive boundary threshold sequence.

[0039] The phase space transformation module extracts the pre-stored multibody dynamics parameters of the robotic arm, which include inertia matrix parameters, Coriolis force matrix parameters, and gravity load vector parameters. The module then performs matrix multiplication on the first-order geometric partial derivative vectors of the continuous spatial pose equations with respect to the path scalar, the second-order geometric partial derivative vectors of the continuous spatial pose equations with respect to the path scalar, and the multibody dynamics parameters of the robotic arm. This generates a first state coefficient matrix associated with the path scalar and a second state coefficient matrix associated with the path scalar. The first and second state coefficient matrices are then merged and set as the phase space state coefficient matrix.

[0040] The phase space transformation module constructs a set of generalized phase space linear inequality constraints from the phase space state coefficient matrix and the integrated boundary threshold sequence. The formula for calculating the set of generalized phase space linear inequality constraints is as follows: ; In the formula, This is the first state coefficient matrix associated with the path scalar; The second time derivative of the path scalar; This is the second state coefficient matrix associated with the path scalar; Let be the square of the first time derivative of the path scalar; This is a comprehensive boundary threshold sequence.

[0041] After the generalized phase space linear inequality constraint set is constructed, the phase space transformation module defines this set as a static inequality constraint set and outputs it to the optimization solution module. Upon receiving this set, the optimization solution module sets the path scalar as the horizontal axis variable and the first-order time derivative of the path scalar as the vertical axis variable, constructing a phase space state plane in the two-dimensional numerical plane.

[0042] The optimization solution module sets multiple discrete path scalar sampling nodes along the horizontal axis in the phase space state plane. For each path scalar sampling node, the known values ​​corresponding to the path scalar sampling node are substituted into the static inequality constraint set for algebraic operations. Based on the operation results, the admissible solution set region formed by the intersection of the first-order time derivative and the second-order time derivative of the path scalar is calculated. The maximum positive real value that the first-order time derivative of the path scalar can obtain is extracted within the admissible solution set region.

[0043] The optimization solution module connects the maximum positive real values ​​corresponding to each path scalar sampling node in ascending order of the horizontal axis variable to generate line segments, thereby generating a continuous static reference safety envelope boundary line in the phase space state plane. The closed two-dimensional graphic region enclosed by the horizontal axis, vertical axis and static reference safety envelope boundary line is defined as the initial feasible region.

[0044] The initial feasible region contains all phase space state nodes that simultaneously satisfy the set of static inequality constraints. The optimization solution module sets the initial feasible region as the initial search boundary condition for the linear programming solution, strictly limiting the optimization space of the linear programming solution to within the initial feasible region. When solving this linear programming problem, if the internally generated test variable node is located outside the initial feasible region, the optimization solution module directly discards the test variable node. The initial feasible region ensures that the generated joint position control commands and velocity feedforward control commands are strictly within the safety boundaries set by the prior pseudo-energy threshold sequence and the limit value of the servo driver during servo execution.

[0045] The optimization and solution module sends joint position control commands and velocity feedforward control commands to multiple robotic arm servo drives. These servo drives receive the control commands and drive the robotic arm body to move. During the robotic arm's movement, the feedback compensation module synchronously collects measured torque and current data and performs inverse mapping. Multiple robotic arm servo drives record the operating status of their internal servo motors in real time and return the measured torque and current data and measured joint position data to the feedback compensation module inside the real-time controller via the field communication bus. The synchronous communication period of the field communication bus is fixed at 1ms.

[0046] The feedback compensation module is equipped with a resampling unit. The resampling unit receives measured torque and current data and measured joint position data at the same time node according to the synchronous communication cycle. It uses a pre-set fixed time window to perform moving average filtering on the measured torque and current data to eliminate electromagnetic noise interference from the sensor and outputs smoothed measured torque and current data.

[0047] The feedback compensation module extracts the smoothed measured torque and current data and measured joint position data, reads the continuous spatial pose equation from the shared memory data area, sets the measured joint position data as the target equation parameters, and performs numerical iterative approximation based on this continuous spatial pose equation to calculate the current path scalar value. The current path scalar value and the measured joint position data correspond completely at the spatial pose level and are independent variables without a time dimension.

[0048] The feedback compensation module aligns the smoothed measured torque and current data with the current path scalar values ​​at the time nodes, eliminating the original time dimension attribute of the measured torque and current data. It then maps the measured torque and current data into the phase space coordinate system constructed by the phase space transformation module, completing the transformation from time-dimensional data to the phase space coordinate system and generating a phase space measured torque sequence. The numerical nodes within the phase space measured torque sequence are strictly aligned with the path scalar nodes defined by the continuous spatial pose equations and stored in the shared memory data area to characterize the actual load state of the robotic arm at specific path scalar nodes.

[0049] After generating the phase space measured torque sequence, the feedback compensation module quantifies the total system delay and predicts the look-ahead injection point, extracts the internally recorded computation time value (the computation time value represents the time consumed by the optimization solution module to perform linear programming solution calculation), reads the synchronous communication cycle of the field communication bus, sets the time span corresponding to the instruction issuance transmission process and the data reading transmission process as one synchronous communication cycle, and adds the two synchronous communication cycles to the computation time value to calculate the total system delay time.

[0050] The feedback compensation module extracts the current path scalar value, reads the first and second time derivative values ​​synchronized with the current path scalar value from the shared memory data area, and performs motion state extrapolation by combining the total system delay time, the current path scalar value, and the first and second time derivative values ​​to calculate and generate the look-ahead path scalar value. The formula for calculating the look-ahead path scalar value is: ; In the formula, Forward path scalar value; This is the scalar value for the current path; This is the value of the first-order time derivative; This represents the total system latency. This is the value of the second-order time derivative.

[0051] After the scalar numerical calculation of the look-ahead path is completed, the feedback compensation module defines it as a look-ahead injection point parameter. Based on the hardware clock reference signal of the real-time controller, a timestamp is added to this look-ahead injection point parameter before it is sent to the optimization solution module. The optimization solution module receives this look-ahead injection point parameter, sets it as the starting variable node for the next linear programming solution operation, and introduces the dynamic constraint boundary data output by the feedback compensation module at this starting variable node. Starting from this starting variable node, it re-solves the time-deterministic linear programming problem and performs instruction interpolation, generating new joint position control commands and velocity feedforward control commands covering the subsequent time interval of the starting variable node. By introducing the look-ahead injection point parameter, the instruction lag caused by the total system delay time is offset, and the new joint position control commands and velocity feedforward control commands are sent to multiple robotic arm servo drives.

[0052] The feedback compensation module includes a real-time evaluation unit for actual energy dissipation rate. This unit receives the phase space measured torque sequence and the current path scalar value generated by the feedback compensation module, extracts values ​​from the phase space measured torque sequence based on the current path scalar value, and obtains the measured torque column vector corresponding to the current spatial pose node.

[0053] The real-time actual energy dissipation rate assessment unit reads the continuous spatial pose equation from the shared memory data area, calculates the first-order partial derivative of the continuous spatial pose equation with respect to the path scalar based on the current path scalar value, generates the geometric Jacobian column vector corresponding to the current spatial pose node, and calculates the inner product of the measured torque and the geometric Jacobian column vector to generate the actual energy dissipation rate value. The formula for calculating the actual energy dissipation rate value is as follows: ; In the formula, This represents the actual energy dissipation rate. This is the column vector of measured torques; This is the matrix transpose operator; It is a geometric Jacobian column vector.

[0054] After the actual energy dissipation rate is calculated, the real-time evaluation unit sets the actual energy dissipation rate value as the actual load state parameter of the robotic arm. The actual load state parameter is used to characterize the actual power consumption and load level of the robotic arm at the current spatial pose node. The feedback compensation module sets the actual load state parameter as the input variable for updating the dynamic constraint boundary data. The feedback compensation module has a boundary dynamic deformation calculation unit, and the real-time evaluation unit directly outputs the actual load state parameter, i.e., the actual energy dissipation rate value, to the boundary dynamic deformation calculation unit.

[0055] The boundary dynamic deformation calculation unit reads the prior pseudo-energy threshold sequence from the shared memory data area, extracts data from the prior pseudo-energy threshold sequence based on the current path scalar value, and obtains the node prior pseudo-energy value corresponding to the current spatial pose node. To avoid the risk of system division by zero, the boundary dynamic deformation calculation unit adds the node prior pseudo-energy value to a preset small positive number, and compares the actual energy dissipation rate value with the added node prior pseudo-energy value by division to calculate the deformation factor value. The formula for calculating the deformation factor value is: ; In the formula, This represents the deformation factor value; This represents the actual energy dissipation rate. The prior pseudo-energy value of the node; This is a preset small positive number used to prevent the denominator from being zero.

[0056] After calculating and generating the deformation factor values, the boundary dynamic deformation calculation unit extracts the current path scalar value and the look-ahead path scalar value, defining the numerical interval between them as the delay blind zone. Within the delay blind zone, a spline smoothing function is constructed. The comprehensive boundary threshold sequence is read from the shared memory data area, and the comprehensive boundary threshold sequence, deformation factor values, and spline smoothing function are combined to perform smooth deformation multiplication and addition calculations, generating dynamic constraint boundary data. The calculation formula for the dynamic constraint boundary data is as follows: ; In the formula, For dynamically constrained boundary data; This is a comprehensive boundary threshold sequence; This represents the deformation factor value; It is a spline smoothing function; The scalar independent variable is the blind path within the delay blind zone.

[0057] The boundary dynamic deformation calculation unit controls the attenuation weight distribution of deformation factor values ​​within the delay blind zone through a spline smoothing function. The spline smoothing function outputs a value of 1 at the current path scalar value and a value of 0 at the look-ahead path scalar value. This unit ensures that the dynamic constraint boundary data perfectly matches the actual load state at the current path scalar value, while simultaneously ensuring a smooth transition from the dynamic constraint boundary data at the look-ahead path scalar value to the original baseline value of the comprehensive boundary threshold sequence. This smooth transition eliminates numerical step phenomena, preventing servo acceleration command oscillations caused by numerical step jumps, and outputs the generated dynamic constraint boundary data to the optimization solution module. The optimization solution module receives the dynamic constraint boundary data and solves a time-deterministic linear programming problem for the subsequent intervals of the look-ahead path scalar value.

[0058] The optimization solution module receives the dynamic constraint boundary data and look-in injection point parameters output by the feedback compensation module, sets the look-in injection point parameters as the solution start node of the current calculation cycle, and extracts a fixed-length path scalar interval along the horizontal axis in the phase space state plane starting from the solution start node. The extracted fixed-length path scalar interval is defined as the local rolling optimization window.

[0059] The optimization solution module establishes a time minimization objective function within the local scrolling optimization window. The formula for calculating the time minimization objective function is as follows: ; In the formula, The total time consumption for minimizing the objective function in time is calculated. The scalar value of the look-ahead path corresponding to the look-ahead injection point parameters; Optimize the scalar value of the solution termination node set inside the local scrolling window; The first-order time derivative of the path scalar; Let be the differential independent variable of the path scalar.

[0060] The optimization solution module extracts the phase space state coefficient matrix that matches the position of the local rolling optimization window, performs algebraic combination operations on the dynamic constraint boundary data and the phase space state coefficient matrix, generates an updated set of local inequality constraints within the local rolling optimization window, extracts the pre-constructed initial feasible region, and uses the set of local inequality constraints to perform external boundary truncation and correction operations on the external boundary of the initial feasible region to generate the local feasible region boundary for the current calculation cycle.

[0061] The optimization solution module sets the local feasible region boundary as the optimization constraint and minimizes the total computation time as the monotonic optimization direction. Before execution, the optimization solution module monitors the geometric area parameter of the local feasible region boundary. If the geometric area parameter is less than the minimum tolerance, the system determines that the current constraint has conflicted and formed an empty set. The optimization solution module then terminates the current calculation cycle and retrieves the instructions from the previous cycle stored in the data cache queue to maintain the robot arm's deceleration and braking state. If the geometric area parameter meets the solution requirements, the optimization solution module solves a time-deterministic linear programming problem on the first-order time derivative of the path scalar within the local feasible region boundary. By solving the linear programming problem, it generates the optimal phase space state node sequence located within the local rolling optimization window.

[0062] The optimization solution module extracts the optimal phase space state node sequence, performs instruction interpolation on the optimal phase space state node sequence, converts the time-dimensional phase space data into a discrete time domain sequence, generates new joint position control instructions and velocity feedforward control instructions covering the local scrolling optimization window range, writes the new joint position control instructions and velocity feedforward control instructions into the internally set data buffer queue, and sends the new joint position control instructions and velocity feedforward control instructions in the data buffer queue to multiple robotic arm servo drives when the next synchronization communication cycle arrives.

[0063] The optimization solution module internally includes a timing instruction reconstruction and low-level execution coverage unit. After extracting the optimal phase space state node sequence, the timing instruction reconstruction and low-level execution coverage unit reads the optimal phase space state node sequence from the data cache queue and the continuous space pose equation from the shared memory data area. It then divides the optimal sequence into discrete time nodes, extracts the discrete path scalar values ​​and the first-order time derivative values ​​of the path scalars corresponding to each discrete time node, and substitutes these discrete path scalar values ​​into the continuous space pose equation to calculate and generate new joint position control commands. Simultaneously, the timing instruction reconstruction and low-level execution coverage unit performs vector multiplication of the discrete path scalar values, the first-order time derivative values ​​of the path scalars, and the continuous space pose equation to generate velocity feedforward control commands. The calculation formula for the velocity feedforward control commands is as follows: ; In the formula, For the first A column vector of velocity feedforward control commands corresponding to each discrete time node; For discrete time node variables; The time node index is an integer; The vector of first-order geometric partial derivatives of the continuous spatial pose equation with respect to the path scalar, corresponding to the discrete path scalar numerical values. For discrete path scalar values; This is the numerical value of the first-order time derivative of the path scalar.

[0064] The timing instruction reconfiguration and underlying execution overlay unit encapsulates the new joint position control instructions and speed feedforward control instructions into data bits according to the protocol format of the field communication bus, generating a data packet sequence conforming to the underlying communication standard. It then acquires the hardware clock reference signal from the real-time controller, reads the timestamp value of the look-ahead injection point parameter based on the hardware clock reference signal, and locates the target write memory address in the circular buffer within multiple robotic arm servo drives based on the timestamp value of the look-ahead injection point parameter. The target write memory address is aligned with the effective time node indicated by the look-ahead injection point parameter.

[0065] The timing instruction refactoring and low-level execution overlay unit writes the data packet sequence to the target memory address, controlling the data packet sequence to replace the original old instruction data inside the circular buffer in ascending time order, starting from the target memory address. The timing instruction refactoring and low-level execution overlay unit resolves instruction overlap conflicts during dynamic trajectory updates by locating the target memory address and using continuous data overlay, achieving stepless data replacement of low-level driving timing instructions based on local rolling optimization.

[0066] The real-time controller continuously acquires data, performs calculations, and writes instructions to maintain the closed-loop dynamic refresh action of the robotic arm body during its spatial movement, ensuring that the annular buffers inside the multiple robotic arm servo drives always contain execution instructions that meet the safety boundary conditions.

[0067] Specific application examples: Based on the specific high-speed, high-load handling operation nodes of a six-axis industrial robotic arm, the system configuration constants and initial parameters are as follows: The synchronous communication cycle of the EtherCAT field communication bus is set to 1ms; the single calculation time of the optimization solution module for performing linear programming calculations is set to 1ms; the preset small positive number for preventing division by zero inside the boundary dynamic deformation calculation unit is set to 0.01; and the load mass parameter of the robotic arm end effector is set to 5kg.

[0068] During the continuous spatial transport trajectory executed by the robotic arm, the feedback compensation module quantifies the total system delay. The module sets the time span for both the instruction transmission and data reading transmission processes to 1ms, and adds these two transmission times to the 1ms computation time of the optimization and solution module, calculating the total system delay to be 3ms, or 0.003s. At a certain moment, the feedback compensation module performs numerical iterative approximation using measured joint position data and the continuous spatial pose equation, extracting the current path scalar value as 0.5. Simultaneously, the module extracts the first-order time derivative as 2.0 and the second-order time derivative as -1.0. The module then substitutes these values ​​into the formula to calculate the look-ahead path scalar value. ; The feedback compensation module uses the calculated 0.5059955 as the look-ahead injection point parameter, adds a timestamp, and sends it to the optimization solution module. The optimization solution module then re-solves the linear programming problem from this node, completely eliminating the underlying execution lag caused by the 3ms communication and total computation delay.

[0069] When the robotic arm grasps the load and passes through the spatial region near the singular point, the actual load state of the body surges. The real-time evaluation unit for actual energy dissipation rate extracts the measured torque column vector based on the current path scalar value and calculates the geometric Jacobian column vector. After performing an inner product operation on the measured torque column vector and the geometric Jacobian column vector, the actual energy dissipation rate is calculated to be 55.0 J / s. The boundary dynamic deformation calculation unit extracts the node prior pseudo-energy value corresponding to the current spatial pose node from the prior pseudo-energy threshold sequence, which is 50.0 J / s. The boundary dynamic deformation calculation unit substitutes the values ​​into the formula to calculate the deformation factor value: ; Because the actual energy dissipation rate exceeded the prior model's estimate, the calculated deformation factor value was greater than 1. The boundary dynamic deformation calculation unit constructed a spline smoothing function within a delay blind zone of 0.5 to 0.5059955, incorporating the deformation factor value greater than 1 into the comprehensive boundary threshold sequence for smooth deformation multiplication and addition calculations, generating inwardly contracting dynamic constraint boundary data. After receiving this dynamic constraint boundary data, the optimization solution module reduced the amplitude of the velocity feedforward control command in subsequent time intervals during linear programming, actively controlling the robotic arm to slow down and avoiding overload alarms from the underlying servo driver.

[0070] Conduct actual operation tests and compare data; refer to the appendix for specific verification results. Figure 2 and attached Figure 3 The experiment compared continuous operation data of a traditional static trajectory planning system and the system of this invention under the same path and the same 5kg load conditions. The experimental data is shown in Table 1.

[0071] Table 1. Comparison of Operating Data between Traditional Static Trajectory Planning System and the System of the Invention Traditional static trajectory planning system 4.85 12.4 1.85 15.6 This invention system 4.62 2.1 0.32 1.2 See attached document Figure 2 And Table 1, Appendix Figure 2 This figure compares the dynamic deformation of the constraint boundary in the phase space coordinate system. The horizontal axis represents the path scalar, which is dimensionless and ranges from 0.4 to 0.6. The vertical axis represents the constraint boundary limit, in N·m. The figure includes the comprehensive boundary threshold curve and the dynamic constraint boundary curve generated by this invention.

[0072] When the path scalar is between 0.4 and 0.48, the robot arm's load is stable, the deformation factor is close to 1, and the dynamic constraint boundary curve completely coincides with the comprehensive boundary threshold curve. When the path scalar reaches around 0.5, the actual energy dissipation rate suddenly increases due to changes in the robot arm's pose. The feedback compensation module of this invention calculates a deformation factor greater than 1 in real time. The comprehensive boundary threshold curve, as a static parameter, remains unchanged, while the dynamic constraint boundary curve generated by this invention, under the constraint of the spline smoothing function, exhibits a smooth inward concave contraction trend starting from the 0.5 node, dynamically tightening the optimization space of the optimization solution module. This mechanism forces the system to slow down in advance, directly suppressing the instantaneous overshoot surge of the servo current.

[0073] See attached document Figure 3 And Table 1, Appendix Figure 3 This diagram compares the trajectory tracking errors during the execution of the underlying robotic arm's servo driver. The horizontal axis represents the execution time in seconds (s), and the vertical axis represents the spatial trajectory tracking error in millimeters (mm). The diagram includes error curves from both the traditional control method and the look-ahead feedback error curve of this invention.

[0074] During the accelerated startup phase, from 0s to 1.0s, the error curve of the traditional control method, due to the lack of compensation for total communication delay and computation time, suffers from instruction lag, causing the tracking error to rapidly climb to over 1.5mm, accompanied by severe high-frequency oscillations. The look-ahead feedback error curve of this invention, thanks to the feedback compensation module's precise calculation of the look-ahead injection point parameters and the implementation of seamless timing instruction coverage based on the target memory address within the circular buffer, effectively offsets the 3ms total system delay blind spot. Throughout the entire runtime, the look-ahead feedback error curve of this invention exhibits extremely strong robustness, with the spatial trajectory tracking error strictly suppressed within a narrow 0.4mm band, completely eliminating the periodic trajectory jitter caused by instruction overlap and lag in traditional systems, ensuring absolute motion accuracy under high-speed, heavy-load conditions.

Claims

1. A robotic arm motion trajectory optimization system based on neural networks, characterized in that, include: An industrial computer runs a non-real-time operating system and has a feature generation module for receiving parameters from a host computer. The industrial computer calculates and infers through the neural network model inside the feature generation module and outputs static parameter data to a shared memory data area. A real-time controller is connected to the industrial computer. It runs an operating system kernel with real-time patches and is equipped with a phase space transformation module, a feedback compensation module, and an optimization solution module. The real-time controller reads the static parameter data through the shared memory data area. Multiple robotic arm servo drives are connected to the real-time controller via a field communication bus; The phase space transformation module is used to convert the static parameter data into a set of static inequality constraints; The feedback compensation module is used to calculate and generate dynamic constraint boundary data based on the measured torque and current data, measured joint position data, and static parameter data returned by the multiple robotic arm servo drives. The optimization solution module is used to receive the set of static inequalities and the dynamic constraint boundary data, re-solve the linear programming problem with time determinism, and generate joint position control commands and velocity feedforward control commands to send to the multiple robotic arm servo drives.

2. The neural network-based robotic arm motion trajectory optimization system according to claim 1, characterized in that, The feature generation module is internally equipped with a task parameter parsing and preprocessing unit; The task parameter parsing and preprocessing unit is used to receive the upper-level task parameters, which include the starting point spatial coordinate parameters, the ending point spatial coordinate parameters, and the end load quality parameters, and to perform engineering normalization processing on the upper-level task parameters to generate normalized task parameters. The neural network model adopts a multilayer perceptron structure. The neural network model does not contain output neurons associated with time variables. The feature generation module is used to extract features based on the normalized task parameters and generate the static parameter data.

3. The neural network-based robotic arm motion trajectory optimization system according to claim 2, characterized in that, Before being deployed to the industrial computer, the neural network model is pre-trained offline by the feature generation module in an offline computing environment. In the offline computing environment, the feature generation module runs a pre-built robotic arm dynamics simulation model to perform iterative optimization calculations on multiple sets of different normalized task parameter samples, generating non-uniform rational B-spline curve parameters and prior pseudo-energy threshold sequences corresponding to the normalized task parameter samples. The feature generation module sets the normalized task parameter samples as the training input data of the neural network model, and merges the non-uniform rational B-spline curve parameters and the prior pseudo-energy threshold sequence as the training label data of the neural network model. The feature generation module stops training and saves the neural network model when the mean square error value drops below a preset error threshold. The error threshold is set based on the steady-state background noise amplitude of the robotic arm joint position sensor.

4. The neural network-based robotic arm motion trajectory optimization system according to claim 1, characterized in that, The static parameter data includes non-uniform rational B-spline curve parameters and a priori pseudo-energy threshold sequence. The feature generation module constructs a continuous spatial pose equation based on the parameters of the non-uniform rational B-spline curve. The feature generation module then packages and integrates the continuous spatial pose equation with the prior pseudo-energy threshold sequence and writes it into the shared memory data area.

5. The neural network-based robotic arm motion trajectory optimization system according to claim 4, characterized in that, The phase space transformation module reads the continuous spatial pose equation and the prior pseudo-energy threshold sequence from the shared memory data area; The phase space transformation module calculates the first and second derivatives of the continuous spatial pose equation with respect to the time variable according to the chain rule, and generates joint velocity conversion formulas for the joint velocity column vector and joint acceleration conversion formulas for the joint acceleration column vector, respectively. The phase space transformation module performs numerical transformation through the joint velocity transformation formula and the joint acceleration transformation formula, converting the joint velocity column vector and the joint acceleration column vector into mathematical expressions consisting only of the path scalar, the first time derivative of the path scalar, and the second time derivative of the path scalar, and constructing a phase space coordinate system; The phase space transformation module compares the pre-set rated maximum torque value and rated minimum torque value of the robotic arm servo driver with the prior pseudo-energy threshold sequence at the same path scalar node to generate a comprehensive boundary threshold sequence.

6. The neural network-based robotic arm motion trajectory optimization system according to claim 5, characterized in that, The phase space transformation module extracts the pre-stored multibody dynamics parameters of the robotic arm; The phase space transformation module combines the first-order geometric partial derivative vector of the continuous spatial pose equation with respect to the path scalar, the second-order geometric partial derivative vector of the continuous spatial pose equation with respect to the path scalar, and the multibody dynamics parameters of the robotic arm to perform matrix multiplication to generate a phase space state coefficient matrix. The phase space transformation module constructs a set of generalized phase space linear inequality constraints from the phase space state coefficient matrix and the integrated boundary threshold sequence, and defines the set of generalized phase space linear inequality constraints as the set of static inequality constraints. The optimization solution module generates a continuous static reference safety envelope boundary line in the phase space state plane according to the static inequality constraint set, and defines the closed two-dimensional graphic region enclosed by the horizontal axis, the vertical axis and the static reference safety envelope boundary line as the initial feasible region.

7. The neural network-based robotic arm motion trajectory optimization system according to claim 5, characterized in that, The feedback compensation module is equipped with a resampling unit. The resampling unit uses a pre-set fixed time window to perform a moving average filter on the measured torque and current data returned by the multiple robotic arm servo drives, and outputs smoothed measured torque and current data. The feedback compensation module extracts the measured joint position data and performs numerical iterative approximation calculations based on the continuous spatial pose equation to obtain the current path scalar value. The feedback compensation module aligns the smoothed measured torque and current data with the current path scalar value at the time nodes to generate a phase space measured torque sequence. The feedback compensation module quantifies the total system delay time and extracts the first-order and second-order time derivative values ​​synchronized with the current path scalar value. It then combines the total system delay time, the current path scalar value, the first-order time derivative value, and the second-order time derivative value to perform motion state extrapolation calculations to generate a look-ahead path scalar value. The feedback compensation module defines the look-ahead path scalar value as a look-ahead injection point parameter and sends it to the optimization solution module.

8. The neural network-based robotic arm motion trajectory optimization system according to claim 7, characterized in that, The feedback compensation module is equipped with a real-time evaluation unit for the actual energy dissipation rate. The real-time evaluation unit of actual energy dissipation rate extracts values ​​from the phase space measured torque sequence based on the current path scalar value to obtain the measured torque column vector corresponding to the current spatial pose node. The real-time evaluation unit of actual energy dissipation rate calculates the first-order partial derivative of the continuous spatial pose equation with respect to the path scalar based on the current path scalar value to generate a geometric Jacobian column vector. The real-time evaluation unit for actual energy dissipation rate calculates the inner product of the measured torque column vector and the geometric Jacobian column vector, generates the actual energy dissipation rate value, and sets the actual energy dissipation rate value as the actual load state parameter characterizing the actual load state of the robotic arm body.

9. The neural network-based robotic arm motion trajectory optimization system according to claim 8, characterized in that, The feedback compensation module is equipped with a boundary dynamic deformation calculation unit. The boundary dynamic deformation calculation unit extracts data from the prior pseudo-energy threshold sequence based on the current path scalar value to obtain the node prior pseudo-energy value. The boundary dynamic deformation calculation unit performs a division comparison on the result of adding the actual energy dissipation rate value to the node prior pseudo-energy value and a preset small positive number to generate a deformation factor value. The boundary dynamic deformation calculation unit constructs a spline smoothing function within the delay blind zone between the current path scalar value and the look-ahead path scalar value. The boundary dynamic deformation calculation unit combines the comprehensive boundary threshold sequence, the deformation factor value, and the spline smoothing function to perform smooth deformation multiplication and addition calculations to generate the dynamic constraint boundary data.

10. The neural network-based robotic arm motion trajectory optimization system according to claim 9, characterized in that, The optimization solution module is equipped with a timing instruction reconstruction and low-level execution overlay unit, and each of the robotic arm servo drivers is equipped with a circular buffer. The optimization solution module receives the look-injection point parameters, sets the look-injection point parameters as the solution start node of the current calculation cycle, extracts and generates a local rolling optimization window, and re-solves the time-deterministic linear programming problem on the first time derivative of the path scalar within the local rolling optimization window to generate the optimal phase space state node sequence. The timing instruction reconstruction and underlying execution overlay unit divides the optimal phase space state node sequence into discrete time nodes, calculates and generates the joint position control instruction and the velocity feedforward control instruction, and encapsulates data bits to generate a data packet sequence. The timing instruction reconstruction and underlying execution overlay unit locates the target write memory address in the circular buffer according to the timestamp value of the look-ahead injection point parameter, and controls the data packet sequence to replace the original old version instruction data inside the circular buffer in ascending order of time, starting from the target write memory address.

Images