A robot control method and system based on surface bioelectric signals, a terminal device, and a medium

Abstract

The application discloses a robot control method and system based on surface bioelectric signals, a terminal device and a medium, and relates to the technical field of robot control. The method comprises the following steps: collecting multi-channel surface bioelectric signals of an operator to generate a continuous time sequence signal stream; extracting a time sequence signal sequence of a preset historical time step and capturing a time sequence dependency relationship between signals to obtain a time sequence feature representing an action evolution trend of the operator, and inputting a pre-trained conditional generation model to generate a parameterized action primitive sequence containing future preset step number action primitives in a continuous action space, wherein the action primitive is a minimum execution unit of a robot containing a continuous adjustable parameter; selecting a preset number of action primitives to be converted into a control instruction for execution; repeatedly generating an updated sequence to replace an unexecuted part to realize continuous adaptive control of the robot. The application realizes smooth and continuous control of the robot without gesture labels or fixed scripts, and improves the adaptability to changes in the operator's intention.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a robot control method, system, terminal device and medium based on surface bioelectric signals. Background Technology

[0002] With the rapid development of human-robot collaboration, rehabilitation assistance, teleoperation, and dexterity technologies, utilizing natural, low-latency, and continuously adjustable human physiological signals for robot control has become an important research direction in the field of robot control. Surface bioelectrical signals can directly reflect muscle activation state and movement intention, and have significant advantages such as non-invasiveness, wearability, and fast response, making them ideal input signals for robot motion generation and execution control.

[0003] Existing robot control schemes based on surface bioelectrical signals mainly fall into three categories: discrete gesture recognition triggering methods, which output limited category labels, struggle to express motion amplitude and duration, and are prone to abrupt motion changes; single-step continuous regression methods, which map control quantities based solely on the current signal, lack the ability to model future motion evolution, and are susceptible to muscle fatigue, electrode misalignment, and environmental noise; and fixed-script task execution methods, which cannot correct trajectories online based on real-time force changes, resulting in poor environmental adaptability. Furthermore, existing technologies generally lack a unified parametric motion representation between high-level task intent and low-level execution control.

[0004] Therefore, there is an urgent need for a robot control method based on the generation of parameterized action primitive sequences and the execution of rolling motion based on surface bioelectric signals, in order to fill the gap in the existing technology. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, in the field of robot control based on surface bioelectric signals, existing technologies cannot simultaneously achieve continuous and smooth control, future action evolution modeling, and real-time environmental adaptability within the same architecture. Discrete triggering methods result in discontinuous control, single-step regression lacks sequence modeling capabilities, fixed-script methods have insufficient adaptability, and there is a lack of unified parameterized intermediate action representations, leading to low control robustness, naturalness, and task execution success rate. Therefore, an effective solution is urgently needed to address these technical problems.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a robot control method based on surface bioelectrical signals, the method comprising: Multi-channel surface bioelectrical signals from the operator are acquired at a fixed sampling frequency to generate a continuous time-series signal stream of surface bioelectrical signals; The temporal signal sequence with a preset historical time step is extracted from the temporal signal stream of the surface bioelectric signal, and the temporal dependency between signals is captured by an encoder to output temporal sequence features, wherein the temporal sequence features are used to characterize the evolution trend of the operator's actions. The time-series features are used as conditions and input into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space. The parameterized action primitive sequence includes action primitives for a preset number of future steps. The action primitive is the smallest execution unit of a robot containing continuously adjustable parameters. The continuous action space refers to the space consisting of all possible values ​​of the action primitives. Each dimension in the continuous action space is a continuous real number. A preset number of motion primitives are selected from the parameterized motion primitive sequence in chronological order, and the selected motion primitives are converted into robot control commands to control the robot's motion. Based on the time-series signal stream of continuously acquired surface bioelectric signals, the parameterized action primitive sequence is repeatedly generated and updated, replacing the unexecuted action primitives in the original sequence, thereby realizing continuous adaptive motion control of the robot. The conditional generation model is a conditional diffusion generation model, and the training steps of the conditional diffusion generation model include: Collect surface bioelectrical signals during operator instruction, as well as actual robot motion trajectory data; Based on the motion characteristics of the robot's actual motion trajectory data, the robot's actual motion trajectory data is divided into several motion trajectory segments, wherein each motion trajectory segment corresponds to a unique action category, and the action categories include movement, hovering, and operation. Each motion trajectory segment is converted into a corresponding real parameterized action primitive sequence, wherein the real parameterized action primitive sequence contains several action primitives; Construct a set of paired samples of surface bioelectrical signals and corresponding real parameterized action primitive sequences; The conditional diffusion generative model is trained using the paired sample set.

[0007] In one implementation, the robot is a robotic arm with grippers, and each motion primitive corresponds to a parameter vector. The parameter vector includes the Cartesian space position increment, velocity increment, gripper opening and closing amount, and motion primitive duration parameters of the robotic arm end effector.

[0008] In one implementation, the preset historical time step and the preset future number of steps are adaptively determined by the current signal quality, task stage, or control mode, wherein the preset historical time step is used to characterize the length of the observation field of view, and the preset future number of steps is used to characterize the length of the predicted field of view.

[0009] In one implementation, after using the temporal sequence features as conditions to input into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space, the method further includes: Calculate the confidence level of the generated parameterized action primitive sequence; When the confidence level is lower than a preset threshold, or the surface bioelectric signal quality is lower than a preset threshold, a safety control action primitive is generated and the parameterized action primitive sequence is updated. The safety control action primitive includes a hovering primitive, a limiting primitive, or a deceleration primitive.

[0010] In one implementation, the robot includes an operation module, wherein the robot's actual motion trajectory data is segmented into several motion trajectory segments based on motion characteristics, including: The motion characteristics of the robot's actual motion trajectory data are acquired and judged, wherein the motion characteristics include the displacement of the operation module, the direction of the displacement of the operation module, the speed of the operation module, the duration of the operation module, and the state of the operation module; When the displacement direction of the operation module is continuous and consistent, and the movement speed of the operation module is greater than a preset speed threshold, it is determined to be a motion trajectory segment of the moving type. When the displacement of the operation module is less than a preset displacement threshold and the movement duration of the operation module exceeds a preset duration threshold, it is determined to be a hovering motion trajectory segment. When the state change of the operation module is greater than a preset change threshold, it is determined to be a motion trajectory segment of the operation type; When the displacement direction of the operation module changes, the movement speed of the operation module crosses zero, the operation module experiences a continuous pause, or the state of the operation module changes, the boundary point of adjacent motion trajectory segments is determined.

[0011] In one implementation, training the conditional diffusion generative model using the paired sample set includes: Perform forward Gaussian noise addition on the real parameterized action primitive sequences in the paired sample set to obtain noisy action sequences; Based on the corresponding surface bioelectric signals, time-series features are extracted, and a loss function is constructed by combining the time-series features with noisy action sequences. The conditional diffusion generative model is trained by iterative parameter optimization based on the loss function.

[0012] Secondly, embodiments of the present invention also provide a robot control system based on surface bioelectrical signals, the system comprising: The time-series signal stream acquisition module is used to acquire multi-channel surface bioelectric signals from the operator at a fixed sampling frequency and generate a continuous time-series signal stream of surface bioelectric signals. The temporal sequence feature acquisition module is used to extract a temporal signal sequence with a preset historical time step from the temporal signal stream of the surface bioelectric signal, and capture the temporal dependency between signals through an encoder to output temporal sequence features, wherein the temporal sequence features are used to characterize the evolution trend of the operator's actions; The parameterized action primitive sequence acquisition module is used to input the time sequence features as conditions into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space. The parameterized action primitive sequence includes action primitives for a preset number of future steps, and the action primitive is the robot's smallest execution unit containing continuously adjustable parameters. The motion primitive control module is used to select a preset number of motion primitives from the parameterized motion primitive sequence in time order, and convert the selected motion primitives into robot control commands to control the robot's motion. The motion primitive update control module is used to repeatedly generate updated parameterized motion primitive sequences based on the time-series signal stream of continuously acquired surface bioelectric signals, and replace the motion primitives that were not executed in the original sequence, so as to realize continuous adaptive motion control of the robot.

[0013] Thirdly, embodiments of the present invention also provide a terminal device, the terminal device including a memory, a processor, and a robot control program based on surface bioelectric signals stored in the memory and executable on the processor, wherein when the processor executes the robot control program based on surface bioelectric signals, it implements the steps of the robot control method based on surface bioelectric signals as described in any of the above schemes.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a robot control program based on surface bioelectric signals. When the robot control program based on surface bioelectric signals is executed by a processor, it implements the steps of the robot control method based on surface bioelectric signals as described in any of the above schemes.

[0015] Beneficial Effects: This invention discloses a robot control method, system, terminal device, and medium based on surface bioelectrical signals, relating to the field of robot control technology. The method first acquires multi-channel surface bioelectrical signals from an operator at a fixed sampling frequency, generating a continuous temporal signal stream of surface bioelectrical signals. Then, a temporal signal sequence with a preset historical time step is extracted from the temporal signal stream of the surface bioelectrical signals, and the temporal dependencies between signals are captured by an encoder, outputting temporal sequence features, which characterize the operator's action evolution trend. Next, the temporal sequence features are used as conditions and input into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space. The parameterized action primitive sequence includes action primitives for a preset number of future steps, and each action primitive is a robot's smallest execution unit containing continuously adjustable parameters. Subsequently, a preset number of action primitives are selected from the parameterized action primitive sequence in chronological order, and the selected action primitives are converted into robot control commands to control the robot's actions. Finally, based on the time-series signal stream of continuously acquired surface bioelectric signals, an updated parameterized action primitive sequence is repeatedly generated to replace the unexecuted action primitives in the original sequence, thereby realizing continuous adaptive motion control of the robot.

[0016] The method proposed in this invention does not rely on discrete gesture label mapping, avoiding abrupt action changes and achieving smooth and continuous robot control. Furthermore, by capturing temporal dependencies and modeling action evolution trends, it improves the stability and naturalness of control. By employing parameterized action primitives as a unified intermediate representation, it connects high-level task intent with low-level execution control. Combined with future multi-step sequence generation and rolling execution mechanisms, it significantly enhances the robot's real-time adaptability to changes in operator intent. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a specific implementation of the robot control method based on surface bioelectric signals provided in this embodiment of the invention.

[0018] Figure 2 This is a flowchart illustrating the robot control method based on surface bioelectric signals provided in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of the robot control device based on surface bioelectric signals provided in an embodiment of the present invention.

[0020] Figure 4 This is a block diagram illustrating the internal structure of the terminal device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, 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 merely illustrative of the invention and are not intended to limit the invention.

[0022] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0023] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, the first control information and the second control information are only used to distinguish different control information and do not limit their order.

[0025] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0026] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] With the continuous development of human-robot collaboration, rehabilitation assistance, teleoperation, and dexterous manipulation applications, how to utilize natural, low-latency, and continuously adjustable human physiological signals to achieve robot control has become an important research direction in the field of robot control. Surface bioelectrical signals can directly reflect muscle activation state and movement intention, and have advantages such as being non-invasive, wearable, and having a fast response, making them suitable as input signals for robot motion generation and execution control.

[0028] There are several existing robot control schemes based on surface bioelectrical signals. The first is the discrete gesture recognition triggering method, which typically outputs a limited number of category labels, and then triggers robot actions based on a preset mapping relationship. The second is the continuous regression method, which directly maps the control quantity from the current signal. The third is the task execution method based on predefined scripts or fixed action templates. However, each of these methods has its limitations. Discrete triggering methods struggle to express the direction, amplitude, and duration of actions, are prone to abrupt changes in action, and have poor continuity. While single-step continuous regression methods can output continuous control quantities, they lack the ability to model future action evolution, requiring the operator to maintain continuous and stable force application, and are susceptible to muscle fatigue, electrode misalignment, and environmental noise. Fixed-script methods lack adaptability, making it difficult to correct the trajectory online based on real-time force changes, and are prone to task execution failure when the environment or target position changes.

[0029] Furthermore, existing technologies generally lack a unified action representation between high-level task intent and low-level execution control. For example, some existing technologies encode electromyography (EMG) signals and input them into a diffusion model to estimate a semi-structured latent space representation of hand movement states, and further decode finger movements, action strategies, and interaction object information. The core focus of these technologies lies in decoding hand labor strategies. Other existing technologies convert filtered and denoised EMG signals into robotic arm control parameters after feature extraction and mapping, and achieve continuous control through interpolation optimization. The focus of these technologies is on control parameter mapping and continuous optimization. Neither of these two types of technologies establishes a unified intermediate representation around parameterized action primitives, nor do they provide a generative execution mechanism that can adaptively switch between single-step high-frequency fine-tuning and multi-step coherent task execution.

[0030] Therefore, existing technologies cannot simultaneously achieve continuous and smooth control, future action evolution modeling, and real-time environmental adaptability under the same architecture. Discrete triggering methods result in discontinuous control, single-step regression lacks sequence modeling capabilities, fixed script methods have insufficient adaptability, and there is a lack of unified parameterized intermediate action representations, leading to low control robustness, naturalness, and task execution success rate.

[0031] To address the problems of discontinuous control in discrete triggering methods, lack of sequence modeling capability in single-step continuous regression, and insufficient adaptability of fixed script methods in existing technologies, this invention proposes a robot control method based on surface bioelectrical signals. The implementation process of this method is as follows: Figure 2As shown, this method acquires multi-channel surface electromyography (EMG) signals from the user, extracts instantaneous and sequential features based on short-time sliding window signals and historical time-series signals, and divides the continuous operation process into several parameterized motion primitive segments according to the end displacement direction, displacement amplitude, velocity zero-crossing point, pause interval, and gripper state changes in the taught trajectory, thus constructing a motion primitive library. Subsequently, based on the current signal quality, task stage, or control mode, the observation field of view and predicted field of view length are adaptively determined. A conditional diffusion generation network directly generates the sequence of motion primitives for the next one or more steps in the continuous motion space. For moments with insufficient generation confidence, hovering primitives, amplitude-limiting primitives, or deceleration primitives are triggered to improve execution safety. Finally, a rolling temporal execution strategy is used to convert the first step or the first few steps in the predicted sequence into robot control quantities for execution, and EMG signals are reacquired and the generation results are updated at the next time step.

[0032] This embodiment provides a robot control method based on surface bioelectrical signals, such as... Figure 1 As shown, the specific steps include the following: Step S100: Acquire multi-channel surface bioelectric signals of the operator at a fixed sampling frequency to generate a continuous time-series signal stream of surface bioelectric signals.

[0033] In this embodiment, a fixed sampling frequency refers to a sampling rate that remains constant throughout the entire signal acquisition process. The sampling frequency can be selected from 100 Hz to 2000 Hz. This embodiment uses a sampling frequency of 1000 Hz, which ensures signal time resolution while avoiding excessive data volume, balancing acquisition accuracy and system real-time performance. Multi-channel surface bioelectrical signals refer to physiological signals reflecting the electrical activity of different muscle groups in the human body, acquired simultaneously through multiple electrodes. In this embodiment, surface electromyography (EMG) signals can be used as surface bioelectrical signals. Surface EMG signals reflect the activation state and movement intention of muscles, offering advantages such as non-invasiveness, wearability, and fast response. The device for acquiring bioelectrical signals can be a wearable armband sensor. The number of channels can be specifically set to 8, 16, or 32 channels based on application requirements. Different channels correspond to different muscle groups in different parts of the arm, such as the biceps brachii, triceps brachii, forearm flexors, and extensors.

[0034] A continuous temporal signal stream of surface bioelectrical signals refers to a continuous data sequence formed by arranging multi-channel signal data collected at each sampling time in chronological order. The data at each sampling point is a multi-dimensional vector, with the dimension equal to the number of channels, and each element representing the signal amplitude of the corresponding channel at that moment. By generating a temporal signal stream, discrete sampled data is converted into a continuous data stream with a defined time dimension. Using a fixed sampling frequency to acquire multi-channel surface bioelectrical signals ensures temporal consistency and spatial integrity of the signals, accurately capturing the dynamic changes in the operator's muscle activity.

[0035] Step S200: Extract a time-series signal sequence with a preset historical time step from the time-series signal stream of the surface bioelectric signal, and capture the temporal dependency between signals through an encoder to output time-series sequence features, wherein the time-series sequence features are used to characterize the evolution trend of the operator's actions.

[0036] In this embodiment, the preset historical time step refers to the number of sampling points traced back from the current moment. In this embodiment, the preset historical time step is used to characterize the length of the observation field, and its value range can be selected from 10 to 100 sampling points, corresponding to historical signals of 10 to 100 milliseconds. When the preset historical time step is set to 1, a single frame signal at the current moment is extracted, corresponding to the robot's single-step high-frequency control mode. When the preset historical time step is greater than 1, a signal sequence containing multiple historical moments is extracted, corresponding to the robot's sequence generation mode. By adjusting the preset historical time step, signal extraction at dual time scales is achieved, enabling simultaneous fine-grained real-time fine-tuning and coherent task execution within the same architecture.

[0037] Specifically, when the preset historical time step is set to 1, the current time is extracted. Short-term sliding window signal Encoded as instantaneous features by a one-dimensional convolutional network When the preset historical time step is greater than 1, extract the past. Signal sequence at each historical time step The encoder captures temporal dependencies and outputs sequence features. .

[0038] An encoder is a neural network model that converts an input temporal signal sequence into a high-dimensional feature vector. In this embodiment, encoder types that can be used include Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), and Transformer networks. This embodiment uses a Transformer network. Transformer networks, through a self-attention mechanism, capture long-range temporal dependencies, modeling the correlation between signals at different time points, such as the order of muscle activation, intensity trends, and the transition process of movement. Temporal dependencies between signals refer to the statistical correlation between signal values ​​at different time points in a temporal signal sequence. This statistical correlation reflects the continuity and regularity of human movement.

[0039] Temporal sequence features are high-dimensional vectors output by the encoder. In this embodiment, the dimension can be set to 128, 256, or 512. The temporal sequence feature vector extracts key information related to the motion intention from the original temporal signal, characterizing the operator's action evolution trend, such as the transition from a static to a moving state, changes in movement direction, and the preparation process for the action. By capturing temporal dependencies and extracting temporal sequence features, historical signal information is effectively utilized to compensate for the insufficient information in a single frame, improving the accuracy and stability of action prediction and reducing interference caused by muscle fatigue, electrode misalignment, and environmental noise.

[0040] Step S300: Using the time sequence features as conditions, input the pre-trained conditional generation model to generate a parameterized action primitive sequence in the continuous action space. The parameterized action primitive sequence includes action primitives for a future preset number of steps. The action primitive is the robot's smallest execution unit containing continuously adjustable parameters. The continuous action space refers to the space consisting of all possible values ​​of the action primitives. Each dimension in the continuous action space is a continuous real number.

[0041] In this embodiment, the conditional generation model is a deep learning model that generates sample data conforming to a specific distribution under the guidance of given conditional information. Conditional generation models that can be used in this embodiment include conditional diffusion generation models, flow matching generation models, and variational autoencoders, with the conditional diffusion generation model specifically used in this embodiment. Because the conditional diffusion generation model has the characteristics of high generation quality, stable training, and the ability to handle high-dimensional continuous data, it is suitable for generating continuous robot action sequences. A pre-trained conditional generation model refers to a model that has been trained using a large amount of teaching data and can generate reasonable action sequences based on the input conditional features.

[0042] The continuous action space refers to the space consisting of all possible values ​​of action primitives. Each dimension in this space is a continuous real number, rather than a discrete category label. By representing actions continuously, the amplitude, direction, and duration of actions can be continuously adjusted, avoiding the abrupt changes and unnaturalness caused by discrete actions. A parameterized action primitive sequence is a sequence composed of multiple parameterized action primitives arranged in chronological order. The future preset number of steps refers to the number of future action primitives predicted by the model, also known as the prediction field of view length, and its value can be selected from 1 to 10 steps. When the future preset number of steps is 1, the model generates single-step action primitives for high-frequency real-time fine-tuning. When the future preset number of steps is greater than 1, the model generates multi-step coherent action sequences for completing task segments such as grasping, moving, and placing.

[0043] Action primitives are the smallest units of action executed by a robot. They are not fixed, discrete category labels, but rather parameterized action segments containing a set of continuously adjustable parameters. Each action primitive corresponds to a complete basic action with clear semantics, such as movement along a certain direction, hovering to maintain the current position, or the opening and closing operation of an end effector. By decomposing complex continuous operation processes into a series of parameterized action primitives, a unified intermediate representation is established between high-level task intent and low-level execution control, simplifying the complexity of robot control while ensuring the continuity and naturalness of actions. It should be noted that, unlike schemes that treat up, down, left, right, grasp, and release as fixed categories, in this embodiment, action primitives are only a set of example semantics for actions. The same action primitive can express different intensities and rhythms of execution processes through variations in amplitude and duration.

[0044] In one implementation, the robot is a robotic arm with grippers, and each motion primitive corresponds to a parameter vector. The parameter vector includes the Cartesian space position increment, velocity increment, gripper opening and closing amount, and motion primitive duration parameters of the robotic arm end effector.

[0045] In this embodiment, the gripper-equipped robotic arm is a common configuration in industrial and service robots. It typically has a 6-DOF robotic arm body and a 1-DOF gripper end effector, capable of performing most grasping, handling, and manipulation tasks. Each motion primitive corresponds to a 5-dimensional parameter vector, which fully describes all the motion characteristics of that motion primitive.

[0046] The specific form of the parameter vector is:

[0047] in, These represent the position increments of the robotic arm's end effector along the X, Y, and Z axes in a Cartesian coordinate system, respectively, in meters or millimeters. These three parameters determine the spatial displacement of the robotic arm's end effector during the execution of this motion element. As an equivalent deformation scheme, It can also be expressed as the velocity increment of the end effector in Cartesian space, in meters per second. In this case, the execution effect of the motion primitive is determined by both the velocity increment and the duration. This indicates the opening / closing amount of the grippers, and its value ranges from 0 to 1, where 0 represents the grippers fully closed and 1 represents the grippers fully open. This is achieved through continuous adjustment. The value of can enable the gripper to flexibly grasp objects of different shapes and materials. This indicates the duration of the action primitive, in seconds, and determines how fast the action primitive is executed. It represents the real number space field.

[0048] Besides the Cartesian space parameter representation mentioned above, the parameter vectors of motion primitives can also use joint space increments, joint velocities, accelerations, impedance parameters, or combinations thereof as motion descriptions. Joint space increments represent the angular changes of each joint of the robotic arm; this representation is suitable for robots with joint space control and is directly converted into control commands for the joint motors. Impedance parameters are used to achieve compliant robot control, enabling the robot to exhibit appropriate stiffness and damping characteristics when interacting with the environment, improving operational safety and stability. Using parameter vectors to represent motion primitives transforms complex robot motion into a set of concise, continuous parameters, facilitating model learning and generation, while simultaneously achieving continuous adjustability and smooth transitions in motion.

[0049] In one implementation, the preset historical time step and the preset future number of steps are adaptively determined by the current signal quality, task stage, or control mode, wherein the preset historical time step is used to characterize the length of the observation field of view, and the preset future number of steps is used to characterize the length of the predicted field of view.

[0050] In this embodiment, current signal quality refers to the quality of the real-time acquired surface bioelectrical signals, and its evaluation indicators mainly include signal-to-noise ratio (SNR) and root mean square (RMS) value. A higher SNR indicates more effective components and less noise in the signal, while the RMS value reflects the overall activation level of the muscle. When the signal quality is good, the predicted visual field length is increased to generate longer, more coherent action sequences, reducing the computational load of the control cycle. When the signal quality is poor, the predicted visual field length is decreased, the observation visual field length is increased, and more historical signals are utilized to improve the reliability of the prediction.

[0051] The task phase refers to the current stage of the task the robot is performing, such as the grasping phase, the moving phase, the placement phase, and the fine manipulation phase. In high-precision phases like grasping and placement, a smaller predicted field of view (RFV) and a larger observed field of view (AVV) are chosen to ensure control accuracy and real-time performance. In continuous phases like moving, a larger RAV is used to generate multi-step, coherent movements, making the robot's motion smoother and more natural.

[0052] Control mode refers to the current operating mode of the system, mainly including single-step high-frequency fine-tuning mode and multi-step continuous task mode. In single-step high-frequency fine-tuning mode, the preset historical time step is set to 1, and the preset number of future steps is also set to 1. Single-step action primitives are generated and executed at the highest frequency to achieve fine control of the robot. In multi-step continuous task mode, both the preset historical time step and the preset number of future steps are set to values ​​greater than 1. Multi-step action sequences are generated and executed to complete continuous task operations.

[0053] By adaptively determining the observation field length and prediction field length based on the current signal quality, task stage, or control mode, the real-time performance, accuracy, and consistency of control can be dynamically balanced under the same model architecture. This achieves the unification of single-step high-frequency fine-tuning and multi-step coherent task execution, significantly improving the adaptability and task execution efficiency of the control method.

[0054] In one implementation, after using the temporal sequence features as conditions to input into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space, the specific steps include: Step S310: Calculate the confidence level of the generated parameterized action primitive sequence; Step S320: When the confidence level is lower than a preset threshold, or the surface bioelectric signal quality is lower than a preset threshold, generate a safety control action primitive and update the parameterized action primitive sequence, wherein the safety control action primitive includes a hovering primitive, a limiting primitive, or a deceleration primitive.

[0055] In this embodiment, a confidence score is calculated on the generated parameterized action primitive sequence to evaluate the reliability of the model output. The confidence score can be calculated based on the output variance of the generated model; a larger variance indicates higher uncertainty about the output and lower confidence. Alternatively, it can be indirectly evaluated based on the quality of the surface bioelectrical signal; poorer signal quality results in lower confidence. A preset threshold is a lower confidence limit pre-set according to actual application requirements. When the confidence score of the generated result falls below this threshold, the result is considered unreliable, triggering a safety control mechanism.

[0056] Safety control motion primitives are special motion primitives specifically designed to ensure system safety, mainly including hovering primitives, amplitude limiting primitives, and deceleration primitives. The parameter vector of the hovering primitive is... ,in The current gripper opening / closing amount is used. When executing the hovering primitive, the robot's end effector remains stationary in its current position, and the gripper state also remains unchanged. The limiting primitive limits the position increment, velocity increment, and other parameters of the generated motion primitive within a preset safety range to prevent the robot from producing excessive motion amplitude. The deceleration primitive multiplies the velocity parameter of the motion primitive by a deceleration coefficient less than 1, reducing the robot's movement speed and improving response time and safety.

[0057] When the confidence level of the generated parameterized action primitive sequence is lower than a preset threshold, or the quality of the real-time surface bioelectric signal is lower than a preset threshold, a corresponding safety control action primitive is generated. This safety control action primitive replaces the action primitives in the original generated sequence, updates the parameterized action primitive sequence, and then executes it. This safety execution strategy based on confidence level and signal quality effectively avoids erroneous operations caused by model prediction errors or signal interference, significantly improving the robustness and security of the system in actual deployment.

[0058] Step S400: Select a preset number of motion primitives from the parameterized motion primitive sequence in chronological order, and convert the selected motion primitives into robot control commands to control the robot's motion.

[0059] In this embodiment, the preset number refers to the number of action primitives actually selected and executed from the generated sequence of future multi-step action primitives. In this embodiment, the value range can be selected from 1 to 3 steps. Predicting a longer sequence of action primitives and selecting shorter action primitives from it for actual execution can maximize the real-time performance of the system while ensuring the continuity of actions, enabling the robot to respond promptly to changes in the operator's intentions.

[0060] Robot control commands refer to instruction data that can be directly executed by the robot's underlying controller. Specifically, combining the kinematic model and dynamic characteristics of the target robot, parameterized motion primitives are converted into robot control commands. For robots controlled in Cartesian space, the position or velocity increment parameters in the motion primitives are directly converted into target pose or velocity commands for the end effector. For robots controlled in joint space, the Cartesian space parameters are converted into angle or angular velocity commands for each joint through inverse kinematics solving. For robots with grippers, the gripper opening / closing parameters can be directly converted into control current or position commands for the gripper motors. The duration parameter of the motion primitive is used to determine the execution time of the motion primitive. The controller will smoothly execute the corresponding control commands within this duration to ensure the smoothness of the motion.

[0061] Step S500: Based on the time-series signal stream of continuously acquired surface bioelectric signals, repeatedly generate updated parameterized action primitive sequences, replace the unexecuted action primitives in the original sequence, and realize continuous adaptive motion control of the robot.

[0062] In this embodiment, repeatedly generating updated parameterized action primitive sequences refers to the system continuously executing steps S100 to S300 according to a fixed control cycle. The length of the control cycle can be set to be the same as the sampling cycle. Within each control cycle, the latest surface bioelectrical signals are acquired, the latest temporal sequence features are extracted, and a completely new future multi-step action primitive sequence is generated.

[0063] Replacing unexecuted action primitives in the original sequence refers to replacing the unexecuted parts of the sequence generated in the previous cycle with newly generated action primitive sequences. For example, if the previous cycle generated a sequence of action primitives for the next 5 steps, and the system executes step 1, then generates a new sequence of 5 steps in the current cycle, discarding steps 2 through 5 from the previous cycle's sequence and executing step 1 of the new sequence. Steps 2 through 5 of the new sequence are then treated as future actions to be executed. Through this rolling time-domain execution mechanism, the robot's control process is always based on the latest operator intent information, correcting the motion trajectory in real time to adapt to changes in operator force, environmental changes, and target position changes, achieving true continuous adaptive control.

[0064] In one implementation, the conditional generation model is a conditional diffusion generation model, and the training steps of the conditional diffusion generation model specifically include: Step S610: Collect surface bioelectric signals during operator instruction and actual robot motion trajectory data; Step S620: Based on the motion characteristics of the robot's actual motion trajectory data, the robot's actual motion trajectory data is divided into several motion trajectory segments, wherein each motion trajectory segment corresponds to a unique action category, and the action categories include movement, hovering, and operation. Step S630: Convert each motion trajectory segment into a corresponding real parameterized action primitive sequence, wherein the real parameterized action primitive sequence contains several action primitives; Step S640: Construct a set of paired samples of surface bioelectric signals and corresponding real parameterized action primitive sequences; Step S650: Train the conditional diffusion generation model using the paired sample set.

[0065] In this embodiment, the training of the conditional diffusion generative model is an offline process, aiming to enable the model to learn the mapping relationship between surface bioelectrical signals and robot motion primitive sequences. Surface bioelectrical signals and actual robot motion trajectory data are collected during operator teaching to construct the training dataset. The teaching process is performed by the operator, who can guide the robot to perform various typical tasks such as grasping, moving, and placing using a joystick, master-slave manipulator, or direct robot traction. During teaching, multi-channel surface bioelectrical signals from the operator's arm and actual robot motion trajectory data are simultaneously collected. The actual robot motion trajectory data includes at least the end effector's pose information, velocity information, gripper opening and closing state information, and corresponding timestamp information, comprehensively recording the robot's motion state during the teaching process.

[0066] Based on the motion characteristics of the robot's actual motion trajectory data, the continuous teaching trajectory is divided into several motion trajectory segments with single action semantics. Each motion trajectory segment corresponds to a unique action category, which mainly includes movement, hovering, and manipulation. Movement trajectory segments correspond to the translational motion of the robot's end effector in space, hovering trajectory segments correspond to the robot's end effector remaining stationary in its current position, and manipulation trajectory segments correspond to the opening or closing action of the gripper. Each motion trajectory segment is converted into a corresponding realistic parameterized action primitive sequence by further dividing the trajectory segment into multiple smaller action units at fixed time intervals and extracting the corresponding parameter vector for each action unit to form the realistic parameterized action primitive sequence.

[0067] Specifically, for each segmented motion primitive, its parameterized description information is extracted, preferably including motion type, end-effector displacement increment, end-effector velocity increment, gripper opening / closing amount, duration, and start and end times. The parameter vector for each motion primitive segment... It can be represented as:

[0068] in, Indicates the first The displacement or velocity increment of an action primitive in Cartesian space Indicates the opening and closing amount of the grippers. This represents the duration parameter of the motion primitive. As an equivalent variation, the parameter vector can also use joint space increment, joint velocity, acceleration, impedance parameters, or combinations thereof as the motion description.

[0069] Multiple action primitive fragments are combined into an action primitive sequence in their original temporal order to obtain the target sequence sample for training:

[0070] in, This represents the length of the action primitive sequence. An action primitive library is thus constructed, which is used by conditional generative models to learn the correspondence between multi-channel surface electromyography signals and action primitive sequences.

[0071] The acquired surface bioelectrical signal sequences were paired with the converted real parameterized action primitive sequences in a time-synchronized manner to construct a set of paired samples of surface bioelectrical signals and their corresponding real parameterized action primitive sequences. Each paired sample contains a continuous surface bioelectrical signal sequence and the corresponding real action primitive sequence.

[0072] Specifically, after constructing the action primitive library, a conditional generative model is used to learn the correspondence between electromyographic (EMG) conditional features and action primitive sequences. For standardized notation, the EMG conditional features input to the generative model are denoted as... In single-window input mode Based on the current electromyographic characteristics Conditional diffusion models directly generate the future. The sequence of action primitives for each step is represented as:

[0073] in, This represents the generated sequence of action primitives. Indicates the first Step movement basics, The determination is adaptively made based on the current control mode or the predicted field of view. When At that time, output single-step action primitives for high-frequency real-time fine-tuning. The model outputs a coherent sequence of actions consisting of multiple parameterized action primitives, used to complete task segments such as grasping, moving, and placing. It should be noted that although the action sequence can be physically interpreted as a combination of several action primitives, the model output is always a continuous control quantity, not a discrete action category label.

[0074] Finally, the conditional diffusion generative model is trained using the constructed paired sample set, enabling the model to generate action primitive sequences similar to human teaching behavior based on the input surface bioelectric signal characteristics.

[0075] In one implementation, the robot includes an operation module, wherein the step of dividing the robot's actual motion trajectory data into several motion trajectory segments based on the motion characteristics of the robot's actual motion trajectory data specifically includes the following steps: Step S621: Obtain and judge the motion characteristics of the robot's actual motion trajectory data, wherein the motion characteristics include the displacement of the operation module, the direction of the displacement of the operation module, the speed of the operation module, the duration of the operation module, and the state of the operation module. Step S622: When the displacement direction of the operation module is continuous and consistent, and the movement speed of the operation module is greater than a preset speed threshold, it is determined to be a motion trajectory segment of the moving type. Step S623: When the displacement of the operation module is less than a preset displacement threshold and the movement duration of the operation module exceeds a preset duration threshold, it is determined to be a hovering motion trajectory segment. Step S624: When the state change of the operation module is greater than a preset change threshold, it is determined to be a motion trajectory segment of the operation type; Step S625: When the displacement direction of the operation module changes, the movement speed of the operation module crosses zero, the operation module experiences a continuous pause, or the state of the operation module changes, determine the boundary point of adjacent motion trajectory segments.

[0076] In this embodiment, the operation module refers to the component on the robot that directly interacts with the environment or the object being operated on. For a robotic arm with a gripper, the operation module is the end effector and gripper of the robotic arm. Motion characteristics are quantitative indicators reflecting the motion state extracted from the robot's actual motion trajectory data, including operation module displacement, operation module displacement direction, operation module speed, operation module motion duration, and operation module state. Operation module displacement refers to the change in spatial position of the operation module per unit time; operation module displacement direction refers to the direction of the displacement vector in space; operation module speed refers to the rate of change of displacement; operation module motion duration refers to the duration of a certain motion state; and operation module state mainly refers to the opening and closing state of the gripper.

[0077] When the displacement direction of the operating module remains consistent across multiple consecutive sampling points, and the movement speed of the operating module exceeds a preset speed threshold, the trajectory segment is determined to be a moving motion trajectory segment. The preset speed threshold can be set to 0.01 meters per second; when the speed is below this threshold, the robot is considered to be in a stationary or slightly moving state. When the displacement of the operating module is less than a preset displacement threshold, and the duration of this state exceeds a preset duration threshold, the trajectory segment is determined to be a hovering motion trajectory segment. The preset displacement threshold can be set to 0.001 meters, and the preset duration threshold is typically set to 0.5 seconds. When the state change of the operating module exceeds a preset change threshold, the trajectory segment is determined to be an operational motion trajectory segment. For grippers, the preset change threshold is typically 0.1, meaning that when the change in the gripper opening / closing amount exceeds 0.1, an operational action is considered to have occurred.

[0078] The rules for determining the boundary points of adjacent motion trajectory segments specifically include: when the movement direction of the operation module changes significantly; when the movement speed of the operation module crosses zero; when the operation module pauses for a duration exceeding a preset threshold; and when the state of the operation module changes. Based on these rules, continuous teaching trajectories are automatically segmented into multiple motion trajectory segments with single action semantics for constructing an action primitive library. This automatic segmentation eliminates the need for manual intervention, efficiently processing large amounts of teaching data, and the segmentation results align with natural human perception of movement.

[0079] In one implementation, training the conditional diffusion generative model using the paired sample set specifically includes the following steps: Step S651: Perform forward Gaussian noise addition processing on the real parameterized action primitive sequences in the paired sample set to obtain noisy action sequences; Step S652: Extract time-series features based on the corresponding surface bioelectric signals, and construct a loss function by combining the time-series features with noisy action sequences. Step S653: Perform parameter iterative optimization training on the conditional diffusion generation model based on the loss function.

[0080] In this embodiment, during the training phase, real action primitive sequences from the action primitive library are used. As a clean sample, Gaussian noise was gradually added to it using a forward diffusion noise addition process to obtain noise sequences at different diffusion steps. Specifically, the forward noise addition process can be represented as:

[0081] in, Indicates the first A sequence of actions for each diffusion step. Indicates the first The action sequence of a diffusion step Preset noise scheduling parameters are used to control the first... The intensity of the noise added in each diffusion step, It is an identity matrix.

[0082] During the inference phase, i.e., the actual control phase, the network transitions from standard Gaussian noise. Let's begin. Introduce a denoising neural network. And in each step of the denoising process, electromyographic features are included. Injected as a condition. The inverse denoising distribution can be represented as:

[0083] in, This represents the encoded electromyographic condition, calculated by the electromyographic signal encoder using paired raw electromyographic signals. For different modes, It can be or . and They respectively represent the parameters as The denoising neural network outputs the mean and covariance parameters. Specifically, the denoising neural network adopts a U-Net structure, a time-encoded encoder-decoder network, or other network structures for conditional denoising generation. Under the strong guidance of the encoded electromyographic conditions, the conditional diffusion generation module performs multi-step iterative denoising on random noise. At each denoising time step, it combines electromyographic features to remove a portion of the noise, ultimately obtaining the continuous motion prediction result for the current control cycle. .

[0084] Furthermore, define Then the mean parameter It can be represented as:

[0085] in, This represents the noise term predicted by the denoising neural network.

[0086] The goal of model training is to make the noise predicted by the network as close as possible to the actual noise added during the forward noise addition process. Preferably, the training loss function is the mean squared error loss:

[0087] in, It is random noise that follows a standard normal distribution. This represents the 0th diffusion step, i.e., the original clean action sequence.

[0088] The loss function minimizes the mean squared error between the noise predicted by the network and the actual noise added during the forward noise generation process. During training, the network parameters are iteratively optimized using stochastic gradient descent. This process gradually reduces the value of the loss function until the model converges. The trained conditional diffusion generative model, guided by surface bioelectrical signal features, can generate high-quality, parameterized action primitive sequences that conform to human movement habits by starting from random Gaussian noise and progressively denoising it.

[0089] In this embodiment, the denoising neural network attempts to predict the actual noise added during the forward pass through its downsampling, upsampling, and skip connection structure. The network weights of the denoising neural network and the electromyography signal encoder are updated via backpropagation by calculating the mean square error between the predicted noise and the actual added noise.

[0090] In summary, the motion primitives proposed in this embodiment do not output discrete category labels directly, but are implicitly expressed as one or more continuous control quantities among the following parameters: end-effector pose increment, velocity increment, gripper opening / closing amount, joint increment, and duration. This technical solution enables unified implementation of single-step high-frequency fine-tuning and multi-step coherent task execution within the same model framework, reducing the risk of abrupt motion changes and false triggering, and improving motion smoothness, naturalness, environmental adaptability, robustness, and safety. This method is also applicable to rehabilitation assistive actuators, prostheses, or other robotic devices with end-effector manipulation capabilities.

[0091] like Figure 3 As shown in the figure, this embodiment of the invention provides a robot control system based on surface bioelectric signals. The system includes: a time-series signal stream acquisition module 10, a time-series sequence feature acquisition module 20, a parameterized action primitive sequence acquisition module 30, an action primitive control module 40, and an action primitive update control module 50.

[0092] Specifically, the time-series signal stream acquisition module 10 is used to acquire multi-channel surface bioelectrical signals of the operator at a fixed sampling frequency to generate a continuous time-series signal stream of surface bioelectrical signals; the time-series sequence feature acquisition module 20 is used to extract a time-series signal sequence with a preset historical time step from the time-series signal stream of the surface bioelectrical signals, and capture the temporal dependencies between signals through an encoder to output time-series features, wherein the time-series sequence features are used to characterize the operator's action evolution trend; the parameterized action primitive sequence acquisition module 30 is used to input the time-series sequence features as conditions into a pre-trained conditional generation model to generate a continuous action space. A parameterized action primitive sequence is generated, wherein the parameterized action primitive sequence includes action primitives for a preset number of future steps, and the action primitive is the smallest execution unit of the robot containing continuously adjustable parameters; the action primitive control module 40 is used to select a preset number of action primitives from the parameterized action primitive sequence in chronological order, and convert the selected action primitives into robot control commands to control the robot's actions; the action primitive update control module 50 is used to repeatedly generate updated parameterized action primitive sequences based on the time-series signal stream of continuously acquired surface bioelectric signals, and replace the unexecuted action primitives in the original sequence to achieve continuous adaptive action control of the robot.

[0093] Based on the above embodiments, the present invention also provides a terminal device, the principle block diagram of which can be as follows: Figure 4As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a robot control method based on surface bioelectrical signals. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0094] Those skilled in the art will understand that Figure 4 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0095] In one embodiment, a terminal device is provided, including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs including instructions for performing operations as described in the embodiments of the methods above.

[0096] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0097] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A robot control method based on surface bioelectrical signals, characterized in that, The method includes: Multi-channel surface bioelectrical signals from the operator are acquired at a fixed sampling frequency to generate a continuous time-series signal stream of surface bioelectrical signals; The temporal signal sequence with a preset historical time step is extracted from the temporal signal stream of the surface bioelectric signal, and the temporal dependency between signals is captured by an encoder to output temporal sequence features, wherein the temporal sequence features are used to characterize the evolution trend of the operator's actions. The time-series features are used as conditions and input into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space. The parameterized action primitive sequence includes action primitives for a preset number of future steps. The action primitive is the smallest execution unit of a robot containing continuously adjustable parameters. The continuous action space refers to the space consisting of all possible values ​​of the action primitives. Each dimension in the continuous action space is a continuous real number. A preset number of motion primitives are selected from the parameterized motion primitive sequence in chronological order, and the selected motion primitives are converted into robot control commands to control the robot's motion. Based on the time-series signal stream of continuously acquired surface bioelectric signals, the parameterized action primitive sequence is repeatedly generated and updated, replacing the unexecuted action primitives in the original sequence, thereby realizing continuous adaptive motion control of the robot. The conditional generation model is a conditional diffusion generation model, and the training steps of the conditional diffusion generation model include: Collect surface bioelectrical signals during operator instruction, as well as actual robot motion trajectory data; Based on the motion characteristics of the robot's actual motion trajectory data, the robot's actual motion trajectory data is divided into several motion trajectory segments, wherein each motion trajectory segment corresponds to a unique action category, and the action categories include movement, hovering, and operation. Each motion trajectory segment is converted into a corresponding real parameterized action primitive sequence, wherein the real parameterized action primitive sequence contains several action primitives; Construct a set of paired samples of surface bioelectrical signals and corresponding real parameterized action primitive sequences; The conditional diffusion generation model is trained using the paired sample set; The robot includes an operation module, which, based on the motion characteristics of the robot's actual motion trajectory data, divides the robot's actual motion trajectory data into several motion trajectory segments, including: The motion characteristics of the robot's actual motion trajectory data are acquired and judged, wherein the motion characteristics include the displacement of the operation module, the direction of the displacement of the operation module, the speed of the operation module, the duration of the operation module, and the state of the operation module; When the displacement direction of the operation module is continuous and consistent, and the movement speed of the operation module is greater than a preset speed threshold, it is determined to be a motion trajectory segment of the moving type. When the displacement of the operation module is less than a preset displacement threshold and the movement duration of the operation module exceeds a preset duration threshold, it is determined to be a hovering motion trajectory segment. When the state change of the operation module is greater than a preset change threshold, it is determined to be a motion trajectory segment of the operation type; When the displacement direction of the operation module changes, the movement speed of the operation module crosses zero, the operation module experiences a continuous pause, or the state of the operation module changes, the boundary point of adjacent motion trajectory segments is determined.

2. The robot control method based on surface bioelectrical signals according to claim 1, characterized in that, The robot is a robotic arm with grippers. Each motion primitive corresponds to a parameter vector, which includes the Cartesian space position increment, velocity increment, gripper opening and closing amount, and motion primitive duration parameters of the robotic arm end effector.

3. The robot control method based on surface bioelectric signals according to claim 1, characterized in that, The preset historical time step and the preset future number of steps are adaptively determined by the current signal quality, task stage, or control mode. The preset historical time step is used to characterize the length of the observation field of view, and the preset future number of steps is used to characterize the length of the predicted field of view.

4. The robot control method based on surface bioelectric signals according to claim 1, characterized in that, After using the temporal sequence features as conditions to input into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space, the method further includes: Calculate the confidence level of the generated parameterized action primitive sequence; When the confidence level is lower than a preset threshold, or the surface bioelectric signal quality is lower than a preset threshold, a safety control action primitive is generated and the parameterized action primitive sequence is updated. The safety control action primitive includes a hovering primitive, a limiting primitive, or a deceleration primitive.

5. The robot control method based on surface bioelectric signals according to claim 1, characterized in that, The step of training the conditional diffusion generative model using the paired sample set includes: Perform forward Gaussian noise addition on the real parameterized action primitive sequences in the paired sample set to obtain noisy action sequences; Based on the corresponding surface bioelectric signals, time-series features are extracted, and a loss function is constructed by combining the time-series features with noisy action sequences. The conditional diffusion generative model is trained by iterative parameter optimization based on the loss function.

6. A robot control system based on surface bioelectrical signals, characterized in that, The system, used to implement the robot control method based on surface bioelectric signals as described in any one of claims 1-5, comprises: The time-series signal stream acquisition module is used to acquire multi-channel surface bioelectric signals from the operator at a fixed sampling frequency and generate a continuous time-series signal stream of surface bioelectric signals. The temporal sequence feature acquisition module is used to extract a temporal signal sequence with a preset historical time step from the temporal signal stream of the surface bioelectric signal, and capture the temporal dependency between signals through an encoder to output temporal sequence features, wherein the temporal sequence features are used to characterize the evolution trend of the operator's actions; The parameterized action primitive sequence acquisition module is used to input the time sequence features as conditions into a pre-trained conditional generation model to generate a parameterized action primitive sequence in a continuous action space. The parameterized action primitive sequence includes action primitives for a preset number of future steps, and the action primitive is the robot's smallest execution unit containing continuously adjustable parameters. The motion primitive control module is used to select a preset number of motion primitives from the parameterized motion primitive sequence in time order, and convert the selected motion primitives into robot control commands to control the robot's motion. The motion primitive update control module is used to repeatedly generate updated parameterized motion primitive sequences based on the time-series signal stream of continuously acquired surface bioelectric signals, and replace the unexecuted motion primitives in the original sequence to achieve continuous adaptive motion control of the robot.

7. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a robot control program based on surface bioelectric signals stored in the memory and executable on the processor. When the processor executes the robot control program based on surface bioelectric signals, it implements the steps of the robot control method based on surface bioelectric signals as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a robot control program based on surface bioelectric signals. When the robot control program based on surface bioelectric signals is executed by a processor, it implements the steps of the robot control method based on surface bioelectric signals as described in any one of claims 1-5.

Images