Neural computing method for legged robot jumping training

Abstract

This invention relates to the field of pulse neural computation technology and discloses a neural computation method for jumping training of legged robots. The method includes: acquiring foot reaction force and joint angular velocity state data and generating an asynchronous pulse sequence using a first-order differential threshold gating rule; adjusting the synaptic conduction delay constant based on the channel timing jitter standard deviation to generate an aligned pulse feature matrix; inputting the feature matrix into a neural computation network to drive asynchronous transitions in neuronal membrane potentials and trigger weight updates; resetting the membrane potential and truncating weight updates by a charge discharge rule when the pulse firing frequency exceeds an upper limit threshold; and converging the network output pulses to output a control parameter matrix. This invention constructs a wavefront alignment mechanism based on conduction delay correction to achieve precise alignment of pulse flows, eliminate disordered information divergence during non-stationary impact phases, avoid oversaturation of the computational state, and ensure deterministic convergence of the output.

Description

Technical Field

[0001] This invention relates to a neural computation method for jumping training of legged robots, belonging to the field of pulse neural computation technology. Background Technology

[0002] Currently, asynchronous event-driven pulse architecture is the mainstream technology for processing multi-source motion stream data. When a discrete pulse event is received, a state update is triggered to maintain a low baseline power consumption. At this time, discrete pulse sequences are transmitted from each level of synaptic topology to the neuron cell body, driving the temporal integral evolution of the membrane potential state equation. When the membrane potential charge integral exceeds a preset threshold, the subsequent neuron releases a pulse. The entire system is highly dependent on a stable signal input frequency and the deterministic connection of bus transmission timing. However, in the process of solving high-burst discontinuous motion data, the landing transient causes strong physical collisions, resulting in transient bursts in the spatiotemporal pulse sequence. Due to random timing jitter in the asynchronous transmission bus, the time when the pulses from each channel arrive at the neuron cell body undergoes asymmetric phase shift, causing the temporal feature matrix to become disordered. At the same time, the dense pulse flow generated by the physical collision causes a large amount of charge superposition in the subsequent network topology, causing the neuron membrane potential to enter a saturation deadlock state. The synaptic weight iteration produces parameter drift, and the system solution delay increases dramatically.

[0003] To address such phenomena, simply expanding the network topology or raising the sampling threshold frequency using linear correction methods not only leads to computational resource exhaustion due to memory bandwidth limitations but also exacerbates the propagation of noise pulses. This fails to eliminate solution divergence under non-stationary transition states, making it difficult to meet real-time high-precision control requirements. Linear adjustments at the physical components and hardware levels have limitations, and the motion control algorithm also has shortcomings. For example, Chinese invention patent application CN115546547A discloses a motion control method for a legged robot based on a spiking neural network in complex environments. It relies on a front-end camera to collect overall environmental images for terrain classification and adjust gait and trajectory height accordingly. However, in the case of transient collisions during jump training of the legged robot, the visual perception mechanism cannot capture the microsecond-level sudden increase in foot reaction force and surface pulse timing jitter caused by bus transmission due to image processing delays. This results in disordered timing feature matrices and membrane potential deadlock divergence. The overall environment prediction logic and the surface boundary conditions of high-frequency transient collisions are fundamentally mismatched, failing to solve the problems of state saturation and output solution divergence.

[0004] Therefore, how to correct the phase deviation of multi-source pulse streams in asynchronous transmission and block the superposition and divergence of explosive impulse noise in the subsequent network topology has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A neural computation method for jumping training of legged robots, comprising the following steps: Step S1: Obtain the timing characteristic data of the foot contact reaction force and the joint angular velocity state data of the legged robot during the jumping training process; Step S2: The timing characteristic data of foot contact reaction force and joint angular velocity state data are processed using the first-order differential threshold gating rule to generate an asynchronous pulse sequence; Step S3: Obtain the standard deviation of random timing jitter of the data transmission channel, adjust the conduction delay constant of each synaptic node according to the inverse proportional rule to calibrate the arrival time of each pulse in the asynchronous pulse sequence, and generate a timing-aligned pulse feature matrix. Step S4: Input the pulse feature matrix into the neural computing network, drive the asynchronous transition of the membrane potential of each neuron in the neural computing network through pulse events, and trigger the synaptic weight update action according to the asynchronous transition state of the membrane potential of each neuron, while monitoring the pulse firing frequency of each neuron within a 1ms time window. Step S5: When the pulse firing frequency exceeds the preset upper limit threshold of 3000Hz, the membrane potential of the corresponding neuron in each neuron is reset by the charge discharge rule, and the synaptic weight update action at the current moment is cut off. Step S6: Converge the pulses output by the neural computing network, generate a control parameter matrix, and output it.

[0006] Preferably, in step S2, the step of processing the foot contact reaction time-series feature data and joint angular velocity state data using a first-order differential threshold gating rule includes: step S21, obtaining the current values ​​of the foot contact reaction time-series feature data and joint angular velocity state data at the current sampling time; step S22, calculating the absolute value of the discrete difference between the current value and the previous value at the previous sampling time; step S23, comparing the absolute value of the discrete difference with a preset dynamic bias threshold, and generating a corresponding event pulse signal to form an asynchronous pulse sequence when the absolute value of the discrete difference exceeds the dynamic bias threshold.

[0007] Preferably, in step S4, the membrane potential of each neuron in the neural computing network changes at a preset discrete time step. The membrane potential value of each control cycle is obtained by multiplying the value of the membrane potential at the previous moment by a preset membrane potential decay coefficient, and then summing the product of the pulse output state of the preceding neuron and the weight of the corresponding synapse.

[0008] Preferably, in step S3, adjusting the conduction delay constant of each synaptic node according to the random timing jitter standard deviation by the reverse proportional rule includes the following steps: Step S31, making the conduction delay constant of each synaptic node decrease as the random timing jitter standard deviation increases, so as to calibrate the arrival time of each pulse in the asynchronous pulse sequence and generate a timing-aligned pulse feature matrix.

[0009] Preferably, in step S5, the step of monitoring the pulse firing frequency of each neuron within a 1ms time window includes: step S51, counting the total number of pulse firings of each neuron within a 1ms time window; step S52, calculating the ratio of the total number of pulse firings to the length of the 1ms time window to obtain the pulse firing frequency, and activating the charge discharge rule when the pulse firing frequency exceeds a preset upper limit threshold.

[0010] Preferably, in step S52, when the charge discharge rule is activated, the following steps are included: step S521, resetting the membrane potential of the target neuron whose pulse firing frequency exceeds the preset upper limit threshold to the preset resting potential value; step S522, stopping the synaptic weight update action corresponding to the target neuron, and setting the synaptic weight update increment in the synaptic weight update action corresponding to the target neuron to zero.

[0011] Preferably, in step S6, the step of generating the control parameter matrix further includes: step S61, performing boundary monitoring on the rate of change of the control parameter matrix through rate of change constraint rules.

[0012] Preferably, in step S61, the step of monitoring the boundary of the rate of change of the control parameter matrix through the rate of change constraint rule includes: step S611, monitoring whether the rate of change of the control parameter matrix shows a monotonically increasing trend within a series of preset pulse periods; step S612, when the rate of change shows a monotonically increasing trend within three consecutive pulse periods and exceeds the preset critical safety boundary, applying a step-by-step penalty to the rate of change of the control parameter matrix to reduce the rate of change of the control parameter matrix according to the preset penalty step size.

[0013] Preferably, in step S6, after outputting the control parameter matrix, the following step is also included: Step S62, the generated control parameter matrix is ​​transmitted to the downstream drive calculation unit, and the downstream drive calculation unit calculates and generates the drive control current for controlling the power components of each joint in the legged robot according to the control parameter matrix, so as to adjust the posture of the legged robot.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In the neural computation of legged robot jumping training, a differential gating operator is used to adaptively capture the rate of change of physical quantities in multi-source motion flow data. Combined with a spatiotemporal synaptic delay feedforward decoupling operator to track the random timing jitter of the bus in real time, a wavefront alignment mechanism driven by dynamic gradient correction of the conduction delay constant is constructed in the logic space layer. This mechanism transforms the deviation caused by the physical transition state lag in situ into a phase offset with monotonically deterministic correlation, so that the asynchronous pulse flow spontaneously achieves precise alignment of the timing feature matrix before reaching the cell body. This not only eliminates the disordered information divergence generated in the non-stationary impact stage, but also enables the network to obtain adaptive control capability to counteract parasitic interference from external bus transmission.

[0015] 2. By integrating the potential phase space asymmetric reversal rule into the pulse integral evolution topology, the system gains the ability to block sudden changes in response to explosive impact interference. When the transient physical collision causes excessive divergence of local state features, the system periodically monitors the pulse firing frequency within a specific time window. When the firing frequency of the target neuron exceeds the safety upper limit threshold, the system automatically activates the charge discharge operator to forcibly reset the potential value to the resting potential and simultaneously calls the weight update truncation operator to stop the weight accumulation iteration between the target neuron and its connected synapses. This blocks the propagation of noise pulses in the deep layers of the network at the logic layer, preventing the computational state from falling into oversaturation due to physical quantity steps.

[0016] 3. A phase space damping adjustment operator is configured in the output layer. By monitoring the hard boundary of the rate of change of the calculation control parameter tensor, a step-by-step penalty constraint mechanism is constructed for sudden diffusion of multi-axial motion. When the rate of change of the parameter tensor shows a monotonically increasing trend in multiple consecutive pulse cycles and crosses the safety boundary, the operator applies a step-by-step penalty to the slope of the parameter tensor, forcibly restricting the transition rate of the eigenvector in the nonlinear phase space. This forms a closed-loop behavior adjustment at the output terminal, so that the output still maintains deterministic convergence characteristics under extreme conditions such as landing impact, and prevents the calculation results from falling into disordered divergence due to sudden changes in system state. Attached Figure Description

[0017] Figure 1 This is a flowchart of the operation steps of the neural computation method for training a legged robot to jump, as described in this invention. Figure 2 This is a diagram of the perception and control architecture of a legged robot that implements the neural computation method of this invention.

[0018] The implementation, functional characteristics, and advantages of the technical solution of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A neural computation method for training jumping in a legged robot includes the following steps: Step S1: Obtain the timing characteristic data of the foot contact reaction force and the joint angular velocity state data of the legged robot during the jumping training process; Step S2: The timing characteristic data of foot contact reaction force and joint angular velocity state data are processed using the first-order differential threshold gating rule to generate an asynchronous pulse sequence; Step S3: Obtain the standard deviation of random timing jitter of the data transmission channel, adjust the conduction delay constant of each synaptic node according to the inverse proportional rule to calibrate the arrival time of each pulse in the asynchronous pulse sequence, and generate a timing-aligned pulse feature matrix. Step S4: Input the pulse feature matrix into the neural computing network, drive the asynchronous transition of the membrane potential of each neuron in the neural computing network through pulse events, and trigger the synaptic weight update action according to the asynchronous transition state of the membrane potential of each neuron, while monitoring the pulse firing frequency of each neuron within a 1ms time window. Step S5: When the pulse firing frequency exceeds the preset upper limit threshold of 3000Hz, the membrane potential of the corresponding neuron in each neuron is reset by the charge discharge rule, and the synaptic weight update action at the current moment is cut off. Step S6: Converge the pulses output by the neural computing network, generate a control parameter matrix, and output it.

[0021] Preferably, in step S2, the step of processing the foot contact reaction time-series feature data and joint angular velocity state data using a first-order differential threshold gating rule includes: step S21, obtaining the current values ​​of the foot contact reaction time-series feature data and joint angular velocity state data at the current sampling time; step S22, calculating the absolute value of the discrete difference between the current value and the previous value at the previous sampling time; step S23, comparing the absolute value of the discrete difference with a preset dynamic bias threshold, and generating a corresponding event pulse signal to form an asynchronous pulse sequence when the absolute value of the discrete difference exceeds the dynamic bias threshold.

[0022] Preferably, in step S4, the membrane potential of each neuron in the neural computing network changes at a preset discrete time step. The membrane potential value of each control cycle is obtained by multiplying the value of the membrane potential at the previous moment by a preset membrane potential decay coefficient, and then summing the product of the pulse output state of the preceding neuron and the weight of the corresponding synapse.

[0023] Preferably, in step S3, adjusting the conduction delay constant of each synaptic node according to the random timing jitter standard deviation by the reverse proportional rule includes the following steps: Step S31, making the conduction delay constant of each synaptic node decrease as the random timing jitter standard deviation increases, so as to calibrate the arrival time of each pulse in the asynchronous pulse sequence and generate a timing-aligned pulse feature matrix.

[0024] Preferably, in step S5, the step of monitoring the pulse firing frequency of each neuron within a 1ms time window includes: step S51, counting the total number of pulse firings of each neuron within a 1ms time window; step S52, calculating the ratio of the total number of pulse firings to the length of the 1ms time window to obtain the pulse firing frequency, and activating the charge discharge rule when the pulse firing frequency exceeds a preset upper limit threshold.

[0025] Preferably, in step S52, when the charge discharge rule is activated, the following steps are included: step S521, resetting the membrane potential of the target neuron whose pulse firing frequency exceeds the preset upper limit threshold to the preset resting potential value; step S522, stopping the synaptic weight update action corresponding to the target neuron, and setting the synaptic weight update increment in the synaptic weight update action corresponding to the target neuron to zero.

[0026] Preferably, in step S6, the step of generating the control parameter matrix further includes: step S61, performing boundary monitoring on the rate of change of the control parameter matrix through rate of change constraint rules.

[0027] Preferably, in step S61, the step of monitoring the boundary of the rate of change of the control parameter matrix through the rate of change constraint rule includes: step S611, monitoring whether the rate of change of the control parameter matrix shows a monotonically increasing trend within a series of preset pulse periods; step S612, when the rate of change shows a monotonically increasing trend within three consecutive pulse periods and exceeds the preset critical safety boundary, applying a step-by-step penalty to the rate of change of the control parameter matrix to reduce the rate of change of the control parameter matrix according to the preset penalty step size.

[0028] Preferably, in step S6, after outputting the control parameter matrix, the following step is also included: Step S62, the generated control parameter matrix is ​​transmitted to the downstream drive calculation unit, and the downstream drive calculation unit calculates and generates the drive control current for controlling the power components of each joint in the legged robot according to the control parameter matrix, so as to adjust the posture of the legged robot.

[0029] Example 1: When the current-legged robot performs a longitudinal jump on unstructured gravel ground, the joint transmission mechanism generates transient high-frequency vibrations due to the physical impact of the ground, causing numerical fluctuations in the pressure signal collected by the foot contact sensor. The data acquisition module acquires this pressure signal and sends the raw value to the neural computing network at a sampling frequency of 1000Hz. The input unit executes a first-order differential threshold gating rule, setting the dynamic bias threshold of the pressure signal to 2.5N. When the change in the ground reaction force of the left forefoot exceeds 2.5N between two consecutive sampling times, the input unit generates a discrete asynchronous pulse signal and marks it as a highly sparse event feature, transforming the simulated pressure quantity into a pulse time feature matrix. To calibrate the transmission phase deviation generated by the main control bus, the central controller acquires the data transmission phase deviation of the main control bus in the data transmission channel. The random timing jitter standard deviation is calculated using a linear deduction calibration rule based on a linear time delay attenuation physical model. This converts random timing disturbances in the bus communication link into a reduction in the conduction time of the corresponding synaptic node. The central controller multiplies the collected random timing jitter standard deviation with a preset time cascade calibration coefficient to obtain the timing delay increment. The system's preset reference conduction delay constant is subtracted from the timing delay increment to output the actual conduction delay constant of the synaptic node within the current sampling period. The actual conduction delay constant is greater than zero and less than the reference conduction delay constant. The time cascade calibration coefficient is a fixed constant that balances the relationship between the bus physical impedance and the pulse transmission rate. When the random timing jitter standard deviation of the data bus communication link increases, the synaptic conduction delay constant decreases, allowing the asynchronous pulse flow that arrives late due to bus jitter to receive phase feedforward compensation before reaching the cell body.

[0030] The system introduces a spatiotemporal synaptic delay feedforward decoupling operator. The central controller measures the standard deviation of random timing jitter generated by the main control bus in real time and adjusts the endogenous conduction delay constant of each neuron's synaptic node in the neural computing network in reverse proportionally based on this parameter. This causes the conduction delay constant of the synaptic node to automatically decrease as the standard deviation of random timing jitter increases. Specifically, in the adjustment control link of each synaptic node, the system presets a reference conduction delay constant. The central controller multiplies the currently measured standard deviation of random timing jitter by a fixed time cascade calibration coefficient to calculate the conduction delay constant of the bus caused by parasitic interference. The timing delay increment is obtained by directly subtracting the reference conduction delay constant from the timing delay increment, which is used as the current actual conduction delay constant of the synaptic node. Through this reverse subtraction mechanism, when the jitter of the data bus transmission increases, the conduction time of the synapse itself is actively compressed, so that the asynchronous pulse stream that arrives late due to bus jitter can achieve phase feedforward compensation before reaching the cell body, ensuring deterministic convergence of wavefront alignment. This phase calibration action converts the signal timing offset caused by the lag of the physical actuator into a phase compensation amount, ensuring that the pulse completes timing alignment before reaching the neuron cell body, generating a timing-aligned pulse feature matrix.

[0031] The neural computational network receives a pulse feature matrix. The membrane potential state of each neuron undergoes decay integral evolution according to the discrete time step. The membrane potential decay coefficient is set to 0.95. The decay integral evolution is based on the leakage integral firing physical model of the spiking neural network, simulating the natural leakage physical and electrical characteristics of the cell membrane and establishing a time-series charge integral loop. The membrane potential value in each control cycle is updated according to the following recursive rule: the membrane potential value of the previous moment is multiplied by the preset membrane potential decay coefficient to obtain the charge leakage residual value; the pulse output state of the preceding neuron is multiplied by the weight of the corresponding synapse to obtain the charge injection increment at the current moment; the charge leakage residual value and the charge injection increment are summed to output the membrane potential value at the current moment. The attenuation coefficient ranges from 0.85 to 0.99 within a closed interval. The pulse output state of the preceding neuron is a discrete binary variable of 0 or 1, used to characterize the presence or absence of pulses. During this period, if the target neuron triggers a pulse firing frequency of more than 3000Hz within a 1ms time window, it indicates that the neuron is in a state saturation stage caused by the superposition of high-frequency noise peaks. At this time, the charge discharge operator resets the membrane potential value of the neuron to 0mV resting potential and cuts off the update logic of the corresponding synaptic weights, making the synaptic weight update increment at the current moment zero. The output unit gathers the pulse distribution characteristics of the hidden layer, generates the calculated control parameter matrix through linear weight mapping, and transmits it to the downstream attitude drive solution unit.

[0032] Example 2: This example verifies the performance of a neural computation method for legged robot jumping training in unstructured complex terrain. A hybrid simulation test platform containing gravel, mud, and uneven slopes was constructed. The experiment simulates random load disturbances generated by foot contact reaction force on complex ground to evaluate the response logic of the spiking neural network to transient impacts. The experimental group applied the method of this invention, setting the sampling frequency of the foot contact pressure sensor to 1000Hz. The input unit applied first-order differential threshold gating logic to the contact reaction force data, setting the dynamic bias threshold to 2.5N. When the logic detects a change in value exceeding 2.5N between adjacent sampling times, it generates a discrete pulse signal to filter background disturbances caused by surface texture. The synaptic conduction delay constant is calibrated in real time by the controller based on the standard deviation of bus transmission jitter to ensure that the pulse flow completes phase alignment before reaching the neuron cell body. The charge discharge operator is set to a trigger upper limit of 3000Hz to distinguish between physical collision noise generated by robot landing and control signals, ensuring... To assess the stability of the asynchronous membrane potential transition process, control group 1 used a traditional pulse network with a fixed sampling frequency of 500Hz, while control group 2 used a conventional network without integrated charge discharge and weight truncation logic. It should be noted that the 15μs feature calculation response delay refers to the computation time consumed by a single pulse event driving network state update and performing a single forward propagation. The 1ms time window on which the pulse firing frequency depends belongs to a parallel-maintained sliding-time counting stream. Each neuron is equipped with a ring counting buffer array driven by the master clock. This array is independent of the forward propagation process and is dedicated to continuously accumulating the total number of pulse firings in the historical interval 1ms prior to the current moment at a refresh rate of 15μs. After the forward calculation is completed within 15μs, the monitoring unit obtains the resident value in the ring counting buffer array through a single-cycle read instruction and immediately performs an out-of-limit comparison. This ensures complete self-consistency between the microsecond-level response speed and the millisecond-level overall statistical window on the physical running time scale while blocking high-frequency collision noise.

[0033] Performance evaluation data are as follows: Feature computation response latency: 15μs for the experimental group, 5ms for control group 1, and 120μs for control group 2; State saturation rate under unstructured landing impact: 1.2% for the experimental group, feature information loss for control group 1, and a high 84.5% for control group 2 under impact overload; Energy consumption for single jump computation: 0.12J for the experimental group, 12.5J for control group 1, and 0.8J for control group 2; Terrain adaptive training success rate: 94.0% for the experimental group, 62.0% for control group 1, and 45.5% for control group 2. Experimental data indicate that when dense impulses are generated at the moment of landing in control group 2, the neurons... The membrane potential continuously accumulates and causes saturation. The weight update mechanism deadlocks due to overload. The experimental group triggers discharge within a 1ms window using a charge discharge operator to cut off weight iteration, keeping the output value of the calculated control parameter matrix in a stable range of 0.8 to 1.2. Further verification shows that when the charge discharge monitoring threshold is lowered to 1500Hz, the saturation rate drops to 0.8%, but the attitude calculation accuracy decreases by 12.0% due to high-frequency feature truncation. When it is raised to 4500Hz, the saturation rate rises to 5.5%, causing a lag in attitude calculation. The experiment confirms that the threshold setting of 3000Hz achieves a match between computational energy consumption and attitude calculation response accuracy.

[0034] Example 3: This example combines Figures 1 to 2 This paper describes a neural computation method for training legged robots to jump, such as... Figure 1 As shown, the steps are as follows: S1, obtaining the temporal characteristic data of foot contact reaction force and joint angular velocity state data during the jumping training process of the legged robot; S2, processing the temporal characteristic data of foot contact reaction force and joint angular velocity state data with a first-order differential threshold gating rule to generate an asynchronous pulse sequence; S3, obtaining the standard deviation of random temporal jitter of the data transmission channel and adjusting the conduction delay constant of each synaptic node according to the inverse proportional rule to calibrate the arrival time of each pulse in the asynchronous pulse sequence to generate a temporally aligned pulse feature matrix; S4, inputting the pulse feature matrix into the neural computing network and driving the membrane potential of each neuron in the neural computing network to undergo asynchronous transitions through pulse events, and triggering synaptic weight update actions according to the asynchronous transition state of the membrane potential of each neuron while monitoring the pulse firing frequency of each neuron within a 1ms time window; S5, resetting the membrane potential of the corresponding neuron in each neuron and truncating the synaptic weight update action at the current moment through a charge discharge rule when the pulse firing frequency exceeds the preset upper limit threshold of 3000Hz; and S6, converging the pulses output by the neural computing network to generate a control parameter matrix and outputting it.

[0035] like Figure 2As shown, the sensor sensing hardware peripherals include a foot contact pressure sensor and a joint angular velocity sensor, which are connected to a data acquisition module via sensor data lines to a signal acquisition and transmission bus. The data acquisition module transmits the acquired data to the operating environment of the central controller core computing platform, namely the neural computing network operating environment, via the main control bus, i.e., the asynchronous transmission bus. The central controller core computing platform also includes a storage area: a non-volatile storage area, and the storage area: the non-volatile storage area and the operating environment: the neural computing network operating environment are connected through an online adaptive bias calibration procedure. The storage area: the non-volatile storage area is equipped with a bias correction register and a zero-point offset register. The operating environment: the neural computing network operating environment integrates an input unit, a hidden layer, a monitoring unit, and an output unit. The output unit is connected to the downstream drive calculation unit of the power drive and actuator via a control parameter matrix transmission channel. The downstream drive calculation unit outputs drive control current and acts on the joint power component.

[0036] Example 4: To address the issue of sensor noise caused by transient impact during high-burst jump training of legged robots in unstructured terrain such as gravel, leading to computational model saturation and solution divergence, this invention provides a neural computational training scheme with targeted compensation. The system acquires foot-to-ground reaction force sensor data and joint angular velocity sensor data from the legged robot. In the data processing stage, the input unit calculates the rate of change of the ground reaction force data within adjacent time steps using first-order differential logic. A threshold of 2.5 N per millisecond is set for the rate of change. When the detected pressure change rate exceeds the threshold, the input unit determines that a transient interference induced by physical impact exists and triggers a response to this interference. The potential reset logic forcibly resets the membrane potential of neurons affected by the shock to a resting state, blocking the spread of noise spikes within the network. To eliminate timing deviations caused by high-frequency shock signals in neural computing networks, the computing architecture introduces a dynamic phase alignment mechanism for pulse streams. The controller collects data on the standard deviation of bus transmission jitter in real time and dynamically adjusts the endogenous conduction delay parameters of synaptic nodes accordingly. If an increase in the jitter amplitude of the transmission channel is detected, the conduction delay value of the synaptic node decreases proportionally. This phase calibration action ensures that even if there is bus timing instability, the pulse events received by each neuron remain within a preset time window, maintaining the semantic consistency of the pulse feature matrix.

[0037] When dealing with state saturation during training, the computational model integrates charge discharge rules. The monitoring unit continuously monitors the firing frequency of each neuron, setting an upper limit threshold of 3000Hz. When the monitored frequency reaches this threshold, it determines that the current processing load or transient noise has caused the computational unit to enter an overload state. It then blocks the data path of the synaptic weight update operator, resetting the current weight update increment to zero. This logic ensures that the computational control parameter tensor remains convergent in a non-stationary transition state. Specifically, to eliminate the mismatch in dimensions and dimensionality between multidimensional data structures and one-dimensional control operations, during boundary monitoring, the system subtracts the control parameter matrix of the previous pulse cycle from the control parameter matrix of the current pulse cycle to construct a difference matrix. The Frobenius norm of this difference matrix is ​​then calculated to extract the scalar rate of change, representing the overall fluctuation amplitude of the parameter matrix. This achieves mathematical dimensionality reduction from high-dimensional tensors to single-valued parameters. When this scalar rate of change exhibits monotonically increasing behavior over three consecutive pulse cycles and exceeds the safety boundary, the system invokes one-dimensional... The preset penalty step size is used as a normalization subtraction operator. The penalty step size value is subtracted synchronously from each corresponding matrix element of the control parameter matrix. This achieves absolute equivalence of physical dimensions and closed-loop control penalty while maintaining the anisotropic control characteristics of the multi-axis motion space. Under the condition of gravel road surface, by applying phase alignment and state saturation hindrance logic, the training system shows stable solution performance. Experimental data shows that during the landing transient, the original sensor pressure value suddenly increased from 12.4N to 15.2N. After processing by the rate of change judgment logic, the impact-induced characteristics were effectively isolated. The output value of the calculation control parameter tensor of the calculation model when dealing with non-stationary impacts is stable between 0.8 and 1.2, without any divergence. Tests show that if the charge discharge trigger threshold is lowered to 1500Hz, the loss of feature information leads to a 12.0% decrease in attitude solution accuracy. If it is raised to 4500Hz, the attitude solution lags due to the failure to suppress high-frequency noise impact in time. The threshold setting of 3000Hz achieves a match between the real-time response of attitude solution and anti-interference stability.

[0038] Example 5: When the legged robot training system faces state observation reference drift due to sensor aging or drastic changes in environmental temperature and humidity, the system activates the online adaptive bias calibration procedure, the robot enters a static calibration state, the control module reads the original output signal of the foot pressure sensor and subtracts it from the factory-preset zero-point reference potential. The difference is defined as the zero-point drift deviation. The controller writes the deviation into the bias correction register in the non-volatile storage area in real time. In subsequent jump training cycles, the negative number of the deviation is added to all collected pressure values ​​in real time, thereby ensuring that the pressure characteristic data input to the neural computing network is always anchored to the physical zero-point reference.

[0039] To ensure adaptability to unstructured ground disturbances during jumping, the system employs dynamic threshold reconstruction logic. During robot jumping training, the monitoring unit continuously analyzes the instantaneous fluctuation variance of the foot pressure sensor signals and updates the gating threshold in the first-order differential logic accordingly. This update logic is set such that if the ratio of the current pressure fluctuation variance to the historical training baseline variance is within the range of 0.8 to 1.2, the current gating threshold remains unchanged. If the ratio exceeds this range, the system initiates a threshold smoothing adjustment procedure based on the current road surface impedance feedback. This road surface impedance feedback data is obtained through online estimation using existing joint angular velocity sensors and foot ground contact pressure sensors within the system. The moment the foot contact sensor generates a pressure step signal, the central controller simultaneously reads the angular velocity change rate of the corresponding joint dynamic component and calculates the ratio of the real-time increment of the foot contact reaction force to the joint angular velocity change rate as the dynamic impedance characteristic value of the current terrain. This quantitatively characterizes the absorption and damping characteristics of gravel or soft soil on impact energy. When this dynamic impedance characteristic value crosses a preset threshold, the system initiates a smooth reconstruction of the first-order differential gating threshold. Thus, without adding additional external sensing hardware, the closed-loop acquisition and processing of road surface impedance data is completed. By introducing this discrete gradient adjustment mechanism based on environmental noise characteristics, false triggering pulses caused by changes in the softness of the environmental surface are suppressed while ensuring sensitivity to real ground impact.

[0040] In the neural computation network, the evolution of the membrane potential integral is synchronously integrated with an anti-aging smoothing operator. The system is equipped with a training cycle counter to count the cumulative number of jumps since the model started. When the cumulative number of jumps exceeds 5000, the network unit automatically adjusts the decay coefficient of the membrane potential integral from 0.95 to 0.97. This adjustment process is triggered based on the average firing activity of neurons during long-term training. By appropriately extending the membrane potential memory cycle, the shift in electrical signal distribution caused by long-term high-frequency updates of synaptic weights is offset, maintaining the stability of the neuron's feature expression at different training stages. This dynamic compensation mechanism addresses the aforementioned issue. Engineering verification shows that during 10,000 training cycles, the zero-point drift of the foot pressure signal was controlled within 0.2N through bias calibration and adaptive gating triggering logic, with no cumulative error divergence. In complex road training switching scenarios, the adaptive threshold reconstruction logic completed feature matching of road disturbances within 3 jump cycles, keeping the output fluctuation amplitude of the neural network's computational control parameter matrix within a stable range of 0.8 to 1.2. Data comparison proves that this calibration and verification process effectively decouples environmental noise and physical ground contact characteristics, achieving performance stability of the computational model under long-term training conditions.

[0041] Example 6: Before performing high-frequency continuous jumping actions, the current legged robot needs to be pre-calibrated for the zero-point deviation of the sensor group and environmental characteristics. The system keeps the robot in a static posture and collects static voltage data within 5 seconds through the foot contact pressure sensor. The average value is calculated and taken as the zero-point voltage reference. If the deviation between this reference and the factory preset value exceeds 0.05V, the system automatically records the difference to the zero-point offset register in the non-volatile storage area as a real-time compensation for subsequent sensor data readings. This calibration process ensures that the sensor input signal is completely aligned with the input reference zero point of the neural network under physical zero force conditions.

[0042] To ensure accurate state judgment of the system under different terrain environments, the training system deploys adaptive working condition monitoring logic. When the robot is running on a gravel road, the controller records the high-frequency noise statistical distribution of the sensor in real time and calculates the standard deviation of the signal envelope. If the standard deviation deviates from the flat ground training benchmark data by more than 20% within 50 consecutive jump cycles, the system determines that the current road condition has changed substantially. At this time, the calculation unit triggers the threshold update process, adjusting the first-order differential gating threshold of the pulse generation logic from 2.5N per millisecond to 3.2N per millisecond after correction based on the current road characteristics. This adjustment path is based on the quantitative evaluation of the unstructured characteristics of the ground to prevent background perturbations from inducing redundant pulse generation.

[0043] To address the potential neuronal membrane potential drift that may occur during long-term training of computational networks, the model maintenance unit establishes a dynamic compensation mechanism based on an integral decay coefficient. The system monitors the average pulse firing frequency of neurons and correlates it with the temporal phases of the training cycle. In the early stages of training, the model uses a first-stage decay coefficient to maintain a rapid response to high-burst movements. In actual engineering calibration, the first-stage decay coefficient is specifically set to 0.92. This lower decay value accelerates the discharge of neuronal membrane potential charge, enabling the network to sensitively capture the high-frequency discontinuous motion characteristics generated at the moment of physical landing during the initial jump phase. The preset critical value is specifically set at 3000 cumulative jumps. When the model surpasses this training node, the system automatically and smoothly adjusts the decay coefficient to the corresponding value of 0.97 in the second stage. By increasing the decay coefficient, the charge memory period is extended, offsetting the random electrical signal distribution shift caused by long-term high-frequency weight iteration. If the first-stage decay coefficient is lower than the engineering lower bound of 0.85, it will lead to excessive loss of energy from previous pulses. The fast-paced nature of the process prevents the formation of effective potential time-series integrals. If the attenuation coefficient in the second stage exceeds the engineering upper limit of 0.99, the membrane potential will remain in a deadlock saturation state for a long time, resulting in a loss of spatiotemporal resolution. The above-mentioned interval setting ensures the stability of long-term training behavior. As the number of training cycles reaches the preset critical value, the attenuation coefficient smoothly transitions to a higher value in the second stage, thereby suppressing the disturbance of cumulative charge leakage on synaptic weight updates. The aforementioned parameter calibration procedure was verified in field tests. Data records show that after zero-point offset calibration, the deviation of the foot pressure reading in the stationary state of the robot decreased from 0.8N to less than 0.1N. When the surface features change, the false trigger pulse rate is maintained below 1% after adaptive threshold reconstruction. In 5000 consecutive jump cycles, with the help of dynamic attenuation coefficient adjustment, the standard deviation of the calculated control parameter matrix output value is stably controlled within 0.05, eliminating the risk of model divergence under long-term operation. The test results show that the calibration procedure is engineering feasible in terms of ensuring computational stability.

[0044] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A neural computation method for training jumping in a legged robot, characterized in that, Includes the following steps: Step S1: Obtain the timing characteristic data of the foot contact reaction force and the joint angular velocity state data of the legged robot during the jumping training process; Step S2: The timing characteristic data of foot contact reaction force and joint angular velocity state data are processed using the first-order differential threshold gating rule to generate an asynchronous pulse sequence; Step S3: Obtain the standard deviation of random timing jitter of the data transmission channel, adjust the conduction delay constant of each synaptic node according to the inverse proportional rule to calibrate the arrival time of each pulse in the asynchronous pulse sequence, and generate a timing-aligned pulse feature matrix. Step S4: Input the pulse feature matrix into the neural computing network, drive the asynchronous transition of the membrane potential of each neuron in the neural computing network through pulse events, and trigger the synaptic weight update action according to the asynchronous transition state of the membrane potential of each neuron, while monitoring the pulse firing frequency of each neuron within a 1ms time window. Step S5: When the pulse firing frequency exceeds the preset upper limit threshold of 3000Hz, the membrane potential of the corresponding neuron in each neuron is reset by the charge discharge rule, and the synaptic weight update action at the current moment is cut off. Step S6: Converge the pulses output by the neural computing network, generate a control parameter matrix, and output it.

2. The neural computation method for jumping training of a legged robot according to claim 1, characterized in that, In step S2, the steps of processing the temporal feature data of foot contact reaction force and the joint angular velocity state data using the first-order differential threshold gating rule include: step S21, obtaining the current values ​​of the temporal feature data of foot contact reaction force and the joint angular velocity state data at the current sampling time; step S22, calculating the absolute value of the discrete difference between the current value and the previous value at the previous sampling time; step S23, comparing the absolute value of the discrete difference with a preset dynamic bias threshold, and generating a corresponding event pulse signal to form an asynchronous pulse sequence when the absolute value of the discrete difference exceeds the dynamic bias threshold.

3. The neural computation method for jumping training of a legged robot according to claim 1, characterized in that, In step S4, the membrane potential of each neuron in the neural computing network changes at a preset discrete time step. The membrane potential value of each control cycle is obtained by multiplying the value of the membrane potential at the previous moment by a preset membrane potential decay coefficient, and then adding the product of the pulse output state of the preceding neuron and the weight of the corresponding synapse.

4. The neural computation method for jumping training of a legged robot according to claim 1, characterized in that, In step S3, adjusting the conduction delay constant of each synaptic node according to the random timing jitter standard deviation through the inverse proportional rule includes the following steps: Step S31, making the conduction delay constant of each synaptic node decrease as the random timing jitter standard deviation increases, so as to calibrate the arrival time of each pulse in the asynchronous pulse sequence and generate a timing-aligned pulse feature matrix.

5. The neural computation method for jumping training of a legged robot according to claim 1, characterized in that, In step S5, the step of monitoring the pulse firing frequency of each neuron within a 1ms time window includes: step S51, counting the total number of pulse firings of each neuron within a 1ms time window; step S52, calculating the ratio of the total number of pulse firings to the length of the 1ms time window to obtain the pulse firing frequency, and activating the charge discharge rule when the pulse firing frequency exceeds a preset upper limit threshold.

6. The neural computation method for jumping training of a legged robot according to claim 5, characterized in that, In step S52, when the charge discharge rule is activated, the following steps are included: Step S521, the membrane potential of the target neuron whose pulse firing frequency exceeds the preset upper limit threshold is reset to the preset resting potential value; Step S522, the synaptic weight update action corresponding to the target neuron is stopped, and the synaptic weight update increment in the synaptic weight update action corresponding to the target neuron is set to zero.

7. The neural computation method for jumping training of a legged robot according to claim 1, characterized in that, In step S6, the step of generating the control parameter matrix further includes: step S61, performing boundary monitoring on the rate of change of the control parameter matrix through the rate of change constraint rule.

8. The neural computation method for jumping training of a legged robot according to claim 7, characterized in that, In step S61, the step of monitoring the boundary of the rate of change of the control parameter matrix through the rate of change constraint rule includes: step S611, monitoring whether the rate of change of the control parameter matrix shows a monotonically increasing trend within a series of preset pulse periods; step S612, when the rate of change shows a monotonically increasing trend within three consecutive pulse periods and exceeds the preset critical safety boundary, applying a step-by-step penalty to the rate of change of the control parameter matrix to reduce the rate of change of the control parameter matrix according to the preset penalty step size.

9. A neural computation method for jumping training of a legged robot according to claim 1, characterized in that, In step S6, after outputting the control parameter matrix, the following steps are also included: Step S62, the generated control parameter matrix is ​​transmitted to the downstream drive calculation unit, and the downstream drive calculation unit calculates and generates the drive control current for controlling the power components of each joint in the legged robot according to the control parameter matrix, so as to adjust the posture of the legged robot.

Images