Seven-degree-of-freedom motion control method and system for brain-like processor

Abstract

This invention discloses a seven-degree-of-freedom motion control method and system for neuromorphic processors, comprising: S1. Converting the desired joint position, velocity, and acceleration of a robotic arm into a pulse sequence using an incremental encoder; S2. Inputting the pulse sequence into a liquid state machine (LSM) and extracting the spatiotemporal features of the motion trajectory using its dynamic reservoir; S3. Inputting the output features of the LSM and the actual torque value from the previous moment into a multilayer perceptron (MLP) to predict the current joint torque; S4. Adjusting the predicted torque through a feedback controller and performing trajectory tracking. This invention reduces the need for parameter adjustment, can be directly deployed on neuromorphic processors, and features low power consumption and high computational efficiency. Furthermore, it achieves high-precision motion control by learning complex nonlinear relationships in high degrees of freedom through cyclic connections. This invention has broad application prospects in industrial robots, intelligent manufacturing, and medical rehabilitation.

Description

Technical Field

[0001] This invention relates to the field of robot motion control technology, and in particular to a seven-degree-of-freedom motion control method and system for neuromorphic processors. Background Technology

[0002] For an N-DOF (degrees of freedom) robotic arm, its inverse dynamics can be modeled as follows:

[0003] ;

[0004] in , , These are joint position, velocity, and acceleration, respectively. It is the inertia matrix. It is the Coriolis term, which depends on and .at last, The force or torque generated by gravity acting on the system. This is the joint torque vector.

[0005] In inverse dynamics modeling of a seven-DOF robotic arm, obtaining precise parameters is difficult, and the system is susceptible to external disturbances, posing a significant challenge to physical modeling. In recent years, data-driven methods have gradually emerged. While Locally Weighted Projective Regression (LWPR) offers high real-time performance, it requires manual adjustment of many parameters. Kernel methods such as Gaussian Process Regression (GPR) offer high accuracy but have high computational complexity, making them difficult to apply in real-time.

[0006] In contrast, Liquid State Machines (LSMs) extract features from reservoir dynamics and only require processing the output layer during prediction, reducing parameter tuning requirements and offering high computational efficiency. Compared to traditional Artificial Neural Networks (ANNs), SNNs inherently possess the ability to process spatiotemporal information, enabling them to better utilize the temporal continuity between robotic arm motion trajectory data. Summary of the Invention

[0007] To address the problems existing in the prior art, the present invention aims to provide a seven-degree-of-freedom motion control method for neuromorphic processors. This method reduces the need for parameter adjustment, can be directly deployed on neuromorphic processors, and features low power consumption and high computational efficiency. Another objective of the present invention is to provide a seven-degree-of-freedom motion control system for neuromorphic processors that implements the above method.

[0008] To achieve the above objectives, the present invention provides a seven-degree-of-freedom motion control method for neuromorphic processors, comprising:

[0009] S1. The desired joint position, velocity, and acceleration of the robotic arm are converted into a pulse sequence using an incremental encoder;

[0010] S2. Input the pulse sequence into the liquid state machine (LSM) and use its dynamic reservoir to extract the spatiotemporal characteristics of the motion trajectory;

[0011] S3. Input the output characteristics of the liquid state machine (LSM) and the actual torque value of the previous moment into the multilayer perceptron (MLP) to predict the current joint torque;

[0012] S4. Adjust the predicted torque and perform trajectory tracking via the feedback controller.

[0013] Furthermore, the input robotic arm motion trajectory data is projected into a high-dimensional space using a liquid state machine (LSM) to extract and retain the features of the motion trajectory.

[0014] Furthermore, the liquid state machine (LSM) contains 100±20 spiking neurons, with an excitatory to inhibitory neuron ratio of 4:1. The neurons are distributed in a three-dimensional cubic topology. and The connection probabilities between them satisfy:

[0015] ;

[0016] in, The distance between neurons is the Euclidean distance. Here, e is the attenuation coefficient, and C is a mathematical constant. C is a hyperparameter.

[0017] Furthermore, the incremental encoder determines the pulse firing status based on the difference between the current time and the previous time. The difference is calculated as follows:

[0018] ;

[0019] in, It is the motion information of the robotic arm at time t. It is the motion information of the robotic arm at time t-1. When the value is >0, an excitatory neuronal impulse is triggered. When <0, inhibitory neuronal impulses are triggered.

[0020] Furthermore, the number of pulses emitted is:

[0021] ;

[0022] in, The threshold is adjustable.

[0023] Furthermore, the multilayer perceptron (MLP) receives features extracted by the liquid state machine (LSM), takes the extracted features and the actual torque value applied to the robotic arm in the previous time step as input, and performs regression based on these inputs to output the predicted torque for generating the robotic arm's state trajectory.

[0024] Furthermore, the input dimension of the multilayer perceptron (MLP) is the output feature dimension of the liquid state machine (LSM) plus 7-dimensional torque history values, and the output layer consists of 7 linear neurons corresponding to 7 degrees of freedom torque prediction.

[0025] Furthermore, the step of using the particle swarm optimization (PSO) algorithm to search for the optimal structural parameters of the liquid state machine (LSM) includes:

[0026] 1) The connection probability from the input layer to the liquid layer;

[0027] 2) Within the liquid layer, there are four types of connection probabilities: excitation → excitation / inhibition → excitation / excitation → inhibition / inhibition → inhibition.

[0028] 3) Number of neurons.

[0029] Furthermore, in step S1, the input trajectory has 21 dimensions, including 7-dimensional joint angles, 7-dimensional joint velocities, and 7-dimensional joint accelerations.

[0030] A seven-DOF motion control system for a neuromorphic processor that implements the above control method includes:

[0031] 1) Incremental encoding module: includes 21 input channels, with 2 spiking neurons per channel;

[0032] 2) Liquid State Machine (LSM) processing module: containing a reservoir of spiking neurons arranged in a three-dimensional cube;

[0033] 3) Multilayer Perceptron (MLP) Torque Prediction Module: Fully Connected Neural Network;

[0034] 4) PD feedback controller;

[0035] 5) Torque history value caching unit;

[0036] Each module achieves asynchronous event-driven operation through the neuromorphic computing core of the neuromorphic processor.

[0037] This invention reduces the need for parameter adjustment, can be directly deployed on neuromorphic processors, and features low power consumption and high computational efficiency. Furthermore, it achieves high-precision motion control by learning complex nonlinear relationships in high-degree-of-freedom environments through cyclic connections. This invention has broad application prospects in fields such as industrial robotics, intelligent manufacturing, and medical rehabilitation. Attached Figure Description

[0038] Figure 1 A schematic diagram of the activity of a single spiking neuron;

[0039] Figure 2 This is a schematic diagram of the overall control architecture of the present invention;

[0040] Figure 3This is a diagram of a spiking neural network structure.

[0041] Figure 4 This is a schematic diagram of incremental pulse coding. Detailed Implementation

[0042] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0043] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0044] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0045] The following is in conjunction with the appendix Figures 1-4 Specific embodiments of the present invention will be described in detail below. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0046] SNNs are a type of artificial neural network inspired by biological nervous systems. Unlike traditional artificial neural networks, SNNs transmit information through pulse sequences. Similar to biological neurons, when the voltage of a spiking neuron exceeds a threshold, the neuron emits a pulse, and its membrane voltage is reset to the resting potential. Figure 1The activity of a single neuron is demonstrated. Before a neuron fires a pulse, the membrane potential accumulates over time, influenced not only by the current input but also by the retention of past information. This characteristic gives spiking neural networks a natural ability to process spatiotemporal data. Compared to simple feedforward SNN motion control networks that control low degrees of freedom (two or three degrees of freedom), the method of this invention uses recurrent connections, enabling it to better learn complex nonlinear relationships in high-degree-of-freedom robotic arms.

[0047] like Figures 2-4 As shown, the present invention provides a seven-degree-of-freedom motion control method for neuromorphic processors, comprising:

[0048] S1. The desired joint position, velocity, and acceleration of the robotic arm are converted into a pulse sequence using an incremental encoder;

[0049] S2. Input the pulse sequence into the liquid state machine (LSM) and use its dynamic reservoir to extract the spatiotemporal characteristics of the motion trajectory;

[0050] S3. Input the output characteristics of the liquid state machine (LSM) and the actual torque value of the previous moment into the multilayer perceptron (MLP) to predict the current joint torque;

[0051] S4. Adjust the predicted torque and perform trajectory tracking via the feedback controller.

[0052] Liquid State Machine (LSM):

[0053] Liquid state machines (LSMs) are a type of recurrent neural network based on spiking neurons. For example... Figure 3 As shown, the Liquid State Machine (LSM) mainly consists of an input layer, a liquid layer, and a readout layer. The connections between the input and liquid layers are random, and the input layer only connects to the excitatory neurons in the liquid layer. The liquid layer is the main part of the LSM, and the excitatory and inhibitory neurons within it are also randomly connected, including self-to-self connections and self-to-other neuron connections. The input current of the excitatory neuron k in the liquid layer can be obtained from the following equation:

[0054] ;

[0055] in, It is the input current of the input layer at the current time step. This is the output of the liquid layer itself at the previous time step. n is the number of neurons in the input layer, and m is the number of neurons in the liquid layer. These are the weights from the input layer to the liquid layer. This is the input for the current time step. It is the connection weight between the liquid layer and itself. It is the final output of the liquid layer at the previous time step.

[0056] Due to the recurrent connections and nonlinear activation between spiking neurons, the liquid state machine (LSM) has the ability to extract features, dynamically memorize and utilize past information, thus enabling it to better process time series data.

[0057] The overall framework of the control system is as follows Figure 2 As shown, the trajectory generator generates the desired state of the robotic arm (joint angles, velocity, and acceleration) for the next moment, converts it into pulse input via an incremental encoder, and feeds it into the neural network. The network's readout layer outputs the predicted torque.

[0058] The trajectory generator in this invention is implemented using the Ruckig library in ROS. The Ruckig trajectory generator generates smooth trajectories that satisfy motion constraints in real time using a high-order polynomial interpolation algorithm. Its core lies in using a combination of seven stages with different jerk rates to precisely connect any initial state with the target state. This invention is based on analytical solution calculation rather than numerical iteration, enabling trajectory planning to be completed in microseconds. It achieves continuous state transitions through a state-space model, providing a complete desired state including position, velocity, and acceleration as needed according to the control cycle.

[0059] Then, the feedback controller adjusts the predicted torque based on the current and desired states to ensure accurate trajectory tracking of the robotic arm. Finally, the actual torque applied to the robotic arm at the previous moment is directly encoded and provided to the network to help predict the torque at subsequent moments.

[0060] The feedback controller is implemented by a PD controller. The proportional gain P is a crucial parameter in the PD controller, used to adjust the proportional relationship between the system output and the error. Generally, increasing the P value can speed up the system's response, but setting it too high may cause system oscillations or even instability. The derivative gain D is used to compensate for system damping and suppress oscillations, its function being to reduce system overshoot and stabilize oscillations. Typically, increasing the derivative gain D can improve system stability, but setting it too high may lead to a slow system response or even overshoot. By properly adjusting the proportional gain P and derivative gain D, the PD controller can achieve optimal control performance in the robotic arm system, realizing precise position and attitude control.

[0061] In the PID control of a robotic arm, to optimize the settings of the proportional gain (P) and derivative gain (D), a stepwise strategy of adjusting P first and then D is adopted, and simulation and experimental verification are combined to evaluate the tracking accuracy and steady-state performance. The initial range of the proportional gain P is recommended to be set between [1, 10]. A larger P value can improve the system's response speed and reduce steady-state error, but if P is too large, it may lead to overshoot or continuous oscillation. Therefore, during the simulation phase, the P value should be gradually increased, and the step response of the robotic arm should be observed to select a critical value that allows the system to stabilize quickly without significant overshoot. Subsequently, the P value is further fine-tuned in experimental verification to ensure that the tracking accuracy in actual motion meets the requirements. The derivative gain D is used to suppress overshoot and oscillation, and its typical range is between [0.005, 1]. A smaller D value can avoid high-frequency noise amplification, while appropriately increasing D can improve dynamic performance. When adjusting D, it should be gradually increased based on a fixed P, and the damping effect of the system response should be observed to ensure that the robotic arm can still converge smoothly during rapid movement. Ultimately, the optimal P and D combination should enable the robotic arm to exhibit good tracking accuracy, fast response, and stable, oscillation-free characteristics in both simulation and experiment.

[0062] Network structure:

[0063] Spiking neural networks used for inverse dynamics solutions, such as Figure 3 As shown, the network consists of two parts:

[0064] Liquid State Machine (LSM): The input robotic arm motion trajectory data is projected into a high-dimensional space using a liquid state machine, and the features of the motion trajectory are extracted and preserved.

[0065] Multilayer Perceptron (MLP): It receives features extracted by the LSM (Level Scale Mining) and uses these features along with the actual torque value applied to the robotic arm at the previous time step as input. Based on these inputs, it performs regression and outputs a predicted torque to generate the robotic arm's state trajectory.

[0066] The connection weights between input neurons and excitatory neurons in the Liquid State Machine (LSM), as well as the connection weights between neurons in the LSM, are constant after initialization and do not participate in network training. This reduces training complexity and accelerates model convergence. The connection between the LSM and the Multilayer Perceptron (MLP) is feedforward and updated according to the backpropagation rule.

[0067] Incremental encoder:

[0068] The input trajectory is encoded using incremental encoding, such as... Figure 4 As shown, the incremental encoder determines the pulse output based on the difference between the current time and the previous time.

[0069] ;

[0070] in, It is the input at time t (the motion information of the robotic arm at time t). It is the input at time t-1 (the motion information of the robotic arm at time t-1).

[0071] For a seven-DOF robotic arm, the input trajectory has 21 dimensions (including 7 dimensions of joint angles, 7 dimensions of joint velocities, and 7 dimensions of joint accelerations). Each dimension of the input corresponds to two spiking neurons. When the difference is positive, the excitatory neuron pulses; conversely, the inhibitory neuron pulses. The network has hyperparameters. ( The adjustable threshold controls the number of pulses fired. :

[0072] ;

[0073] Incremental encoders capture the dynamic changes in the input robot motion trajectory data, making it easier for the network to capture the complex changes and trends in the robot's motion trajectory.

[0074] In spiking neural networks, a timestep refers to a unit of time segment in which the network performs one iteration update. It can be understood as the neural network receiving and processing input and generating output at a discrete point in time. Spiking neural networks represent neuron activity as discrete pulse signals by discretizing the input and neuron states to simulate the information transmission process between neurons. In this invention, an input is processed in multiple timesteps. That is, the network receives the current input at multiple discrete points in time.

[0075] Searching for the structure and parameters of a liquid state machine (LSM):

[0076] The Liquid State Machine (LSM) uses approximately 100 spiking neurons. We randomly selected a portion of them as inhibitory neurons and the remainder as excitatory neurons. The LSM neurons form a cubic structure. We define neurons according to the following formula. and Distance between:

[0077] ;

[0078] in, It is the x-coordinate of the i-th neuron in the liquid layer. It is the x-coordinate of the j-th neuron in the liquid layer. It is the y-coordinate of the i-th neuron in the liquid layer. It is the y-coordinate of the j-th neuron in the liquid layer. It is the z-coordinate of the i-th neuron in the liquid layer. It is the z-coordinate of the j-th neuron in the liquid layer.

[0079] Then, the distance is used to calculate the neuron according to the following formula. and The probability of connections between them.

[0080] ;

[0081] in, The distance between neurons is the Euclidean distance. Here, is the attenuation coefficient, e is a mathematical constant, and C is a hyperparameter. As shown in the formula, the smaller the distance between neurons, the higher the probability of connection between them. This also improves the biological rationale of the network. The parameter search range is shown in Table 1:

[0082] Table 1: Parameter Search Range

[0083] ;

[0084] Where n is the number of neurons in the liquid layer. It represents the connection probability between neurons in the input layer and neurons in the liquid layer. It represents the connection probability between excitatory neurons in the liquid layer. It represents the connection probability between excitatory neurons and inhibitory neurons in the liquid layer. It represents the connection probability between inhibitory neurons and excitatory neurons in the liquid layer. It represents the connection probability between inhibitory neurons in the liquid layer.

[0085] Since different liquid state machine (LSM) structures exhibit varying feature extraction capabilities, we employ the particle swarm optimization (PSO) algorithm for the search. Key search parameters include the number of neurons, the connection probabilities from input layer neurons to the liquid layer, and the four connection probabilities between excitatory and inhibitory neurons within the liquid layer.

[0086] Particle Swarm Optimization (PSO) utilizes the information sharing among individuals in a swarm to cause the movement of the entire swarm to evolve from disorder to order in the problem solution space, thereby obtaining the optimal solution.

[0087] PSO is initialized with a swarm of random particles (random solutions). The optimal solution is then found iteratively. In each iteration, the particles update themselves by tracking two "extremes" (pbest, gbest). After finding these two optimal values, the particles update their velocity and position using the following formula.

[0088] ;

[0089] Where i = 1, 2, ..., N, and N is the total number of particles in this group;

[0090] It is the velocity of the particle. It is a random number between (0,1). This represents the particle's current position. and It is a learning factor. (Individual Optimal): Represents the historical optimal position found by the i-th particle during the iteration process (i.e., the solution with the best fitness in all previous iterations). (Global Optimal): This refers to the optimal position found by all particles in the entire particle swarm during the iteration process (i.e., the best solution found by the entire swarm so far).

[0091] This invention employs the Particle Swarm Optimization (PSO) algorithm for automatic hyperparameter optimization. The optimization process follows these steps: First, the PSO algorithm initializes a set of hyperparameter combinations and calculates the error index of the torque prediction model based on the current parameter configuration. Then, the PSO algorithm dynamically adjusts the particle positions and velocities according to the error evaluation results to generate a new generation of hyperparameter combinations. Through iterative optimization, the algorithm continuously reduces the prediction error and eventually converges to the hyperparameter configuration that optimizes the torque prediction accuracy.

[0092] A multilayer perceptron (MLP) is a feedforward neural network consisting of an input layer, one or more hidden layers (we'll use one hidden layer in this paper), and an output layer. An MLP is a fully connected neural network, where every neuron in each layer is connected to every neuron in the next layer. The following is the computational formula for a perceptron:

[0093] ;

[0094] Where X is the input of the layer, W is the weight of the layer, B is the bias of the layer, T is the activation function of the layer (the Tanh function is used in this invention), and y is the output of the layer.

[0095] The present invention also provides a seven-degree-of-freedom motion control system for a neuromorphic processor that implements the above-described control method, comprising:

[0096] 1) Incremental encoding module: includes 21 input channels, with 2 spiking neurons per channel;

[0097] 2) Liquid State Machine (LSM) processing module: containing a reservoir of spiking neurons arranged in a three-dimensional cube;

[0098] 3) Multilayer Perceptron (MLP) Torque Prediction Module: Fully Connected Neural Network;

[0099] 4) PD feedback controller;

[0100] 5) Torque history value caching unit;

[0101] Each module achieves asynchronous event-driven operation through the neuromorphic computing core of the neuromorphic processor.

[0102] Example:

[0103] The LSM has 174 neurons, with 80% excitability, and 42 input neurons. The PSO search hyperparameters are configured as follows:

[0104]

[0105] This MLP (Multilayer Perceptron) network employs a classic three-layer fully connected structure. The input layer receives a 181-dimensional composite feature vector composed of features extracted from a 174-dimensional LSM and 7-dimensional actual joint torque values. The hidden layer has 128 neurons, using the Tanh activation function for nonlinear transformation. The output layer has 7 neurons corresponding to the torque prediction values ​​of each joint of the robotic arm. Model training uses a batch processing strategy with step_size=50, iteratively optimizing for 300 epochs using the Adam optimizer (learning rate lr=0.001, weight decay coefficient weight_decay=0.01). A step learning rate scheduler (StepLR) is also introduced, which works by multiplying the current learning rate by 0.1 (gamma=0.1) after every 50 epochs (step_size=50) of training for decay.

[0106] The technical problem addressed by this invention is the limitation of inverse dynamics physical modeling for robotic arms, which is constrained by the difficulty in obtaining parameters and external interference. Previous data-driven methods each have their advantages, but they suffer from limitations such as high manual parameter tuning requirements, high computational complexity, high resource consumption, and low control degrees of freedom. The main technical solution of this invention is to propose a motion control method and system oriented towards neuromorphic processors. This method combines a liquid state machine and a multilayer perceptron to predict the torque required for robotic arm movement. This method reduces the need for parameter adjustment, can be directly deployed on neuromorphic processors, and features low power consumption and high computational efficiency. Furthermore, by learning complex nonlinear relationships in high degrees of freedom through cyclic connections, it achieves high-precision motion control. This invention has broad application prospects in industrial robotics, intelligent manufacturing, and medical rehabilitation.

[0107] Any process or method described in the flowcharts of this invention or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, which can be implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device. The computer-readable medium can be any medium containing a program for storage, communication, propagation, or transmission for use by an execution system, apparatus, or device, including read-only memory, magnetic disks, or optical disks.

[0108] In the description of this specification, references to terms such as "embodiment," "example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, those skilled in the art can combine or combine the different embodiments or examples described in this specification and the features therein without causing contradiction.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A seven-degree-of-freedom motion control method for neuromorphic processors, characterized in that, The method includes: S1. The desired joint position, velocity, and acceleration of the robotic arm are converted into a pulse sequence using an incremental encoder; S2. Input the pulse sequence into the liquid state machine (LSM) and use its dynamic reservoir to extract the spatiotemporal characteristics of the motion trajectory; S3. Input the output characteristics of the liquid state machine (LSM) and the actual torque value of the previous moment into the multilayer perceptron (MLP) to predict the current joint torque; S4. Adjust the predicted torque via the feedback controller and perform trajectory tracking; The method further includes: (a) The trajectory generator uses the Ruckig library in ROS to generate smooth trajectories in real time through a high-order polynomial interpolation algorithm, completes trajectory planning based on analytical solutions, and provides a complete desired state including position, velocity, and acceleration. (b) The multilayer perceptron (MLP) is trained using the Adam optimizer with a learning rate of 0.001, a weight decay coefficient of 0.01, a batch size of 50, and iterative optimization for 300 epochs. A step-learning rate scheduler is used to multiply the learning rate by 0.1 every 50 epochs to decay the learning rate. (c) The formula for calculating the input current of the excitatory neurons in the liquid state machine (LSM) is as follows: ； in, It is the input current of the input layer at the current time step. It is the output of the liquid layer itself at the previous time step; n is the number of neurons in the input layer, and m is the number of neurons in the liquid layer; These are the weights from the input layer to the liquid layer. This is the input for the current time step. It is the connection weight between the liquid layer and itself. It is the final output of the liquid layer at the previous time step.

2. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 1, characterized in that, The input robotic arm motion trajectory data is projected into a high-dimensional space using a liquid state machine (LSM) to extract and retain the features of the motion trajectory.

3. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 2, characterized in that, The liquid state machine (LSM) contains 100±20 spiking neurons, with an excitatory to inhibitory neuron ratio of 4:

1. The neurons are distributed in a three-dimensional cubic topology. and The connection probabilities between them satisfy: ； in, The distance between neurons is the Euclidean distance. Here, e is the attenuation coefficient, and C is a mathematical constant. C is a hyperparameter.

4. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 1, characterized in that, The incremental encoder determines the pulse firing status based on the difference between the current time and the previous time. The difference is calculated as follows: ； in, It is the motion information of the robotic arm at time t. It is the motion information of the robotic arm at time t-1. When the value is >0, an excitatory neuronal impulse is triggered. When <0, inhibitory neuronal impulses are triggered.

5. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 4, characterized in that, The number of pulses emitted is: ； in, The threshold value is used.

6. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 1, characterized in that, The multilayer perceptron (MLP) receives features extracted by the liquid state machine (LSM), takes the extracted features and the actual torque value applied to the robotic arm in the previous time step as input, performs regression based on these inputs, and outputs the predicted torque that generates the state trajectory of the robotic arm.

7. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 6, characterized in that, The input dimension of the multilayer perceptron (MLP) is the output feature dimension of the liquid state machine (LSM) plus 7-dimensional torque history values, and the output layer consists of 7 linear neurons corresponding to 7 degrees of freedom torque prediction.

8. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 1, characterized in that, The optimal structural parameters of the liquid state machine (LSM) were searched using the particle swarm optimization (PSO) algorithm, including: 1) The connection probability from the input layer to the liquid layer; 2) Within the liquid layer, there are four types of connection probabilities: excitation → excitation / inhibition → excitation / excitation → inhibition / inhibition → inhibition. 3) Number of neurons.

9. The seven-degree-of-freedom motion control method for neuromorphic processors according to claim 1, characterized in that, In step S1, the input trajectory has 21 dimensions, including 7-dimensional joint angles, 7-dimensional joint velocities, and 7-dimensional joint accelerations.

10. A seven-degree-of-freedom motion control system for neuromorphic processors, used to implement the seven-degree-of-freedom motion control method for neuromorphic processors according to any one of claims 1-9, characterized in that, The system includes: 1) Incremental encoding module: includes 21 input channels, with 2 spiking neurons per channel; 2) Liquid State Machine (LSM) processing module: containing a reservoir of spiking neurons arranged in a three-dimensional cube; 3) Multilayer Perceptron (MLP) Torque Prediction Module: Fully Connected Neural Network; 4) PD feedback controller; 5) Torque history value caching unit; Each module achieves asynchronous event-driven operation through the neuromorphic computing core of the neuromorphic processor.

Images