Robot skill simulation method, device and equipment and storage medium

Abstract

The invention relates to the field of robots, and discloses a skill simulation method, device and equipment of a robot and a storage medium. Comprising the following steps: acquiring a clamping jaw state when a robot completes an operation task, inputting the clamping jaw state into a clamping jaw simulation environment for analogue simulation to obtain a first joint action position, inputting the first joint action position into a joint simulation environment for analogue simulation to obtain a second joint action position and a visual image, and taking the second joint action position and the visual image as input data; and training and optimizing a preset initial prediction formula through a preset loss function to obtain a prediction formula, inputting the initial joint action position and the initial visual image of the robot in the joint simulation environment into the prediction formula to obtain a target joint action sequence, and driving the robot to execute corresponding actions. According to the scheme, the action sequence of the robot is predicted by learning the joint action information and the visual information, the technical difficulty of the robot during operation behavior learning is reduced, and the intelligent degree of the robot and the generalization ability of skill simulation are improved.

Description

Technical Field

[0001] The present invention relates to the field of robots, and in particular to a robot skill simulation method, device, equipment and storage medium. Background Art

[0002] With the rapid development of robotics technology, robots have shown great application potential in various industries due to their high flexibility and adaptability. However, there are many limitations in the skill learning of traditional robots. For example, traditional robots can only perform repetitive movements according to planned motion trajectories, cannot obtain environmental information, can only perform relatively single task types and are inefficient. In addition, traditional robots lack the ability to adapt and learn autonomously, which makes it difficult to autonomously plan motion trajectories and make flexible responses and adjustments when facing complex and changing environments or similar tasks. That is, in the existing technology, the skill learning of robots is achieved by the system issuing specific work instructions, and it is impossible to perform adaptive operations according to the actual environment. Summary of the invention

[0003] The main purpose of the present invention is to solve the technical problems existing in the existing robot skill simulation method, that is, the action execution is single and the operation cannot be adaptively operated according to the environment.

[0004] The first aspect of the present invention provides a method for simulating the skills of a robot, the method comprising: constructing a simulation environment based on an operation task, parsing and splitting the operation task, and obtaining the gripper state of the robot when completing the operation task, wherein the simulation environment comprises a gripper simulation environment based on gripper control and a joint simulation environment based on joint control; inputting the gripper state into the gripper simulation environment, and performing simulation based on the gripper state to obtain a first joint action position of the robot in the gripper simulation environment; inputting the first joint action position into the joint simulation environment, and performing simulation based on the first joint action position to obtain a second joint action position and a visual image of the robot in the joint simulation environment; training a preset initial prediction formula with the second joint action position and the visual image as input data, and optimizing the preset initial prediction formula with a preset loss function to obtain a prediction formula for the robot; obtaining the initial joint action position and the initial visual image of the robot in the joint simulation environment, and inputting the initial joint action position and the initial visual image into the prediction formula to obtain a target joint action sequence of the robot in the joint simulation environment, and driving the robot to perform corresponding actions based on the target joint action sequence.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the clamping jaw state includes a position, a rotation vector, and a gripping state of the clamping jaw, and the simulation of the clamping jaw simulation environment is:

[0006] in, represents the gripper simulation environment based on gripper control, represents the first joint action position output in the gripper simulation environment based on gripper control, represents the position and rotation vector of the gripper, Indicates the gripping state of the gripper. 0 means closed, 1 means open. Respectively represent the left arm and right arm of the dual-arm robot, Indicates time.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the visual image includes a left arm image, a right arm image, and a side image of the robot, and the simulation of the joint simulation environment is:

[0008] in, represents the joint simulation environment based on joint control, represents the second joint action position output in the joint simulation environment based on joint control, Represents a visual image in RGB, Respectively represent the robot's left arm, right arm and side, Indicates time.

[0009] Optionally, in a third implementation method of the first aspect of the present invention, the preset initial prediction formula is trained with the second joint action position and the visual image as input data, including: obtaining the second joint action position and the visual image output by the joint simulation environment, and generating input data based on the second joint action position and the visual image, and using the input data as training data to train the preset initial prediction formula.

[0010] Optionally, in a fourth implementation of the first aspect of the present invention, the input data is in the form of: ,

[0011] in, represents the input at time t, Indicates that The combined joint action position, for The combined visual image, dThe dimension representing the position of the joint action, that is, the dimension of the left arm + the right arm, Indicates Cameras, ; Represents the width × height × number of channels of the visual image, Indicates the time; the preset initial prediction formula is: ,

[0012] in, represents the predicted joint action sequence from time t to time t+k, represents the input data at the tth moment, is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence.

[0013] Optionally, in a fifth implementation of the first aspect of the present invention, the use of a preset loss function to optimize a preset initial prediction formula to obtain a prediction formula for the robot includes: calculating an error value between a predicted joint motion sequence and a collected joint motion sequence based on a preset loss function, and optimizing the preset initial prediction formula based on the error value to obtain a prediction formula for the robot, wherein the preset loss function is: ,

[0014] represents the predicted joint action sequence from time t to time t+k, represents the input joint action sequence, and the prediction formula is: ,

[0015] in, Represents the joint action position at time t represents the visual image of the i-th camera at time t; is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence, and .

[0016] Optionally, in a sixth implementation of the first aspect of the present invention, the driving the robot to perform a corresponding action based on the target joint action sequence includes: calculating the control torque of each joint of the robot based on the target joint action sequence through a preset control calculation algorithm, and driving the robot to perform the corresponding action based on each control torque, wherein the preset control calculation algorithm is: ,

[0017] represents the control torque of joint J at the current moment, represents the proportional gain coefficient of the controller, represents the derivative gain coefficient of the controller, represents the predicted action position of joint J at the next moment, represents the actual action position of joint J at the next moment, Represents the actual velocity of joint J at the next moment.

[0018] A second aspect of the present invention provides a robot skill simulation device, the device comprising: A construction module is used to construct a simulation environment based on an operation task, and parse and split the operation task to obtain the gripper state of the robot when completing the operation task, wherein the simulation environment includes a gripper simulation environment based on gripper control and a joint simulation environment based on joint control; A first simulation module is used to input the gripper state into the gripper simulation environment, and perform simulation based on the gripper state to obtain a first joint action position of the robot in the gripper simulation environment; A second simulation module is used to input the joint action position into the joint simulation environment, and perform simulation based on the first joint action position to obtain a second joint action position and a visual image of the robot in the joint simulation environment; A generation module, used to train a preset initial prediction formula using the second joint action position and the visual image as input data, and optimize the preset initial prediction formula using a preset loss function to obtain a prediction formula for the robot; A driving module is used to obtain the initial joint action position and the initial visual image of the robot in the joint simulation environment, and input the initial joint action position and the initial visual image into the prediction formula to obtain the target joint action sequence of the robot in the joint simulation environment, and drive the robot to perform corresponding actions based on the target joint action sequence.

[0019] Optionally, in a first implementation of the second aspect of the present invention, the first simulation module includes: the clamping state includes the position, rotation vector, and gripping state of the clamping jaw, and the simulation of the clamping jaw simulation environment is:

[0020] in, represents the gripper simulation environment based on gripper control, represents the first joint action position output in the gripper simulation environment based on gripper control, represents the position and rotation vector of the gripper, Indicates the gripping state of the gripper. 0 means closed, 1 means open. Respectively represent the left arm and right arm of the dual-arm robot, Indicates time.

[0021] Optionally, in a second implementation of the second aspect of the present invention, the second simulation module includes: the visual image includes a left arm image, a right arm image and a side image of the robot, and the simulation of the joint simulation environment is: ,

[0022] in, represents the joint simulation environment based on joint control, represents the second joint action position output in the joint simulation environment based on joint control, represents a color image, Respectively represent the robot's left arm, right arm and side, Indicates time.

[0023] Optionally, in a third implementation of the second aspect of the present invention, the generating module includes: A training unit is used to obtain the second joint action position and visual image output by the joint simulation environment, and generate input data based on the second joint action position and the visual image, and use the input data as training data to train a preset initial prediction formula.

[0024] Optionally, in a fourth implementation manner of the second aspect of the present invention, the training unit includes: The input data is in the form of: ,

[0025] in, represents the input at time t, Indicates that The combined joint action position, for The combined visual image, d The dimension representing the position of the joint action, that is, the dimension of the left arm + the right arm, Indicates Cameras, ; Represents the width × height × number of channels of the visual image, Indicates the time; the preset initial prediction formula is: ,

[0026] in, represents the predicted joint action sequence from time t to time t+k, represents the input data at the tth moment, is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence.

[0027] Optionally, in a fifth implementation of the second aspect of the present invention, the generating module further includes: The optimization unit is used to calculate the error value between the predicted joint motion sequence and the collected joint motion sequence based on a preset loss function, and optimize the preset initial prediction formula based on the error value to obtain the prediction formula of the robot, wherein the preset loss function is: ,

[0028] represents the predicted joint action sequence from time t to time t+k, represents the input joint action sequence, and the prediction formula is: ,

[0029] in, Represents the joint action position at time t represents the visual image of the i-th camera at time t; is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence, and .

[0030] Optionally, in a sixth implementation of the second aspect of the present invention, the driving module includes: The torque calculation unit is used to calculate the control torque of each joint of the robot based on the target joint action sequence through a preset control calculation algorithm, and drive the robot to perform corresponding actions based on each control torque, wherein the preset control calculation algorithm is: ,

[0031] represents the control torque of joint J at the current moment, represents the proportional gain coefficient of the controller, represents the derivative gain coefficient of the controller, represents the predicted action position of joint J at the next moment, represents the actual action position of joint J at the next moment, Represents the actual velocity of joint J at the next moment.

[0032] The third aspect of the present invention provides a robot skill simulation device, which includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to enable the robot skill simulation device to execute the robot skill simulation method as described above.

[0033] A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed by a processor, implement the skill simulation method of the robot as described above.

[0034] In the technical solution provided by the present invention, the state of the gripper when the robot completes the operation task is obtained, and input into the gripper simulation environment for simulation to obtain the first joint action position, the first joint action position is input into the joint simulation environment for simulation to obtain the second joint action position and visual image, and it is used as input data, the preset initial prediction formula is trained and optimized by the preset loss function, the prediction formula is obtained, the initial joint action position of the robot in the joint simulation environment and the initial visual image are input into the prediction formula, the target joint action sequence is obtained, and the robot is driven to perform the corresponding action. This solution predicts the action sequence of the robot by learning joint action information and visual information, reduces the technical difficulty of the robot in learning operation behavior, improves the intelligence of the robot and the generalization ability of skill simulation, and realizes the learning of operation behavior by the robot in the simulation environment, and realizes the robot's ability to adapt and learn autonomously, so that when facing complex and changeable environments or tasks, it can autonomously plan the action trajectory, make flexible responses and adjustments, improve the robot's fine operation ability, reduce the cost of the robot's skill simulation, and improve the efficiency and convenience of the robot's skill simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of a first embodiment of a method for simulating a skill of a robot provided by an embodiment of the present invention; Figure 2 A schematic diagram of a second embodiment of a method for simulating a skill of a robot provided by an embodiment of the present invention; Figure 3 A schematic diagram of a flow chart of a method for simulating robot skills provided by an embodiment of the present invention; Figure 4 A schematic diagram of the input and output of the simulation environment and prediction formula provided by an embodiment of the present invention; Figure 5A schematic diagram of the structure of a robot skill simulation device provided by an embodiment of the present invention; Figure 6 Another structural schematic diagram of a robot skill simulation device provided by an embodiment of the present invention; Figure 7 A schematic diagram of the structure of a robot skill simulation device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In response to the existing robot skill simulation methods, the present application predicts the robot's action sequence by learning joint motion information and visual information, thereby reducing the technical difficulty of the robot in learning operational behaviors, improving the robot's intelligence level and the generalization ability of skill simulation, and realizing the robot's learning of operational behaviors in a simulation environment, so that the robot can have the ability to adapt and learn autonomously, so that when facing complex and changing environments or tasks, it can autonomously plan action trajectories, make flexible responses and adjustments, improve the robot's fine operation capabilities, reduce the robot's skill simulation costs, improve the robot's skill simulation efficiency and convenience, and improve the robot's skill simulation efficiency.

[0037] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0038] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 A schematic diagram of a first embodiment of a method for simulating a skill of a robot provided by an embodiment of the present invention, the method specifically comprises the following steps: 101. Build a simulation environment based on the operation task, parse and split the operation task, and obtain the robot's gripper state when completing the operation task.

[0039] The simulation environment includes a gripper simulation environment based on gripper control and a joint simulation environment based on joint control. This solution does not limit the number of mechanical arms of the robot. When a corresponding number of mechanical arms are added, a corresponding number of positions and images need to be collected. This embodiment takes a dual-arm robot as an example.

[0040] The differences between the gripper simulation environment based on gripper control and the joint simulation environment based on joint control are: control object and function. The control object of the gripper simulation environment is the operation and control of the gripper, which is used to simulate the gripping, releasing and other actions of the gripper, as well as the interaction between the gripper and the object. The control object of the joint simulation environment is the joint part of the robot, which is used to simulate the movement of the robot joints, including rotation and displacement, to achieve the overall movement and posture adjustment of the robot. Modeling and simulation requirements. The gripper simulation environment focuses on the mechanical structure, clamping force, clamping stability and other parameters of the gripper, which is used to simulate the gripping effect of the gripper under different conditions, such as gripping force, gripping speed, gripping accuracy, etc. The joint simulation environment needs to establish an overall model of the robot, including the kinematics, dynamics and other characteristics of the joint, which is used to simulate the motion trajectory, speed, acceleration, etc. of the robot in different scenarios, as well as the interaction between the robot and the environment.

[0041] 102. Input the gripper state into the gripper simulation environment, and perform simulation based on the gripper state to obtain the first joint action position of the robot in the gripper simulation environment.

[0042] The gripper state is input into the gripper simulation environment, and the gripper motion trajectory is determined based on the gripper state when the robot completes the operation task, and the motion trajectory is converted into the joint motion position to obtain the first joint motion position of the robot in the gripper simulation environment. The simulation environment based on gripper control is mainly used to convert the motion trajectory of the end gripper of the dual-arm robot in the simulation environment into the joint motion position to obtain the joint motion position for task teaching operation.

[0043] Specifically, start the gripper simulation software or platform, and load the relevant models of the gripper and the robot, configure the simulation environment to simulate the actual working environment and physical conditions of the gripper, determine the initial state of the gripper, such as the degree of opening and closing, clamping force, etc., pass the initial state of the gripper as input data to the simulation environment, and define a mapping relationship based on the results of the actual robot or simulation experiment. The mapping relationship describes the association between the gripper state and the robot joint action position, and uses the defined mapping relationship to calculate the first joint action position based on the input gripper state.

[0044] In another achievable manner, the desired trajectory that the end gripper of the dual-arm robot needs to follow is determined in a simulation environment. The desired trajectory is a series of posture points, each of which contains the position and posture information of the end gripper in three-dimensional space. The state of the dual-arm robot is initialized in the simulation environment, including the initial position and posture of all joints. For each posture point on the trajectory, an inverse kinematics algorithm is used to calculate the angle or position that each joint of the robot should be at when reaching the posture point. Since the inverse kinematics calculation may produce discontinuous or jittery joint trajectories, it is necessary to smooth the calculated joint trajectory. The smoothing method may include interpolation, filtering, spline curve fitting, etc. Further, the first joint action position in the calculated joint trajectory is input into the simulation environment to observe whether the end gripper of the dual-arm robot can move according to the expected trajectory. If there is a deviation or problem, the inverse kinematics algorithm or smoothing method can be adjusted, and recalculated and verified.

[0045] 103. Input the first joint action position into the joint simulation environment, and perform simulation based on the first joint action position to obtain the second joint action position and visual image of the robot in the joint simulation environment.

[0046] The second joint action position is the state position of each joint of the robot based on joint control output after the robot is simulated in a joint simulation environment, and the visual image includes color images of the left arm, right arm and side of the robot.

[0047] Start the joint simulation software or platform, load the robot model, configure the simulation environment to simulate the physical conditions of the real world, input the action position of the first joint, pass the initial action position of the first joint as input data to the simulation environment, and the simulation environment will simulate the movement of the robot based on the input action position of the first joint. During the simulation process, the robot model will calculate the action position of each joint in the joint simulation environment based on the physical laws and kinematic equations, and configure a virtual camera to capture the movement process of the robot and take visual images of the robot at different time points.

[0048] 104. The preset initial prediction formula is trained using the second joint action position and the visual image as input data, and the preset initial prediction formula is optimized using a preset loss function to obtain a prediction formula for the robot.

[0049] Collect a dataset containing the second joint action position and the corresponding visual image, mark each sample in the dataset so that it contains the expected or target joint action sequence, for example, the joint angle of the next time step, preprocess the second joint action position data, such as normalization or standardization, preprocess the visual image, including scaling, cropping, grayscale, binarization, feature extraction, etc., use the prepared dataset to train the prediction formula, pass the second joint action position and the visual image as input data to the prediction formula in each iteration, and calculate the predicted target joint action sequence, use a preset loss function to calculate the error between the predicted value and the actual value, calculate the gradient of the loss function with respect to the prediction formula parameters through the back propagation algorithm, and use the optimization algorithm to update the parameters to minimize the loss function.

[0050] 105. Obtain the initial joint action position and initial visual image of the robot in the joint simulation environment, and input the initial joint action position and initial visual image into the prediction formula to obtain the target joint action sequence of the robot in the joint simulation environment, and drive the robot to perform corresponding actions based on the target joint action sequence.

[0051] In practical applications, first initialize the robot and the simulation environment, start the joint simulation environment, load the robot model, set initial conditions, such as the robot's initial joint action position (initial posture), obtain the initial joint action position and visual image, obtain the robot's initial joint action position (usually a set of joint angles) from the simulation environment, and if there is a visual sensor simulation, obtain the visual image corresponding to the robot's initial state at the same time, perform necessary preprocessing on the initial joint action position and visual image, such as normalization, denoising, etc., extract features useful for prediction, which may include joint angles, angular velocities, specific target positions in the image, etc., take the preprocessed initial joint action position and visual image features as input, pass them to the prediction formula, use the prediction formula to calculate the target joint action sequence based on the input data, the target joint action sequence is a number of joint action positions, convert the target joint action sequence into instructions that can be understood by the robot controller in the simulation environment, send the instructions to the robot controller in the simulation environment, and drive the robot to perform the corresponding action.

[0052] This solution obtains the gripper state of the robot when it completes the operation task, and inputs it into the gripper simulation environment for simulation to obtain the first joint action position, and inputs the first joint action position into the joint simulation environment for simulation to obtain the second joint action position and visual image, and uses it as input data, and optimizes the preset initial prediction formula through the preset loss function to obtain the prediction formula, and inputs the robot's initial joint action position and initial visual image in the joint simulation environment into the prediction formula to obtain the target joint action sequence, and drives the robot to perform the corresponding action. This solution predicts the robot's action sequence by learning joint action information and visual information, which reduces the technical difficulty of the robot in learning operation behaviors and improves the robot's intelligence and the generalization ability of skill simulation.

[0053] See also Figure 2 , a schematic diagram of a second embodiment of a robot skill simulation method provided by an embodiment of the present invention, the method specifically comprises the following steps: 201. Build a simulation environment based on the operation task, obtain the gripper state of the robot when completing the operation task, and build the robot's gripper simulation environment based on gripper control and a joint simulation system based on joint control.

[0054] In this scheme, a simulation environment and a robot model are built based on the robot's motion environment; in the simulation environment, the gripper trajectory is determined based on the operation task, and the joint movement is determined by inverse motion; learning and training are carried out based on the joint movement to determine the joint movement of the operation task and control the robot movement.

[0055] This embodiment takes a dual-arm robot as an example. Figure 3 , a flow chart of a robot skill simulation method provided by an embodiment of the present invention, firstly, designs and determines the tasks that need to be performed by the dual-arm robot in the simulation environment and the simulation model of the dual-arm robot; constructs a simulation environment based on gripper control and two simulation environments based on joint control of the dual-arm robot for performing teaching operations and obtaining simulation data sets; defines a method for calculating the motion trajectory of the end gripper when the dual-arm robot performs a task in the simulation environment, which is used to calculate and obtain the complete motion trajectory of the end gripper when performing the task; in the simulation environment based on gripper control, inputs the motion trajectory of the end gripper to complete the task, outputs the joint motion position of the dual-arm robot, and then inputs the joint motion position into the simulation environment based on joint control, performs teaching operations, outputs the joint motion position and visual image, and forms a simulation data set for task teaching; in the simulation environment based on joint control, inputs the simulation data set to train the imitation learning algorithm; when performing model reasoning in the simulation environment, inputs the initial joint motion position and visual image of the dual-arm robot in the simulation environment into the model, performs reasoning and outputs the joint motion sequence, and drives the dual-arm robot to perform the action.

[0056] In this example, the left arm and the right arm of the dual-arm robot are both mechanical arms containing a number of joints and grippers. Specifically, the left arm and the right arm of the dual-arm robot are mechanical arms with 6 degrees of freedom, each of which includes a base, 6 joints and a fixture, wherein the angular motion limit range of joints 1, 4, and 6 is 360 degrees, and the angular motion limit range of joints 2, 3, and 5 is 180 degrees, and the joints with different angular motion limit ranges are alternately connected, and the opening and closing limit of the grippers of the left and right arms is 180 degrees. It should be noted that the camera used by the dual-arm robot is a color camera, which can obtain color images of the robot in the process of performing the operation task, involving three cameras of the left arm, the right arm and the side. The side of the dual-arm robot refers to a position similar to the human eye, which is used to capture the image of the overall motion state of the two grippers. The image of the gripper position will move with the gripper movement, and then the picture of the gripper movement is obtained, while the side position is fixed, and the image of the two grippers in the motion process can be observed. Specifically, a gripper simulation environment based on gripper control and a joint simulation environment based on joint control are built in proportion in the simulation environment, and the center of the rectangular table top is used as the origin of the world coordinate system to place the dual-arm robot and the corresponding object simulation model. Simulation cameras are set at the end positions of the left arm and the right arm and on the sides of the dual-arm robot. The simulation cameras are fixed at the end positions of the left arm and the right arm and move with the movement of the left arm and the right arm. A camera is set in the middle of the left arm and the right arm and slightly higher than the side of the length of the robot arm. The image resolutions of the three simulation cameras are all 480×640×3.

[0057] Specifically, the operation task performed by the dual-arm robot in this example is to use the right arm of the dual-arm robot to clamp an object between the two arms on a rectangular table, pass it to the left arm and return to the initial state, and the left arm clamps the object passed by the right arm, places it in a specified position and returns to the initial state. In order to build a simulation environment based on the real physical environment where the operation task is located, first clarify the simulation goal, determine the real physical environment that the simulation environment needs to simulate, such as mechanical structure, physical characteristics, environmental interaction, etc., collect and model real physical data, build a virtual physical environment, build an accurate model for the robot, including mechanical structure, sensor system and control system, import the robot model in the simulation environment, configure physical parameters and sensor data, and simulate the corresponding sensor data based on the sensors used in the real physical environment.

[0058] 202. Input the state of the gripper into the gripper simulation environment, determine the motion trajectory of the gripper based on the state of the gripper when the robot completes the operation task, and convert the motion trajectory into a joint motion position to obtain the first joint motion position of the robot in the gripper simulation environment.

[0059] The simulation environment based on gripper control is mainly used to convert the motion trajectory of the end gripper of the dual-arm robot in the simulation environment into joint motion positions, and obtain the joint motion positions for task teaching operations. Specifically, the current state of the gripper is determined, including the degree of opening and closing of the gripper, the clamping force, etc., and the corresponding gripper state parameters are set in the gripper simulation environment. The operation tasks that the robot needs to complete, such as grasping, placing, and carrying, are analyzed, and the key states that the gripper needs to go through in the task are determined, such as the gripper starts to close, clamps the object, and the gripper starts to open. According to the operation task and the gripper state, the gripper's motion trajectory is planned. The trajectory planning includes the speed, acceleration, and path of the gripper's opening and closing. According to the position, speed, and acceleration of the gripper, the angle, speed, and acceleration of each joint of the robot are calculated using the inverse kinematics calculation method, and the gripper's motion trajectory is converted into the robot's joint motion position. The above processing is performed on each joint in turn to determine the first joint motion position of each joint.

[0060] Furthermore, the trajectory planning method of spatial linear interpolation is used to define all the gripper action positions for the task of the end gripper position. Taking the end joint positions A (xA, yA, zA) and B (xB, yB, zB) of the two key points of the right arm of the dual-arm robot in the gripper simulation environment based on gripper control as an example, A is the previous key point, B is the next key point, and the corresponding time intervals are tA and tB, where N = tB-tA is the number of interpolation points, and the calculation method of the end gripper position of the interpolation point at each time step is: ,

[0061] Specifically, first determine the position and time of the key points, calculate the number of interpolation points and the time interval, perform linear interpolation between points A and B to obtain the position of each interpolation point, create an array of size N+1 (including points A and B) to store the positions of the interpolation points, and calculate the interpolation point positions.

[0062] The formula for outputting the first joint action position of the dual-arm robot is as follows: ,

[0063] In the above formula It indicates that the simulation environment is a simulation environment based on gripper control; It represents the joint action position output in the simulation environment based on gripper control; Represents the position and rotation vector of the end; Indicates the gripping state of the gripper, 0 means closed, 1 means open; They represent the left and right arms of the dual-arm robot respectively; Indicates time.

[0064] 203. Input the first joint action position into the joint simulation environment for simulation, and obtain the second joint action position and visual image of the robot in the joint simulation environment.

[0065] The simulation environment based on joint control is mainly used to perform task teaching operations on the simulation environment based on joint control according to the joint action positions output in the simulation environment based on gripper control, and output joint action positions and visual images to form a simulation data set for task teaching. Specifically, define the tasks to be performed in the simulation environment, such as grabbing objects, moving objects to a specified position, etc., obtain joint action position data from the gripper simulation environment based on gripper control, import the joint action position data in the gripper simulation environment into the joint simulation environment based on joint control, and perform task teaching operations. During the simulation process, if it is found that the robot's action is not in line with expectations, adjust or optimize the joint action position. After the task teaching is completed, the simulation environment will generate a simulation data set containing information such as joint action positions and visual images.

[0066] The calculated complete end gripper motion trajectory when executing the task is input into the joint motion position obtained in the gripper control-based simulation environment. The joint motion position is then input into the joint control-based simulation environment for teaching operations. The joint motion positions of the left and right arms of the dual-arm robot and the visual images of the simulation camera are output, i.e., a 480×640×3 image of 14 dimensions + 3 cameras.

[0067] The formula for outputting joint action positions and visual images is:

[0068] In the above formula It indicates that the simulation environment is a simulation environment based on joint control; Represents the joint action position output in the simulation environment based on joint control; Represents a visual image in RGB, Respectively represent the left arm, right arm and side (the side position between the two arms) of the dual-arm robot; Indicates the moment. The side of the dual-arm robot refers to a position similar to the human eye, which is used to capture the image of the overall motion state of the two grippers. The image at the gripper position will move with the gripper movement, and then obtain the picture of the gripper in motion, while the side position is fixed, and the image of the two grippers in motion can be observed.

[0069] 204. Generate input data based on the second joint action position and the visual image, and input the input data into an initial prediction formula to train the initial prediction formula.

[0070] The joint action positions and visual images output by the joint control-based simulation environment constitute the input data; based on the ACT imitation learning algorithm, the joint action positions and visual images are input to predict the output joint action sequence.

[0071] The second joint action position and visual image output by the joint simulation environment based on joint control form the input data as follows:

[0072] In the above formula, represents the input at time t; Indicates that The joint action position after combination; for The combined visual image; d represents the dimension of the joint action position, that is, the dimension of the left arm + the right arm; Indicates Cameras, ; Represents the width of the visual image ( )×Height( )×Number of channels( ), Indicates time.

[0073] The formula for predicting the output joint action sequence is:

[0074] In the above formula, Represents the predicted joint action sequence from time t to time t+k; Represents the input data at the tth moment; is the decoder parameter, L represents the number of samples, l=0, 1, 2, …, L; is the z latent variable corresponding to the i-th target action sequence.

[0075] 205. Calculate the error value between the predicted joint motion sequence obtained based on the initial prediction formula and the collected input joint motion sequence using a preset loss function, and optimize the initial prediction formula based on the error value to obtain the prediction formula of the robot.

[0076] The loss function is used to measure the difference between the predicted value and the actual value. In the prediction of robot joint motion, common loss functions include mean square error, mean absolute error, etc. The calculation formula of the loss function for predicting the output joint motion sequence in this scheme is:

[0077] In the above formula, the mean square error is used as the loss function to calculate the difference between the predicted value and the collected data. Represents the predicted joint action sequence from time t to time t+k; Represents the input joint action sequence.

[0078] Use the initial prediction formula and input data to calculate the predicted joint action sequence, use the defined loss function to calculate the error between the predicted value and the actual value, and use an optimization algorithm (such as gradient descent, stochastic gradient descent, Adam, etc.) to update the parameters of the prediction formula to minimize the loss value. This usually involves calculating the gradient of the loss function with respect to the model parameters and using these gradients to update the parameters.

[0079] See also Figure 4 , an input and output schematic diagram of the simulation environment and prediction formula provided by an embodiment of the present invention, first, into a simulation environment based on terminal control, and input the data output by the simulation environment based on terminal control into a simulation environment based on joint control for joint teaching operation, output the joint movements based on joint control as positions and visual images, thereby realizing the acquisition of simulation data, input the joint movements based on joint control as positions and visual images into the ACT algorithm for encoding and decoding, and finally output the joint movement sequence, the ACT algorithm,. Specifically, the joint action positions in the simulation data set are obtained for action block division to obtain the target action sequence, and the current joint action position, target action sequence, and visual information are constructed as model training inputs. The CVAE encoder (conditional variational autoencoder) in the ACT imitation learning algorithm is constructed, and the joint action position and target action sequence are input to calculate the z latent variables; the CVAE decoder in the ACT imitation learning algorithm is constructed, including the residual network image encoder, encoder and decoder, and the joint action position, the visual image of the camera and the z latent variables are input. The residual network image encoder is first used to process the visual image of the camera to obtain image features, and then combined with the joint action position and z latent variable data, it is input into the encoder to calculate and output the comprehensive features for prediction, and then decoded by the decoder to output the joint action sequence. It should be noted that the state space of the CVAE encoder in the ACT model learning algorithm, that is, the input includes the joint action positions of the two arms of the dual-arm robot, which is 7+7=14, and the visual images of the three simulation cameras are 480×640×3 color images.

[0080] In practical applications, data preprocessing and action segmentation are first performed to obtain a simulation data set containing joint action positions and visual information. The continuous joint action position sequence is divided into smaller action blocks or target action sequences. The joint action position at the beginning of each action block is selected as the current position, and the joint action position sequence in each action block is used as the target action sequence. For each current joint action position, the corresponding visual image or video frame is obtained. The current joint action position and the target action sequence are used as the input of the CVAE encoder, and the input data is encoded into one or more latent variables z, which capture the potential structure and relationship of the input data. The input of the decoder includes latent variables z, the current joint action position and visual information. The image feature extractor is used to encode the visual image, extract image features, and combine the image features, the current joint action position and the latent variable z, and input them into the encoder. The encoder uses a self-attention mechanism to capture the dependency between the inputs and outputs comprehensive features. The decoder receives the output of the encoder as input and uses self-attention and encoder-decoder attention mechanisms to generate predicted joint action sequences.

[0081] 206. The initial joint action position and initial visual image of the robot in the joint simulation environment are input into the prediction formula to obtain the target joint action sequence.

[0082] In the simulation environment based on joint control, the initial joint position information (7+7=14 dimensions) of the dual arms of the dual-arm robot and the visual images (480×640×3) of the three simulation cameras are obtained and input into the model to predict the joint action sequence (14 dimensions) of the next time step (step 1). The joint action sequence is calculated by the Stable PD control calculation method to calculate the control torque of each joint of the dual-arm machine in the MuJoCo simulation to control the dual-arm robot and drive the dual-arm robot to perform the corresponding action. Then, the predicted joint position information (14 dimensions) of the first step plus the visual images (480×640×3) of the three simulation cameras are used as the input of the second step and input into the model for inference output. Then, the joint action information is applied to the dual-arm robot in the simulation environment scene to drive the dual-arm robot to perform the corresponding action. This step is continuously repeated to drive the dual-arm robot in the simulation environment scene to complete the task.

[0083] During reasoning, the CVAE encoder needs to be removed. Then, the joint action position information of the current dual-arm robot and the visual information of the camera are input into the CVAE decoder in the ACT imitation learning algorithm, and the z latent variable is set to 0. The joint action sequence is output for reasoning, and the final prediction formula is expressed as:

[0084] In the above formula, Represents the joint action position at time t represents the visual image of the i-th camera at time t; is the decoder parameter; L represents the number of samples, l=0, 1, 2, …, L; is the z latent variable corresponding to the i-th target action sequence, and .

[0085] In practical applications, the dual-arm robot is run in a simulation environment to collect its initial joint position information (14 dimensions) and the corresponding three camera visual images (each image is 480×640×3), and the visual images are preprocessed, such as cropping, scaling, and normalization, so as to facilitate input into the neural network. A CVAE model is constructed. The CVAE model includes an encoder and a decoder. The encoder receives the current joint position information (14 dimensions) and conditional information (such as the target position or task description) and outputs the latent variable z. The decoder receives the latent variable z, the current joint position information and conditional information, and predicts the joint action sequence (14 dimensions) of the next time step. A ResNet image encoder (Residual Network, ResNet image encoder is an image feature extractor based on a residual network, which is used to convert the input image into high-level semantic features) is constructed to predict the joint action sequence of the next time step. Use the collected data sets to train the CVAE model and ResNet image encoder. During the training process, the model will learn how to predict the joint action sequence of the next time step from the current state (including joint position and visual information). In the simulation environment, the initial joint position information and visual image are input into the trained model, and the model predicts the joint action sequence of the next time step.

[0086] 207. Based on the target joint action sequence, the control torque of each joint of the robot is calculated through a preset control calculation algorithm, and the robot is driven to perform corresponding actions based on each control torque.

[0087] The control calculation method is used to calculate the control torque of each joint according to the predicted joint action sequence, and the control torque is applied to the robot to drive the robot to perform the corresponding action. The simulated dual-arm robot in the gripper control and joint control simulation environment uses proportional-differential control to control the dual-arm robot joints in the simulation environment to perform actions. The formula is as follows:

[0088] In the above formula, Represents the control torque of joint J at the current moment; represents the proportional gain coefficient of the controller; represents the derivative gain coefficient of the controller; represents the predicted action position of joint J at the next moment; Indicates the actual action position of joint J at the next moment; Represents the actual velocity of joint J at the next moment.

[0089] In practical applications, the robot's dynamic model and target joint action sequence are used to calculate the required joint torque. The target joint angle, velocity and acceleration are obtained by differentiating the target action sequence, and are used as input to calculate the torque required for each joint. In addition to directly using inverse dynamics to calculate the torque, the preset control calculation algorithm can also be combined with other control algorithms to obtain a smoother or more stable torque output. For example, proportional-differential control, proportional-integral-differential control, etc. can be used. It can be understood that the joint torque of the robot usually has certain limitations. After the torque is calculated, it is necessary to ensure that it does not exceed the torque limit of the joint. If the calculated torque exceeds the limit, saturation processing is usually required, that is, the torque is limited to the maximum or minimum allowable value. After calculating and adjusting the torque value, each torque value is sent to the robot's controller, and the motor is driven according to the received torque command, thereby generating the required joint torque and acting on the robot's connecting rod and joint, so that it moves according to the target action sequence, and the robot is driven to perform the corresponding action.

[0090] This scheme predicts the robot's action sequence by learning joint motion information and visual information, reducing the technical difficulty of the robot in learning operational behaviors and improving the robot's intelligence level and the generalization ability of skill simulation.

[0091] The skill simulation method of the robot in the embodiment of the present invention is described above. The skill simulation device of the robot in the embodiment of the present invention is described in detail from the perspective of modular functional entities. Figure 5 , a schematic diagram of a structure of a robot skill simulation device provided by an embodiment of the present invention, the device comprises: A construction module 510 is used to construct a simulation environment based on the operation task, and parse and split the operation task to obtain the gripper state of the robot when completing the operation task, wherein the simulation environment includes a gripper simulation environment based on gripper control and a joint simulation environment based on joint control; A first simulation module 520, configured to input the gripper state into the gripper simulation environment, and perform simulation based on the gripper state to obtain a first joint action position of the robot in the gripper simulation environment; A second simulation module 530 is used to input the joint action position into the joint simulation environment, and perform simulation based on the first joint action position to obtain a second joint action position and a visual image of the robot in the joint simulation environment; A generating module 540, configured to train a preset initial prediction formula using the second joint action position and the visual image as input data, and optimize the preset initial prediction formula using a preset loss function to obtain a prediction formula for the robot; The driving module 550 is used to obtain the initial joint action position and the initial visual image of the robot in the joint simulation environment, and input the initial joint action position and the initial visual image into the prediction formula to obtain the target joint action sequence of the robot in the joint simulation environment, and drive the robot to perform corresponding actions based on the target joint action sequence.

[0092] This solution predicts the robot's action sequence by learning joint motion information and visual information, thereby reducing the technical difficulty of the robot in learning operational behaviors, improving the robot's intelligence level and the generalization ability of skill simulation, and realizing the robot's learning of operational behaviors in a simulation environment, so that the robot can have the ability to adapt and learn autonomously. Therefore, when facing complex and changing environments or tasks, it can autonomously plan action trajectories, make flexible responses and adjustments, improve the robot's fine operation capabilities, reduce the robot's skill simulation costs, and improve the robot's skill simulation efficiency and convenience.

[0093] See also Figure 6 , another structural schematic diagram of a robot skill simulation device provided by an embodiment of the present invention, the device comprises: A construction module 610 is used to construct a simulation environment based on the operation task, and parse and split the operation task to obtain the gripper state of the robot when completing the operation task, wherein the simulation environment includes a gripper simulation environment based on gripper control and a joint simulation environment based on joint control; A first simulation module 620, configured to input the gripper state into the gripper simulation environment, and perform simulation based on the gripper state to obtain a first joint action position of the robot in the gripper simulation environment; A second simulation module 630 is used to input the joint action position into the joint simulation environment, and perform simulation based on the first joint action position to obtain a second joint action position and a visual image of the robot in the joint simulation environment; A generating module 640, configured to train a preset initial prediction formula using the second joint action position and the visual image as input data, and optimize the preset initial prediction formula using a preset loss function to obtain a prediction formula for the robot; The driving module 650 is used to obtain the initial joint action position and the initial visual image of the robot in the joint simulation environment, and input the initial joint action position and the initial visual image into the prediction formula to obtain the target joint action sequence of the robot in the joint simulation environment, and drive the robot to perform corresponding actions based on the target joint action sequence.

[0094] In this embodiment, the first simulation module 620 includes: the clamping state includes the position, rotation vector, and gripping state of the clamping jaw, and the simulation of the clamping jaw simulation environment is:

[0095] in, represents the gripper simulation environment based on gripper control, represents the first joint action position output in the gripper simulation environment based on gripper control, represents the position and rotation vector of the gripper, Indicates the gripping state of the gripper. 0 means closed, 1 means open. Respectively represent the left arm and right arm of the dual-arm robot, Indicates time.

[0096] In this embodiment, the second simulation module 630 includes: the visual image includes the left arm image, the right arm image and the side image of the robot, and the simulation of the joint simulation environment is: ,

[0097] in, represents the joint simulation environment based on joint control, represents the second joint action position output in the joint simulation environment based on joint control, represents a color image, Respectively represent the robot's left arm, right arm and side, Indicates time.

[0098] In this embodiment, the generating module 640 includes: The training unit 641 is used to obtain the second joint action position and visual image output by the joint simulation environment, and generate input data based on the second joint action position and the visual image, and use the input data as training data to train a preset initial prediction formula.

[0099] In this embodiment, the training unit 641 includes: The input data is in the form of: ,

[0100] in, represents the input at time t, Indicates that The combined joint action position, for The combined visual image, d The dimension representing the position of the joint action, that is, the dimension of the left arm + the right arm, Indicates Cameras, ; Represents the width × height × number of channels of the visual image, Indicates the time; the preset initial prediction formula is: ,

[0101] in, represents the predicted joint action sequence from time t to time t+k, represents the input data at the tth moment, is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence.

[0102] In this embodiment, the generating module 640 further includes: The optimization unit 642 is used to calculate the error value between the predicted joint motion sequence and the collected joint motion sequence based on a preset loss function, and optimize the preset initial prediction formula based on the error value to obtain the prediction formula of the robot, wherein the preset loss function is: ,

[0103] represents the predicted joint action sequence from time t to time t+k, represents the input joint action sequence, and the prediction formula is: ,

[0104] in, Represents the joint action position at time t represents the visual image of the i-th camera at time t; is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence, and .

[0105] In this embodiment, the driving module 650 includes: The torque calculation unit 651 is used to calculate the control torque of each joint of the robot based on the target joint action sequence through a preset control calculation algorithm, and drive the robot to perform corresponding actions based on each control torque, wherein the preset control calculation algorithm is: ,

[0106] represents the control torque of joint J at the current moment, represents the proportional gain coefficient of the controller, represents the derivative gain coefficient of the controller, represents the predicted action position of joint J at the next moment, represents the actual action position of joint J at the next moment, Represents the actual velocity of joint J at the next moment.

[0107] This scheme predicts the robot's action sequence by learning joint motion information and visual information, reduces the technical difficulty of the robot in learning operational behaviors, improves the robot's intelligence level and the generalization ability of skill simulation, and realizes the robot's learning of operational behaviors in a simulation environment.

[0108] above Figure 5-6 The skill simulation device of the robot in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The skill simulation device of the robot in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0109] See also Figure 7 As shown, the robot skill simulation device includes a processor 700 and a memory 701. The memory 701 stores machine executable instructions that can be executed by the processor 700. The processor 700 executes the machine executable instructions to implement the above-mentioned robot skill simulation method.

[0110] Further, Figure 7 The robot skill simulation device shown also includes a bus 702 and a communication interface 703 , and the processor 700 , the communication interface 703 and the memory 701 are connected via the bus 702 .

[0111] Among them, the memory 701 may include a high-speed random access memory (RAM), and may also include a non-volatile memory (non-volatile memory), for example, at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 703 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 702 can be an ISA bus, a PCI bus or an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0112] The processor 700 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 700. The above processor 700 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present disclosure can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present disclosure can be directly embodied as a hardware decoding processor to be executed, or a combination of hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 701 , and the processor 700 reads the information in the memory 701 and completes the method steps of the above-mentioned embodiment in combination with its hardware.

[0113] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the various steps of the robot skill simulation method provided in the above-mentioned embodiments.

[0114] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described equipment, devices, and units can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0115] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0116] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A robot skill simulation method, characterized in that: The skill simulation method of the robot comprises: Constructing a simulation environment based on the operation task, and parsing and splitting the operation task to obtain the gripper state of the robot when completing the operation task, wherein the simulation environment includes a gripper simulation environment based on gripper control and a joint simulation environment based on joint control; Inputting the gripper state into the gripper simulation environment, and performing simulation based on the gripper state to obtain a first joint action position of the robot in the gripper simulation environment; Inputting the first joint action position into the joint simulation environment, and performing simulation based on the first joint action position to obtain a second joint action position and a visual image of the robot in the joint simulation environment; The preset initial prediction formula is trained using the second joint action position and the visual image as input data, and the preset initial prediction formula is optimized using a preset loss function to obtain a prediction formula for the robot; The initial joint action position and the initial visual image of the robot in the joint simulation environment are obtained, and the initial joint action position and the initial visual image are input into the prediction formula to obtain the target joint action sequence of the robot in the joint simulation environment, and the robot is driven to perform corresponding actions based on the target joint action sequence.

2. The robot skill simulation method according to claim 1, characterized in that: The gripper state includes the position, rotation vector, and gripping state of the gripper. The simulation of the gripper simulation environment is as follows: ， in, represents the gripper simulation environment based on gripper control, represents the first joint action position output in the gripper simulation environment based on gripper control, represents the position and rotation vector of the gripper, Indicates the gripping state of the gripper. 0 means closed, 1 means open. Respectively represent the left arm and right arm of the dual-arm robot, Indicates time.

3. The robot skill simulation method according to claim 1, characterized in that: The visual image includes a left arm image, a right arm image and a side image of the robot, and the simulation of the joint simulation environment is: ， in, represents the joint simulation environment based on joint control, represents the second joint action position output in the joint simulation environment based on joint control, represents a color image, Respectively represent the robot's left arm, right arm and side, Indicates time.

4. The robot skill simulation method according to claim 1, characterized in that: The training of a preset initial prediction formula using the second joint action position and the visual image as input data includes: The second joint action position and visual image output by the joint simulation environment are obtained, and input data is generated based on the second joint action position and the visual image. The input data is used as training data to train a preset initial prediction formula.

5. The robot skill simulation method according to claim 4, characterized in that: The input data is in the form of: ， in, represents the input at time t, Indicates that The combined joint action position, for The combined visual image, d The dimension representing the position of the joint action, that is, the dimension of the left arm + the right arm, Indicates Cameras, ; Represents the width × height × number of channels of the visual image, Indicates the moment; The preset initial prediction formula is: ， in, represents the predicted joint action sequence from time t to time t+k, represents the input data at the tth moment, is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence.

6. The robot skill simulation method according to claim 1, characterized in that: The method of optimizing the preset initial prediction formula by using the preset loss function to obtain the prediction formula of the robot includes: The error value between the predicted joint motion sequence and the collected joint motion sequence is calculated based on the preset loss function, and the preset initial prediction formula is optimized based on the error value to obtain the prediction formula of the robot, wherein the preset loss function is: ， represents the predicted joint action sequence from time t to time t+k, represents the input joint action sequence, The prediction formula is: ， in, Represents the joint action position at time t represents the visual image of the i-th camera at time t; is the decoder parameter, L represents the number of samples, =0, 1, 2, …, L, is the z latent variable corresponding to the i-th target action sequence, and .

7. The robot skill simulation method according to claim 1, characterized in that: The driving the robot to perform corresponding actions based on the target joint action sequence includes: Based on the target joint action sequence, the control torque of each joint of the robot is calculated by a preset control calculation algorithm, and the robot is driven to perform corresponding actions based on each control torque, wherein the preset control calculation algorithm is: ， represents the control torque of joint J at the current moment, represents the proportional gain coefficient of the controller, represents the derivative gain coefficient of the controller, represents the predicted action position of joint J at the next moment, represents the actual action position of joint J at the next moment, Represents the actual velocity of joint J at the next moment.

8. A robot skill simulation device, characterized in that: The skill simulation device of the robot comprises: A construction module is used to construct a simulation environment based on an operation task, and parse and split the operation task to obtain the gripper state of the robot when completing the operation task, wherein the simulation environment includes a gripper simulation environment based on gripper control and a joint simulation environment based on joint control; A first simulation module is used to input the gripper state into the gripper simulation environment, and perform simulation based on the gripper state to obtain a first joint action position of the robot in the gripper simulation environment; A second simulation module is used to input the joint action position into the joint simulation environment, and perform simulation based on the first joint action position to obtain a second joint action position and a visual image of the robot in the joint simulation environment; A generation module, used to train a preset initial prediction formula using the second joint action position and the visual image as input data, and optimize the preset initial prediction formula using a preset loss function to obtain a prediction formula for the robot; A driving module is used to obtain the initial joint action position and the initial visual image of the robot in the joint simulation environment, and input the initial joint action position and the initial visual image into the prediction formula to obtain the target joint action sequence of the robot in the joint simulation environment, and drive the robot to perform corresponding actions based on the target joint action sequence.

9. A robot skill simulation device, characterized in that: The robot's skill simulation device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the robot's skill simulation device executes the robot's skill simulation method as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the skill simulation method of the robot as described in any one of claims 1-7 is implemented.

Images