A Deep Learning-Based Method and System for Precision Force Control of Robots

Abstract

This invention relates to the field of control systems and discloses a deep learning-based method and system for precise force control of robots. The method includes: acquiring a first control command from a deep learning model and performing low-pass filtering to generate a low-frequency control component; acquiring a target force value; acquiring a force error signal between a force sensor feedback signal and the target force value and performing high-pass filtering; generating a high-frequency control component based on the high-frequency error component using a high-bandwidth controller; real-time monitoring of the sign determination conflict conditions between the low-frequency and high-frequency control components; synthesizing the two control components to generate a final control command when the conflict condition is not triggered; and activating a safety degradation mode to prevent the low-frequency control component from contributing to the final control command when the conflict condition is triggered. This invention utilizes a frequency domain decoupling architecture to resolve the temporal contradiction between the inference delay of the deep learning model and the instantaneous response requirements of force control, and establishes an intrinsic safety mechanism to address the risk of model misjudgment.

Description

Technical Field

[0001] This invention relates to a deep learning-based method and system for precise force control of robots, belonging to the field of control system technology. Background Technology

[0002] In general control or regulation systems, the core objective of control law design is to achieve a balance between response speed and system stability. Traditional controllers, such as proportional-integral-derivative (PID) controllers, have a simple logic structure and extremely low computational burden. Their response speed is limited by the hardware sampling rate and can reach the microsecond level, making them suitable for applications requiring fast response. However, in the field of precision force control for robots, such as constant force grinding or precision assembly of complex curved surfaces, the physical contact between the robot end effector and the environment is a complex nonlinear process. Linear controllers like PID controllers often struggle to achieve ideal control accuracy and compliance when handling such complex nonlinear force control tasks.

[0003] To enhance the intelligence and nonlinear processing capabilities of control systems, the industry has begun exploring the introduction of deep learning models with nonlinear fitting capabilities into control loops to generate more precise control commands. However, this introduction introduces new and more fundamental constraints at the control system level: inference latency. Deep learning models are complex computational structures, requiring milliseconds to complete a single forward inference. This represents a difference of several orders of magnitude in objective time scales compared to the microsecond response speed of traditional controllers. While this time scale difference is acceptable in conventional, slow visual sorting tasks, it becomes a significant constraint in precision force control tasks. The essence of this is to handle the physical contact between the robot's end effector and the environment. This process is full of high-frequency events, such as the instantaneous impact of the initial contact, the microscopic stick-slip friction effect, or the sudden hardening of material surfaces. Existing technologies attempt to make the deep learning model, which has a millisecond-level inference delay, directly handle these high-frequency physical events that require microsecond-level instantaneous responses. This attempt has an inherent contradiction in the control principle. When a high-frequency impact occurs, the deep learning model, due to its delay, has missed the opportunity to respond in physical time. The delayed output may instead apply incorrect control force to the system, resulting in chatter, blade bounce, or excessive impact, causing the control system to become unstable.

[0004] In existing technologies, some solutions attempt to use deep learning models directly as the main body of the control loop. However, this cannot fundamentally avoid the aforementioned temporal contradictions. For example, Chinese invention patent CN113977429A discloses a robot constant force grinding system and grinding control method based on deep learning. In this solution, it attempts to input the real-time grinding force data collected by force sensors and the robot's real-time pose data into a deep neural network model. The model then directly analyzes and outputs the robot's pose adjustment strategy and the angle correction information of each axis joint. The fundamental flaw of this architecture is that it forces the response speed of the entire force control closed loop to be limited by the millisecond-level inference delay of the deep learning model. When faced with instantaneous physical impacts or high-frequency friction during the grinding process, the model cannot make an instantaneous response on a microsecond-level time scale. The delayed control commands will become a source of disturbance and are very likely to cause system oscillations. Directly using a slow intelligent model to process high-frequency physical events has an inherent contradiction in principle. Therefore, the technical problem to be solved by this invention is how to resolve the temporal contradiction between the inference latency of deep learning models and the instantaneous response requirements of force control from the perspective of control system architecture, so that the intelligent model with inference latency and the instantaneous response required by the system can work together in the force control loop to achieve stable and high-precision force control. Summary of the Invention

[0005] This invention provides a deep learning-based method and system for precise force control of robots. Its main purpose is to solve the problem that existing control architectures cannot separate the inference delay of deep learning models from the instantaneous response required for force control in the time domain, resulting in the two being unable to work together and causing system instability.

[0006] To achieve the above objectives, this invention provides a deep learning-based method for precise force control of robots, comprising the following steps:

[0007] Obtain the first control command generated by the deep learning model; apply the first control command to the low-pass filter to generate low-frequency control components;

[0008] Acquire the target force value; acquire the force error signal between the force sensor feedback signal and the target force value; apply the force error signal to a high-pass filter to generate a high-frequency error component; based on the high-frequency error component, generate a high-frequency control component through a high-bandwidth controller; monitor the running signs of the low-frequency control component and the high-frequency control component in real time;

[0009] Based on the operating symbol, it is determined whether the low-frequency control component and the high-frequency control component constitute a continuous sign-opposite state. If so, a conflict condition is triggered. In the operating state where the conflict condition is not triggered, the low-frequency control component and the high-frequency control component are added together to generate the final control command for driving the robot. When the conflict condition is triggered, a safety degradation mode is activated, in which the contribution of the low-frequency control component to the final control command is forcibly prevented.

[0010] Preferably, it also includes: real-time quantization of the signal energy of the high-frequency control component; establishing feedback adjustment logic, which generates an adjustment output based on the deviation of the signal energy from a preset nominal energy range; and using the adjustment output to adaptively adjust the cutoff frequency common to the low-pass filter and the high-pass filter.

[0011] Preferably, the step of real-time quantization of the signal energy of the high-frequency control component includes: calculating the high-frequency control component within a continuous short time window. root mean square value The calculation rules follow: ,in, The number of sampling points within the short time window; the root mean square value As signal energy, it is used for feedback regulation logic.

[0012] Preferably, the feedback regulation logic is a proportional-integral controller. The input of the proportional-integral controller is the error between the signal energy and the center value of the preset nominal energy range, and the output of the proportional-integral controller is the regulation output.

[0013] Preferably, the high-bandwidth controller is a proportional controller, where the high-frequency control component is proportional to the high-frequency error component.

[0014] Preferably, before applying the force error signal to the high-pass filter, the method further includes the step of passing the force error signal through a notch filter, wherein the center frequency of the notch filter is set to the known characteristic noise frequency of the robot system servo motor or tool spindle.

[0015] Preferably, it further includes: superimposing a probe sinusoidal signal of a specific frequency into the high-frequency control component; analyzing the force error signal through a bandpass filter, wherein the center frequency of the bandpass filter is consistent with the frequency of the probe sinusoidal signal; determining the contact state between the robot end effector and the workpiece in real time based on the output energy of the bandpass filter; and activating a safety degradation mode when the non-contact state is determined.

[0016] Preferably, the step of determining whether the low-frequency control component and the high-frequency control component constitute a continuous sign-opposite state includes: calculating the product of the low-frequency control component and the high-frequency control component at each time step; when the product is continuously negative within a preset time window, the conflict condition is triggered.

[0017] Preferably, the step of forcibly preventing the contribution of the low-frequency control component to the final control command when the conflict condition is triggered specifically includes: forcibly setting the value of the low-frequency control component to zero, and setting the final control command to be equal to the high-frequency control component, wherein the deep learning model is a model used to generate the force control feedforward command, and the first control command is the force control feedforward command.

[0018] A deep learning-based precision force control system for robots includes:

[0019] The first control command acquisition unit is configured to acquire the first control command generated by the deep learning model;

[0020] The low-pass filtering unit is configured to perform low-pass filtering on the first control command to generate a low-frequency control component; the force information processing unit is configured to acquire the force target value and the force sensor feedback signal, and calculate the force error signal between the force sensor feedback signal and the force target value.

[0021] The high-pass filter unit is configured to perform high-pass filtering on the force error signal to generate high-frequency error components;

[0022] The high-bandwidth control unit is configured to generate high-frequency control components based on high-frequency error components;

[0023] The conflict determination unit is configured to monitor the running symbols of the low-frequency control component and the high-frequency control component in real time, and determine whether the low-frequency control component and the high-frequency control component constitute a continuous opposite symbol state based on the running symbols to trigger the conflict condition.

[0024] The command generation unit is configured to, in the operating state where no conflict condition is triggered, add and synthesize low-frequency control components with high-frequency control components to generate the final control command for driving the robot; and, when a conflict condition is triggered, activate a safety degradation mode, in which the contribution of low-frequency control components to the final control command is forcibly prevented.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. By low-pass filtering the first control command from the deep learning model, it operates in the low-frequency domain of the control system, thus accommodating the inherent inference delay. At the same time, the force error signal is high-pass filtered, and a high-bandwidth controller is used to generate high-frequency control components. By synthesizing the low-frequency control components and the high-frequency control components, the control system retains the ability of the deep learning model to handle low-frequency tasks such as complex contours, while gaining the ability to respond to and suppress high-frequency physical events such as collisions and friction. This solves the inherent problem of the difficulty in balancing intelligence and high-frequency robustness in the control system.

[0027] 2. Based on the above architecture, a high-frequency control component is utilized. By quantizing the signal energy of this high-frequency control component in real time and using it as the input of the feedback adjustment logic, the cutoff frequency shared by the low-pass and high-pass filtering processes is adaptively adjusted. In this way, the high-frequency control component not only undertakes the control task of fast reflection, but its own energy magnitude is also used as a measure to characterize the high-frequency stability margin of the system. This allows the time domain boundary of the control architecture to automatically adapt to changes in physical characteristics such as workpiece stiffness, transforming from static setting to dynamic optimization, so that the system can maintain high-frequency stability under different operating conditions.

[0028] 3. This method also provides a mechanism for cross-validation using signals within the architecture. By monitoring in real time whether there is a continuous opposite state between low-frequency control components and high-frequency control components, it determines whether the slow thinking intention of the deep learning model fundamentally conflicts with the system's fast reflection protection. When a conflict is determined, the system will activate a safety degradation mode to actively prevent the low-frequency control components from contributing to the final control command. This approach utilizes the control system's own logic to achieve online supervision of the deep learning model's credibility, avoiding internal friction or damage caused by the system's continuous reckless actions when the model misjudges, and providing fault safety assurance for the control system. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the frequency domain decoupling and security degradation control architecture of the present invention;

[0030] Figure 2 The waveform diagram shows the response of the system of the present invention to instantaneously suppress high-stiffness contact impact.

[0031] Figure 3 This is the cutoff frequency of the present invention. Timing diagram of adaptive adjustment. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the scope of protection of the invention.

[0033] This invention provides a deep learning-based method and system for precise force control of robots. The system is designed at the control architecture level to resolve the temporal contradiction between the inference latency of the deep learning model and the instantaneous response required for precise force control. Functionally, the system can be divided into several cooperating units, including a first control command acquisition unit for acquiring instructions from the deep learning model, a low-pass filtering unit for processing these instructions, a force information processing unit and a high-pass filtering unit for acquiring and processing force feedback information, a high-bandwidth control unit for generating fast responses, a conflict determination unit for ensuring system safety, and a command generation unit for synthesizing final drive instructions. These units collectively construct a dual-channel parallel control architecture: a slow channel for handling low-frequency control tasks and a fast channel for handling high-frequency instantaneous disturbances. In this scheme, the deep learning model is configured as the source of force control feedforward commands. Its function is to generate a series of control commands offline or infer in real time for a predetermined complex contour or task. These commands constitute the first control command. The first control command acquisition unit is configured to periodically read the first control command from the model's data interface in a preset data frame format, using floating-point torque or speed correction values. The reading cycle is as follows: This corresponds to the inference delay time of the model. The subsequent low-pass filtering processing in this scheme is to operate on this data stream with a specific delay time.

[0034] In the slow channel, the first control command acquisition unit is configured to acquire the generated first control command from the deployed deep learning model. This first control command is typically a force control feedforward command for complex contour tracking or nonlinear torque compensation. Given the inherent millisecond-level inference latency of the deep learning model, the output signal contains both valuable low-frequency control intent and high-frequency distortion and noise caused by the latency. If these high-frequency components directly enter the control loop, they will interfere with system stability. Therefore, the first control command is passed to the low-pass filtering unit, which applies a low-pass filter to objectively filter out all frequencies above a certain cutoff frequency in the command. The components are selected to retain only the reliable low-frequency control intent, thus generating low-frequency control components. Cutoff frequency This is a key system calibration parameter. The initial setting is based on ensuring that this frequency is significantly lower than the Nyquist frequency corresponding to the inference latency of the deep learning model. If the stable inference cycle of the deep learning model is... (Right now ),but An exemplary initial calibration value can be set at to In between, only low-frequency signals that the model can reliably generate are allowed to pass through; in the fast lane, the force information processing unit acquires force feedback signals in real time from the force sensor integrated at the robot's end effector. Obtain the current force target value from the upper-level task planning. The unit then calculates the force error signal between the two. In a real industrial environment, this Inevitably, the signal will be coupled with known characteristic noise from the robot system's servo motors or tool spindle. This noise typically manifests as periodic interference at a specific frequency. To prevent this noise from interfering with the subsequent high-frequency response, this technical solution can selectively configure a preprocessing step before applying the force error signal to the high-pass filter. First, the force error signal is passed through at least one notch filter, whose center frequency is precisely set to the known characteristic noise frequency. For example, by performing spectral analysis, such as Fast Fourier Transform (FFT), on the force signal during the robot's unloaded operation, if it is found that... If a significant servo drive noise peak exists, then the center frequency should be configured as follows: The notch filter precisely removes this specific interference, resulting in a cleaner force error signal.

[0035] Subsequently, the purified or unpurified force error signal is applied to a high-pass filter unit for high-pass filtering. This high-pass filter extracts all high-frequency components of the force error, representing instantaneous disturbance signals from physical collisions, friction, or sudden hard point changes, thus generating high-frequency error components. The cutoff frequency used in this high-pass filtering process is set to be the same as the cutoff frequency used in the low-pass filtering unit. With identical or similar filters, this matched filter configuration ensures seamless integration of the two channels in the frequency domain, maximizing the overall response of the control system. Smooth transition in the vicinity; high-frequency error components The signal is then fed into a high-bandwidth control unit, which is implemented as a computationally simple controller with a response speed in the microsecond range. In a preferred embodiment, this high-bandwidth controller is a proportional controller (P controller), whose control logic follows... ,in For proportional gain, the parameters of the high-bandwidth control unit are, for example, the proportional gain when implemented as a proportional controller. The calibration procedure includes the following steps: First, the output of the low-pass filter unit... The force is forced to zero, causing the system to operate in the fast lane; secondly, the force target value is... While maintaining constant contact, a small, instantaneous force step disturbance with an amplitude of 5% of the target force is applied externally; then, the force error signal is observed and recorded. The response curve is then analyzed. Finally, the proportional gain is adjusted to ensure that the system response curve exhibits a critically damped state (i.e., the fastest response with no overshoot) or a slightly underdamped state, with the first overshoot being less than 10% of the step amplitude. The value is set and locked as the system's working gain. The value can be determined through standard calibration procedures. With the channel closed, gradually increase The gain is adjusted until the system begins to exhibit high-frequency chatter when subjected to slight disturbances, at which point a slightly lower gain value is taken. The critical gain is multiplied by a factor of 1 as the operating gain; this high-bandwidth control unit ultimately outputs a high-frequency control component. This component serves as the system's fast response part and is specifically designed for instantaneous suppression of high-frequency disturbances.

[0036] The command generation unit is configured to combine the outputs of the two channels mentioned above. Under normal system operation, when no conflict condition is triggered, this unit adds the low-frequency control component and the high-frequency control component to generate the final control command. ,Should The signal is sent to the robot's servo drive, enabling the robot's movement to simultaneously possess the low-frequency control capability of a deep learning model and the high-frequency response capability of high-bandwidth control. To address potential misjudgments by the deep learning model, the system also establishes a built-in safety degradation mechanism. The conflict resolution unit is configured to monitor low-frequency control components in real time. With high-frequency control components The unit determines whether a continuous sign-opposite state exists between the two based on the operation symbols. This indicates that a continuous conflict exists between the deep learning model representing low-frequency control intent (if it continues to advance) and the rapid response representing physical protection (if it immediately withdraws). In one specific determination method, the determination is executed by calculating... and For the product at each time step, apply a short-time window to the product sequence. The moving average or integral is used to determine if a conflict condition is triggered when the average product within the time window is continuously negative and its absolute value exceeds a certain preset conflict threshold. The calibration principle is that the duration should be significantly longer than the transient response time of the system during normal contact transition, such as... To avoid false triggering, but must be short enough to intervene promptly in the event of sustained conflict; when a conflict condition is triggered, the command generation unit will immediately activate the security degradation mode, in which the contribution of low-frequency control components to the final control command is forcibly prevented. Specifically, this can be achieved by reducing the low-frequency control components... The value is forcibly set to zero, and the final control command is set to At this point, the system is downgraded to a pure high-frequency damper, relying on rapid response to ensure physical safety.

[0037] Furthermore, the present invention also provides an adaptively adjustable cutoff frequency. This mechanism addresses the challenges posed by variations in workpiece stiffness; this mechanism can utilize... The signal energy serves as a characterization of the system's high-frequency stability margin; the system is configured to quantize high-frequency control components in real time. The signal energy; in one quantization method, this step includes a continuous short time window such as Within each sampling point, calculate root mean square value The calculation rules follow ,in, The number of sampling points within the short time window. For the first The high-frequency control component at each sampling time; and the As signal energy; the system further establishes feedback regulation logic, which is based on signal energy. Compared to the preset nominal energy value representing slight activation, such as Deviations generate adjustment outputs; in one type of adjustment logic, this feedback adjustment logic operates on a slower time scale, such as per A proportional-integral (PI) controller is updated once, with the energy error as the input. Its output is the regulated output. The system utilizes this adjustment output Adaptively adjust the cutoff frequency shared by the low-pass and high-pass filters. The adjustment logic is that when the system tends to oscillate, it encounters a high-rigidity workpiece. Increase When the value is negative, the PI controller outputs a negative adjustment value, making... Automatically lowered, which makes The bandwidth of the channel is narrowed, thus suppressing oscillations; conversely, when the system response is insufficient, such as when contacting a flexible workpiece, Decrease When positive, the PI controller makes Automatic height adjustment, making The channel is more sensitive, and this method allows the frequency domain boundary of the system to be dynamically adjusted.

[0038] The present invention may also include a mechanism for detecting the contact state between the robot end effector and the workpiece, preventing loss of connection when the robot end effector accidentally separates from the workpiece. This component causes the robot to lunge forward uncontrollably and trigger a second collision; the mechanism is achieved by superimposing a specific frequency, such as..., into the high-frequency control component. The system detects a weak sinusoidal signal; it then analyzes the force error signal using a bandpass filter (BPF). The center frequency of the bandpass filter is precisely set to the frequency of the probed sinusoidal signal. Consistent; the system determines the contact state in real time based on the output energy of the bandpass filter: when the end effector is in contact with the workpiece, the high damping of the workpiece absorbs the energy of the detection signal, resulting in very low BPF output energy; when the end effector is disconnected, the end effector vibrates under the excitation of a weak detection signal, causing the BPF output energy to be significantly higher than the preset contact threshold; once the system determines that the contact state is non-contact, it will immediately activate the safety degradation mode and force the connection to be closed. Zeroing prevents loss of connection and forward acceleration; a deep learning-based robot precision force control system is used to implement any of the above methods. The system includes a first control command acquisition unit, a low-pass filtering unit, a force information processing unit, a high-pass filtering unit, a high-bandwidth control unit, a conflict determination unit, and a command generation unit. The functions of these units have been described in detail above. In engineering implementation, these units are logical functional modules that can be implemented on one or more digital processors, such as the real-time central processing unit (CPU), field-programmable gate array (FPGA), or dedicated digital signal processor (DSP) built into the robot controller.

[0039] Example 1: This example demonstrates the operation of the general technical solution described in a specific industrial scenario. The scenario involves using a robot integrated with a deep learning model to perform precision force-controlled grinding on aero-engine turbine blades with complex curved surfaces, material hard points, and internal holes. The goal is to maintain a constant normal contact force throughout the grinding process. In the initial stage of the grinding task, the robot's end effector contacts the gently curving surface of the blade. At this point, the slow-speed channel of the control system plays a major role. The deep learning model instructions acquired by the first control command acquisition unit are processed by the low-pass filter unit to generate smooth low-frequency control components. This component, based on a deep learning model's understanding of the complex aerodynamic shape of the blade, dominates the robot's low-frequency tracking of target forces on a large scale; force error signals in the fast channel. Keeping the value low, the high-frequency components are close to zero, therefore the high-bandwidth control unit generates... It is also close to zero, and the final control command Approximately equal to The robot exhibits compliant and highly precise low-frequency tracking characteristics; when the grinding tool reaches microscopic material hard points on the blade surface, such as welding or repair areas, high-frequency physical events occur instantaneously, and the force sensor provides feedback. The signal generates high-frequency pulse impacts, leading to The signal changes drastically; because the frequency of this change is much higher than the cutoff frequency. It is completely extracted by the high-pass filter unit, enabling the high-bandwidth control unit (in this case, the proportional controller) to instantaneously output a high-frequency control component with a significant amplitude opposite to the direction of the impact. At this moment, the deep learning model, with its millisecond-level latency, has not yet reacted to the shock, and its output... It remains stable, but eventually control commands... Already because The instantaneous intervention produces an immediate suppression effect, which occurs on a microsecond timescale, thus physically preventing the formation of excessive impact and avoiding tool vibration or scratches on the blade surface.

[0040] Next, as the robot grinded to the edge of the blade and crossed the internal holes used for cooling, a disconnection and re-contact condition occurred. Force feedback occurred the instant the tool left the blade surface. It becomes zero. At this point, if only relying on... The channel (which will still output propulsion commands due to the delay) will cause the robot to rush forward at an uncontrolled speed and violently crash into the other edge of the hole; however, in this solution, as in the specific implementation, the contact state detection mechanism is activated, superimposed on... On the detection sinusoidal signal At the moment the tool loses contact, the loss of physical damping from the workpiece causes vibration at the end. This vibration signal is detected by a parallel bandpass filter (BPF), causing the output energy to momentarily exceed the contact threshold. Based on this, the system determines a non-contact state and immediately triggers a safety degradation mode. Forced to zero, thus enabling the final control command It contains only high-frequency components (mainly the detection signal at this point), which eliminates... The robot, initially exhibiting a loss of contact and forward momentum, smoothly traversed the hole with a controlled posture, achieving safe re-contact. In the final stage of the task, the robot ground down to an unexpected rigid mounting base at the blade root, creating a hard stop. At this point, the deep learning model, lacking training data for this scenario, might incorrectly classify the obstacle as a passable, conventional load and continuously output the first control command demanding continued propulsion, leading to... The value remains positive; physical contact leads to The high-frequency control component remains consistently large negative values. Correspondingly, it becomes a negative value that is continuously used for pullback; at this point, the conflict determination unit starts working, by monitoring... and The running symbols were found to be in use beyond the preset time window. For example The system remains in a state of conflict with opposite signs for a specified period of time; the system determines that the conflict condition has been triggered, reactivates the security degradation mode, and forcibly switches to a state of conflict. The contribution is reset to zero, and the final control command is executed. This allows the robot to stop its continuous propulsion and maintain a safe damping state at the contact point using a high-frequency channel, thereby protecting the robot's motor and blades from continuous overload damage caused by model misjudgment. This control architecture, which decouples control commands in the frequency domain and uses the internal signals of the two channels for cross-validation such as conflict determination and active detection such as contact state determination, enables the system to retain the ability of deep learning models to handle low-frequency tasks such as complex contours, while independently gaining the ability to respond to and suppress high-frequency physical events such as collisions and disconnections. Furthermore, it provides a mechanism to avoid the risk of model misjudgment through built-in safety degradation logic.

[0041] Example 2: To objectively verify the stability and high-frequency disturbance suppression capability of the frequency domain decoupling control architecture disclosed in this invention in high-stiffness contact tasks, and to examine the operation of the conflict determination and safety degradation mechanisms, a robot force control test platform was built. This platform includes a six-axis industrial robot with a six-dimensional force sensor installed on its end flange. The sampling frequency of this sensor is [missing information]. The robot controller runs under a real-time operating system, and the control cycle is set to... The test object was a high-rigidity steel workpiece fixed on a base, used to simulate hard contact conditions encountered during precision assembly or grinding; a control group and the sample group of this invention were set up for comparative testing. The control group used a method that tested the deep learning model and found that the average inference latency was... By placing it directly in the force control loop, the final control command is the output of the deep learning model. The prototype of this invention adopts a dual-channel frequency domain decoupling architecture, which decouples the first control command of the deep learning model. Generated by low-pass filter The force error signal After being processed by a high-pass filter, the signal is sent to a high-bandwidth controller, which is configured here as a proportional controller for generation. Final control command To ensure effective separation in the frequency domain and avoid the latency effects of deep learning models, the two filters share a common cutoff frequency. Labeled as The proportional gain of the high-bandwidth controller Offline calibration This value was determined after a trade-off between ensuring the speed and stability of the high-frequency channel response.

[0042] Experiment 1: High stiffness contact stability test; Target contact force was set for both sets of experiments. for Command robot from A constant external velocity is used to contact a steel workpiece, and the results after contact are recorded. The force signal within; in the control group, the force sensor feedback when the robot end effector contacts the workpiece. The signal exhibits violent oscillations and overshoot due to the deep learning model. The inference delay and lagging control commands cannot suppress microsecond-level contact impacts, imposing hysteresis compensation on the system and exacerbating system instability at the contact point. In the prototype of this invention, at the instant of contact, the high-frequency impact force is captured by the high-pass filter unit, and the high-bandwidth controller... The control cycle instantaneously generated a direction opposite to the control cycle. Component, the component in The impact was suppressed before any reaction was made; the key performance indicators of the two control methods are compared as follows, see Table 1.

[0043] Table 1: Comparison of Contact Stability Tests

[0044]

[0045] Referring to Table 1, the experimental data show that the control group oscillated under rigid contact, with a maximum overshoot. More than twice the target force, in The internal convergence is impossible; however, the sample group of this invention uses high-frequency control components. Instantaneous intervention controls overshoot within ,exist Converging to steady state, steady-state oscillation amplitude (standard deviation) The latency was significantly reduced compared to the control group, indicating that the architecture of this invention helps to resolve the contradiction between high-frequency physical events and model latency.

[0046] Experiment 2: Model Misjudgment and Security Degradation Test. Building upon Experiment 1, the two systems were first stabilized at... The contact force, and then At any moment, a human-simulated deep learning model makes a misjudgment, and the output intention is to increase the force. In the control group, the system executed an erroneous instruction from the deep learning model, causing the robot to lunge forward with significant acceleration, resulting in contact force... Internal surge to This caused a physical impact on the workpiece and sensor; in the prototype of this invention, after receiving this erroneous command, the result was low-pass filtered... The component begins to increase slowly, attempting to propel the vehicle forward; however, this results in force errors. A sudden increase, leading to The component also instantly generates a pull-back command; the conflict determination unit (continuous confrontation time window) Set as )exist Detected at all times and The persistent sign-opposite state immediately triggered a conflict condition; the system then activated the security degradation mode, which... Force zeroing, only retain The damping effect ultimately controls the command. The robot stopped its forward movement; the maximum contact force during this process is recorded below, see Table 2.

[0047] Table 2: Comparison Table of Security Degradation Tests

[0048]

[0049] Experimental data shows that the frequency domain decoupling control architecture disclosed in this invention helps maintain system stability in high-stiffness contact scenarios, as shown in Experiment 1; simultaneously, as shown in Experiment 2, the built-in conflict determination and safety degradation mechanisms can respond to and block control commands caused by misjudgments in the simulation model, suppressing peak impact forces within... This provides a security mechanism for the system while maintaining control performance.

[0050] Example 3: This example combines Figures 1 to 3 This document describes a deep learning-based method and system for precise force control of robots. Figure 1 As shown, in this architecture, the first control command force control feedforward command generated by the deep learning model is acquired by the first control command acquisition unit and sent to the low-pass filter unit to generate low-frequency control components. Meanwhile, the target force value and the force sensor feedback are used by the force information processing unit to calculate the force error signal. ,Should The signal is fed into a high-pass filter unit to generate a high-frequency error component, which is then passed through a high-bandwidth control unit to generate a high-frequency control component. The low-pass filter unit and the high-pass filter unit share the cutoff frequency. , Quantity and The components are sent to the conflict determination unit, which monitors the running symbols of both components in real time to determine if a conflict exists. Based on the result of the conflict determination unit, the command generation unit, under normal conditions, will... and Synthesized In a conflict state, a safety degradation mechanism is activated, and the resulting control commands are sent to the robot actuators. Furthermore, the architecture includes feedback regulation logic, which quantizes... The energy, adaptively adjusting the cutoff frequency. .

[0051] like Figure 2As shown in the figure, the horizontal axis represents time in milliseconds (ms), and the vertical axis represents the control quantity or force value. The figure displays four curves: actual contact force (solid line), low-frequency control component, and so on. (Dashed line) High-frequency control component (Dots and lines) and the final control commands (Dotted line) When the actual contact force experiences an instantaneous impact and overshoots at approximately 90ms, the high-frequency control component... It immediately produces an instantaneous suppression effect in the opposite direction, while the low-frequency control component This results in a delayed response, and the final control command... As and The synthesis of the initial response mainly consists of This dominant approach enabled rapid suppression of high-frequency shocks and ultimately led to system convergence; such as Figure 3 As shown, the process begins with the high-frequency control component output by the high-bandwidth controller. The energy quantization module receives this component and calculates its root mean square value. As signal energy, the PI controller receives this signal energy and compares it with the nominal energy value to calculate the energy error. This process then enters a conditional decision branch: if the energy is too high (for high-stiffness workpieces), indicating that the system tends to oscillate, a negative adjustment is generated to lower the cutoff frequency. To suppress oscillations, the high-frequency channel bandwidth is narrowed. If the energy is too low (in the case of low-stiffness workpieces), it indicates insufficient system response, so a positive adjustment is generated to increase the cutoff frequency. To improve sensitivity, the high-frequency channel bandwidth is widened. If the energy is within the nominal range, the current setting is maintained, and the cutoff frequency is kept constant. Cutoff frequency The adjustments are applied to both the low-pass and high-pass filter units to achieve dynamic optimization of the frequency domain boundaries.

[0052] Example 4: To further illustrate the technical effect of the frequency domain decoupling architecture adopted in this invention compared with the prior art, the following comparative experiment was conducted; this comparative experiment used the same test platform and parameter settings as in Example 2, and a six-axis industrial robot was equipped with A six-dimensional force sensor with a sampling rate and a control cycle. The test object was a high-rigidity steel workpiece, and the target contact force was... for The key difference between this comparative experiment and the sample group in Example 2 is that this comparative experiment does not use a frequency domain decoupling architecture, but directly sets the average inference delay to... deep learning model output As the final control command, this approach corresponds to the technical path described in the background section that attempts to allow deep learning models with inference latency to directly process force control closed loops; when performing the same high-stiffness contact task as in Example 2, the robot is instructed to... The robot end effector contacts the workpiece at a constant speed, and the force signal after contact is recorded. The force sensor feedback signal is observed when the robot end effector contacts the workpiece. Oscillations occurred immediately, with the recorded peak overshoot reaching [value missing]. Exceeding the target value Due to the lack of a high-frequency response channel for instantaneous suppression of the initial shock, deep learning models... The delay makes it impossible for it to Effective compensation is provided for high-frequency force fluctuations occurring within the control cycle. The lagging output causes continuous interference to the control loop, resulting in persistent high-frequency chattering near the contact point. During the observation period, the contact force failed to stabilize near the target value, and the calculated standard deviation of the oscillation was... This phenomenon indicates that, without employing a frequency domain decoupling mechanism, directly applying a deep learning model with inference delay to a high-stiffness force-controlled closed loop will lead to system instability due to time scale mismatch.

[0053] Experimental results show that after removing the frequency domain decoupling architecture of the present invention, namely, low-pass filtering to process the first control command, high-pass filtering to process the force error signal, and combining it with a high-bandwidth controller to generate high-frequency control components, the method of force control relying on a deep learning model cannot maintain system stability in high-stiffness contact scenarios, resulting in force overshoot and continuous oscillation. This contrasts with the stable control effect obtained by using the architecture of the present invention in Example 2, illustrating the role of the frequency domain decoupling architecture in resolving the contradiction between the latency of the deep learning model and the instantaneous response requirements of force control.

[0054] Example 5: This example illustrates the adaptive adjustment of the common cutoff frequency of the low-pass and high-pass filters. The specific calibration and operating procedures address situations where the physical characteristics of the workpiece change during precision force control tasks in robots, taking stiffness as an example. In the specific calibration process, the robot system is first configured to perform a constant force contact task on a standard test workpiece with specific stiffness, such as a medium-hardness aluminum alloy plate. At this time, the adaptive tuning logic is temporarily disabled, and the cutoff frequency... Manually set on a series of discrete values, for example from Increment to Step length In each Under the setting, the system runs for a preset period of time, for example... Record high-frequency control components The time-domain waveform; by calculating the root mean square value of each waveform segment. Based on the assessment of contact force stability, such as the standard deviation of force error, a [specific value] is determined. Values, for example and its corresponding Root mean square value of signal energy This setting is the result of a trade-off between high-frequency disturbance suppression and system stability; a narrow band range is defined around this energy value, for example... The target nominal energy range of the feedback regulation logic has a central value of . This calibration process establishes a baseline for adaptive adjustment. When the system is operating close to the calibration state, The energy should be maintained within the nominal range; after calibration, the adaptive tuning logic is activated, which uses a proportional-integral (PI) controller to achieve [the desired energy level]. Slow adjustment; the input to the PI controller is set to the signal energy calculated in real time. With respect to the center value of the nominal energy range Error between ; Output of PI controller Used to adjust the current cutoff frequency The control law can be ,in For discrete time steps, For the update cycle of the PI controller, for example , and These are the proportional and integral gains, respectively. These two gain parameters can be set on a slower time scale by applying a small-range step disturbance to the calibrated workpiece and using control system parameter tuning methods to achieve a stable result. Adjust the response, for example, set , At the same time Set upper and lower limit saturation constraints, for example Updated cutoff frequency ,in The function represents a limit operation.

[0055] Imagine a robot performing a task that transitions from contacting a calibration aluminum alloy plate (of medium stiffness) to contacting a high-hardness steel plate. Due to the increased contact stiffness, the system's response to small disturbances becomes more drastic, leading to more volatile high-frequency control components. energy The energy level increased, exceeding the upper limit of the nominal energy range, for example, reaching... At this point, the energy error... The value is negative; the PI controller calculates a negative adjustment amount based on the negative error input. Make the cutoff frequency Automatically lowered; during the update cycle, if Maintain at Nearby, This will result in an integral term containing negative proportional terms and negative cumulative terms, leading to Shift to lower frequencies, for example from Reduce to A lower cutoff frequency means a correspondingly lower frequency at which the high-pass filter begins to operate, which reduces the frequency at which the high-frequency control channel... The response to high-frequency disturbances becomes less sensitive, suppressing the tendency of the system to oscillate due to increased stiffness. The system returns to near its nominal range; if the robot switches from contacting an aluminum alloy plate to contacting a flexible rubber pad, the high-frequency vibration energy of the system weakens due to the reduced contact stiffness. It will fall below the lower limit of the nominal range, for example, drop to... At this point, the energy error... It is a positive value; the PI controller outputs a positive adjustment amount. Make the cutoff frequency Automatically increased, for example from Rise to The increase in cutoff frequency makes the high-frequency control channel... The increased bandwidth allows for a faster response to potential viscous or springback phenomena under low stiffness, helping to maintain the system's dynamic performance under different contact conditions. This adaptive cutoff frequency tuning procedure based on high-frequency control component energy feedback, through closed-loop adjustment, [further enhances the performance]. By maintaining the frequency domain segmentation point within the nominal range, the control architecture can automatically adapt to changes in physical properties such as workpiece stiffness, thus maintaining the high-frequency stability of the system under different working conditions and eliminating the need for manual parameter tuning for different workpieces.

[0056] Example 6: This example illustrates the specific steps for configuring the notch filter and setting the collision determination parameters, as part of a standardized pre-calibration procedure before deploying the control system of this invention to a specific robot and task. First, system characteristic noise is identified and the notch filter is configured. The robot system and end effector are then operated within the predetermined speed range for the task without contacting the workpiece, while simultaneously using a high sampling rate. Record the raw signal data from the force sensor, and the acquisition duration is set to [specify duration]. Fast Fourier Transform (FFT) was applied to the collected force signal data in various dimensions for spectral analysis to identify specific frequency points in the spectrum that exhibit energy peaks caused by servo motor drive or tool rotation. If, in the spectral analysis results, [the following is true / false] ... A stable narrowband peak was found at [location]. If another peak is detected, two notch filters are configured accordingly, with their center frequencies set to [values ​​to be filled in]. and And set the narrowband width, set to It filters out noise while reducing the impact on effective signals in adjacent frequency bands; after configuration, the notch filter is applied to the processing path before the force error signal enters the high-pass filter.

[0057] Finally, the collision determination parameters were calibrated; a series of test sequences were designed, including typical operations such as start-up, stop, speed change, normal contact and separation; the sequences were run and the results were recorded. and Data; analyze the data to find For intervals with negative values, calculate the maximum duration of these intervals. and the peak value of the absolute value of the product during this period The record shows that the sign-opposite state lasts the longest during a fast contact transition. Then, simulated controlled adversarial scenarios, such as instructing the robot to operate at low speeds. When the value is small and positive, it contacts the rigid limiting block, and the physical reaction force causes... Generates a continuous series of negative values; record this stable adversarial state. product mean Based on the measurement results, a time window for conflict determination is set. The duration is greater than the maximum duration of a normal transient process, which is set here. Conflict threshold Set to a value between the transient peak and the mean of the stable adversarial state; here it is set to... Confirm that the threshold is greater than ; calibrated and The parameters are stored as system parameters for the conflict determination unit to use in real time. By executing the above-mentioned standardized calibration procedures for notch filters and conflict determination parameters, and configuring the control system according to the specific robot hardware platform and application environment, it helps to improve the system's ability to suppress inherent noise, so that the safety degradation mechanism can be triggered when abnormal confrontation occurs, and avoids misjudgment in normal transient operations.

[0058] 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.

[0059] 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 deep learning-based method for precise force control of robots, characterized in that, Includes the following steps: Obtain the first control command generated by the deep learning model; apply the first control command to the low-pass filter to generate low-frequency control components; Acquire the target force value; acquire the force error signal between the force sensor feedback signal and the target force value; The force error signal is applied to a high-pass filter to generate high-frequency error components; High-frequency control components are generated by a high-bandwidth controller based on high-frequency error components. Real-time monitoring of the operating symbols of low-frequency and high-frequency control components; Based on the running symbol, determine whether the low-frequency control component and the high-frequency control component constitute a continuous symbol opposite state, and if so, trigger the conflict condition. In the operating state where the conflict condition is not triggered, the low-frequency control component and the high-frequency control component are added together to generate the final control command used to drive the robot. And when a conflict condition is triggered, a security degradation mode is activated, in which the contribution of low-frequency control components to the final control command is forcibly prevented. In addition, before applying the force error signal to the high-pass filter, the method further includes the step of passing the force error signal through a notch filter, wherein the center frequency of the notch filter is set to the known characteristic noise frequency of the robot system servo motor or tool spindle. The method also includes: superimposing a specific frequency detection sinusoidal signal into the high-frequency control component; analyzing the force error signal through a bandpass filter, wherein the center frequency of the bandpass filter is consistent with the frequency of the detection sinusoidal signal; determining the contact state between the robot end effector and the workpiece in real time based on the output energy of the bandpass filter; and activating a safety degradation mode when the non-contact state is determined. The step of determining whether a continuous sign-opposite state exists between a low-frequency control component and a high-frequency control component includes: calculating the product of the low-frequency control component and the high-frequency control component at each time step; when the product is continuously negative within a preset time window, a conflict condition is triggered.

2. The deep learning-based precision force control method for robots according to claim 1, characterized in that, Also includes: Real-time quantization of the signal energy of high-frequency control components; A feedback regulation logic is established, which generates a regulation output based on the deviation of the signal energy from the preset nominal energy range. It also adaptively adjusts the cutoff frequency shared by the low-pass and high-pass filters by adjusting the output.

3. The deep learning-based precision force control method for robots according to claim 2, characterized in that, The steps for real-time quantization of the signal energy of the high-frequency control component include: calculating the high-frequency control component within a continuous short time window. root mean square value The calculation rules follow: Where N is the number of sampling points within the short time window; the root mean square value As signal energy, it is used for feedback regulation logic.

4. The deep learning-based precision force control method for robots according to claim 2, characterized in that, The feedback regulation logic is a proportional-integral controller. The input of the proportional-integral controller is the error between the signal energy and the center value of the preset nominal energy range, and the output of the proportional-integral controller is the regulation output.

5. The deep learning-based precision force control method for robots according to claim 1, characterized in that, The high-bandwidth controller is a proportional controller, where the high-frequency control component is proportional to the high-frequency error component.

6. The deep learning-based precision force control method for robots according to claim 1, characterized in that, The steps for forcibly preventing the contribution of low-frequency control components to the final control command when a conflict condition is triggered specifically include: forcibly setting the value of the low-frequency control component to zero, and setting the final control command to be equal to the high-frequency control component. The deep learning model is a model used to generate force control feedforward commands, and the first control command is a force control feedforward command.

7. A deep learning-based precision force control system for robots, used to implement the deep learning-based precision force control method for robots as described in claim 1, characterized in that, include: The first control command acquisition unit is configured to acquire the first control command generated by the deep learning model; The low-pass filter unit is configured to perform low-pass filtering on the first control command to generate low-frequency control components. The force information processing unit is configured to acquire the target force value and the feedback signal from the force sensor, and to calculate the force error signal between the force sensor feedback signal and the target force value. The high-pass filter unit is configured to perform high-pass filtering on the force error signal to generate high-frequency error components; The high-bandwidth control unit is configured to generate high-frequency control components based on high-frequency error components; The conflict determination unit is configured to monitor the running symbols of the low-frequency control component and the high-frequency control component in real time, and determine whether the low-frequency control component and the high-frequency control component constitute a continuous opposite symbol state based on the running symbols to trigger the conflict condition. The command generation unit is configured to add and synthesize low-frequency control components and high-frequency control components to generate the final control command for driving the robot in the operating state where the conflict condition is not triggered. And when a conflict condition is triggered, a safety degradation mode is activated, which forcibly prevents low-frequency control components from contributing to the final control command.

Images