Robot cell microoperation method based on imitation learning of security constraint mechanism

Abstract

The invention discloses a robot cell microoperation method based on imitation learning of a security constraint mechanism. The method comprises the following steps: acquiring a teaching video of robot cell microoperation to train an image generation network and an attention network containing a reconstruction and skill information distribution loss function, and extracting discrete potential codes; constructing a robot cell operation model, and constructing safety action constraints according to the discrete potential codes; training the model through a teaching video and a robot action track; when autonomous cell micro-operation is carried out, operation video frames are obtained in real time, the operation video frames and the previous action are jointly processed through a model, then the next action is predicted, and then the robot is controlled to execute. According to the method, by introducing a skill information representation technology and a security constraint mechanism, the problems of multi-task coupling, long-time operation precision and security in a micro-operation task are solved, the success rate is high, operation errors and accumulative errors are effectively reduced, and a complex cell operation task can be executed in a real physical environment.

Description

Technical Field

[0001] The invention relates to a robot cell micro-operation method, relates to the field of robot technology and biological cell operation, and specifically to a robot cell micro-operation method based on imitation learning of a safety constraint mechanism. Background Art

[0002] In the field of robotics, especially in cell micromanipulation tasks, such as delicate operations such as peeling off cell membranes, there is a demand for long-term, high-precision robot operations. These tasks not only require robots to have high dexterity and stability, but also to ensure safety during operation to avoid damage to fragile cell structures. Although imitation learning provides an effective way to control robots by imitating taught behaviors to address this challenge, existing algorithms are prone to accumulated errors when dealing with long-term tasks, leading to operational failures. In addition, these algorithms have limitations in adapting to the deformation and uncertainty of soft tissues such as cells, and face challenges in precise coordination and extended motion planning in micro-scale operations. Therefore, it is necessary to develop a new imitation learning algorithm that can improve the success rate of cell micromanipulation tasks by accurately representing taught skill information and enforcing safe action constraints, especially in scenarios that require long-term operation and high-precision control. Summary of the invention

[0003] In order to solve the problems existing in the background technology, the present invention provides a robot cell micro-manipulation method based on imitation learning of safety constraint mechanism. The present invention aims to quantify the representation of teaching skill information and extract safety action constraints by extracting the potential discrete codes of teaching video frames and modeling the distribution of these codes, and finally predict the action by fusing the action of the previous time step, and execute the action that meets the safety constraints, so as to effectively suppress the compound error.

[0004] The technical solution adopted by the present invention is:

[0005] The robot cell micro-manipulation method based on imitation learning of a safety constraint mechanism of the present invention comprises:

[0006] Step 1: Obtain the teaching video of the robot cell micro-manipulation and extract the video frames as the training set to train the image generation network and the attention network containing the reconstruction and skill information distribution loss functions, obtain the trained image generation network and the encoder of the attention network, and extract the discrete latent code of each teaching video frame.

[0007] Step 2: The trained image generation network, the encoder of the attention network and the multi-layer perceptron MLP are jointly constructed into a robot cell operation model, and the safe action constraints of the robot cell operation model are constructed according to each discrete latent code. The representation of the teaching skill information is quantified by calculating the log-likelihood of the latent discrete code, so as to extract the safe action constraints; the robot cell operation model is trained through the video frames of the teaching video and the action trajectory of the robot cell micro-operation in the teaching video until the loss function of the robot cell operation model converges, and the trained robot cell operation model is obtained.

[0008] Step 3: When the robot performs autonomous cell micro-operation, the video frames of the robot's autonomous cell micro-operation are obtained in real time and input into the trained robot cell operation model together with the robot action at the previous moment. The robot action at the next moment is output while satisfying the safety action constraints, and then the robot is controlled to perform the action to realize robot cell micro-operation.

[0009] In the step 1, the teaching video of the robot cell micro-operation is a video of the robot performing micro-operation on the cell according to a preset action, specifically a video of controlling the remote robot's manipulator to perform cell micro-operation according to a preset action through human teleoperation; the image generation network and the attention network respectively use vector quantization and generative adversarial network VQ-GAN (Vector Quantized Generative Adversarial Network) and autoregressive network Transformer, and the training set is input into the vector quantization and generative adversarial network VQ-GAN for training to obtain the trained vector quantization and generative adversarial network VQ-GAN, and at the same time, the encoder of the vector quantization and generative adversarial network VQ-GAN compresses each teaching video frame to obtain a discrete potential code z of each teaching video frame; each discrete potential code z is input into the autoregressive network Transformer for training, and the distribution of the discrete potential code z is modeled by the autoregressive network Transformer, and the likelihood value of each discrete potential code z is maximized until the reconstruction and skill information distribution loss function converges to obtain the trained autoregressive network Transformer.

[0010] The autoregressive network Transformer optimizes the distribution of discrete potential codes z based on maximum likelihood estimation. The optimization objective function of the autoregressive network Transformer is as follows:

[0011]

[0012] Where θ represents the model parameters of the autoregressive network Transformer; T represents the number of time steps of the teaching video; Z represents the number of dimensions of the discrete potential code; logp θ () represents log-likelihood; represents the encoding value at the i-th position in the discrete potential code at the t-th time, represents the code value at the 1st to i-1th position in the discrete potential code at the tth time; z max(1,t-k):t-1 represents the potential discrete coding from max(1,tk) to t-1 subsequence, k represents the frame length in the subsequence.

[0013] The reconstruction and skill information distribution loss function of the autoregressive Transformer network as follows:

[0014]

[0015] Among them, Cross_entropy() represents the cross entropy loss between the predicted value and the true value, that is, the reconstruction loss; t and Represent the true value and predicted value of the discrete latent code at time t respectively; represents the predicted value of the code value at the i-th position in the discrete potential code at the t-th time, Represents the predicted value of the code value at the 1st to i-1th positions in the discrete potential code at time t.

[0016] In the step 2, the safety action constraints are as follows:

[0017]

[0018] Among them, lnp θ () represents log-likelihood; represents the discrete latent code z at time t+1 t+1 The predicted value, z t-k:t represents the discrete potential code from time tk to time t, k represents the frame length in the subsequence; λ s represents the threshold parameter of the safety constraint; η t+1 represents the safety constraint value at time t+1 during the training process; M represents the total number of samples of potential discrete coding of the teaching video; lnp θ,m () represents the log-likelihood value of the potential discrete code.

[0019] In the step 2, when the robot cell operation model is trained, the video frames of the teaching video and the motion trajectory of the robot cell micro-operation in the teaching video are input into the robot cell operation model. The motion trajectory of the robot cell micro-operation in the teaching video is the motion trajectory of the robot's mechanical arm end and the micro-gripper. The video frames of the teaching video are sequentially input into the encoders of the trained image generation network and the attention network. After processing, the results are fused with the motion trajectory of the robot cell micro-operation in the teaching video, and then the fusion results are input into the multi-layer perceptron MLP for training. During training, the network parameters of the encoder of the attention network are frozen, and the image generation network and the multi-layer perceptron MLP are trained.

[0020] In step 2, the loss function of the robot cell operation model is The details are as follows:

[0021]

[0022] Where MSE() represents mean square error; a t and They represent the real action in the action trajectory of the robot cell micro-manipulation at time t and the predicted action output by the robot cell manipulation model; and λ skill They represent the reconstruction and skill information distribution loss functions of the autoregressive Transformer network and the balance weight parameters of its skill information loss, respectively.

[0023] In step 3, the robot action at the next moment currently output by the trained robot cell operation model is The input is sent to the robot controller, and the controller will determine the robot's next action. The trajectory smoothing method is used for smoothing, and then the robot is controlled to perform the smoothed action.

[0024] The electronic device of the present invention comprises: a memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the method as described above.

[0025] The computer-readable storage medium of the present invention stores program data thereon, and when the program data is executed by a processor, the method described above is implemented.

[0026] The beneficial effects of the present invention are:

[0027] 1. The method proposed in the present invention extracts potential discrete codes by imitating the teaching operation video frames and uses the autoregressive transformer model to model, which can effectively imitate the teaching skills, improve the accuracy and safety of the robot in cell micro-manipulation tasks, and reduce the risk of damage to cells.

[0028] 2. The present invention quantifies the teaching skill information by calculating the log-likelihood of the potential discrete code and extracts the safe action constraints. The present invention can effectively suppress the common compound errors in long-term tasks when executing actions and improve the success rate of the tasks.

[0029] 3. The safety constraint mechanism integrated in the method of the present invention can be dynamically adjusted during operation to ensure that the operation of the robot remains within a safe and effective range even when faced with unknown or changing environmental conditions.

[0030] In summary, by introducing skill information representation technology and safety constraint mechanism, the problems of multi-task coupling, long-term operation accuracy and safety in micro-manipulation tasks are solved. The algorithm shows a high success rate in the embryonic cell membrane peeling task and effectively reduces operational errors and cumulative errors. The present invention can perform complex cell manipulation tasks in a real physical environment and is expected to promote the development of robot skill learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0032] Figure 2 : is a schematic diagram of the overall structure of the robot cell operation model of the present invention, wherein: Figure 2 (a) is a schematic diagram of the encoder and decoder architecture of vector quantization and generative adversarial network VQ-GAN. Figure 2 (b) is a schematic diagram of the autoregressive network Transformer model. Figure 2 (c) is a schematic diagram of the network structure of the robot cell operation model including a multilayer perceptron (MLP) for controlling the left and right mechanical arms and the micro-gripper of the robot during specific implementation;

[0033] Figure 3 It is a schematic diagram of the mechanical structure of the micro-gripper used in the specific implementation of the present invention;

[0034] Figure 4 It is a schematic diagram of the change of loss function during training, where Figure 4 (a) is a schematic diagram of the loss function during vector quantization and generative adversarial network VQ-GAN training. Figure 4 (b) is a diagram of the loss during the training process of the autoregressive network Transformer. Figure 4(c) is a schematic diagram of the loss function during the behavior cloning training process. Figure 4 (b) is a schematic diagram of the normalized log-likelihood value of the training data set after behavior cloning;

[0035] Figure 5 It is a schematic diagram of the process of controlling the robot system to perform cell membrane tearing in the specific implementation of the present invention;

[0036] In the figure: 1. micro claw, 2. Bragg grating, 3. optical fiber, 4. moving sleeve, 5. rotating shaft, 6. first DC servo motor, 7. second DC servo motor, 8. photoelectric gate, 9. cylindrical propulsion cam, 10. rotating gear, 11. sleeve push rod, 12. rotating shaft rotating gear, 13. base. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 1 As shown, the robot cell micro-manipulation method based on imitation learning of the safety constraint mechanism of the present invention is as follows:

[0039] Firstly, the teaching video of robot cell micro-manipulation is obtained and the video frames are extracted as the training set to train the image generation network and the attention network containing the reconstruction and skill information distribution loss function, and the encoders of the trained image generation network and attention network are obtained, and the discrete potential code of each teaching video frame is extracted; the teaching video of robot cell micro-manipulation is a video of the robot performing micro-manipulation on cells according to preset actions, specifically, a video of controlling the remote robot's manipulator to perform cell micro-manipulation according to preset actions through human teleoperation; the image generation network and the attention network use vector quantization and generative adversarial network VQ-GAN and autoregressive network Transformer respectively, and the training set After input into vector quantization and generative adversarial network VQ-GAN for training, the trained vector quantization and generative adversarial network VQ-GAN are obtained. At the same time, the encoder of vector quantization and generative adversarial network VQ-GAN compresses each teaching video frame to obtain the 8×8 discrete latent code z of each teaching video frame; each discrete latent code z is input into the autoregressive network Transformer for training, and the distribution of discrete latent code z is modeled by the autoregressive network Transformer. The likelihood value of each discrete latent code z is maximized until the reconstruction and skill information distribution loss function converges to obtain the trained autoregressive network Transformer.

[0040] The autoregressive network Transformer optimizes the distribution of discrete potential code z based on maximum likelihood estimation. The optimization objective function of the autoregressive network Transformer is as follows:

[0041]

[0042] Where θ represents the model parameters of the autoregressive network Transformer; T represents the number of time steps of the teaching video; Z represents the number of dimensions of the discrete potential code; logp θ () represents log-likelihood; represents the code value at the i-th position in the 8×8 discrete potential code at the t-th time, represents the code value at the 1st to i-1th position in the discrete potential code at the tth time; z max(1,t-k):t-1 represents the potential discrete coding from max(1,tk) to t-1 subsequence, k represents the frame length in the subsequence; Indicates z max(1,t-k):t-1 as well as After given The log-likelihood of .

[0043] Reconstruction and skill information distribution loss function for autoregressive Transformer networks as follows:

[0044]

[0045] Among them, Cross_entropy() represents the cross entropy loss between the predicted value and the true value, that is, the reconstruction loss; t and Represent the true value and predicted value of the discrete latent code at time t respectively; represents the predicted value of the code value at the i-th position in the 8×8 discrete potential code at the t-th time, Represents the predicted value of the code value at the 1st to i-1th position in the discrete potential code at time t; Maximum likelihood estimation for latent coding is optimized for skill information distribution loss.

[0046] like Figure 2 (a) Figure 2 (b) and Figure 2As shown in (c), the trained image generation network, the encoder of the attention network and the multi-layer perceptron MLP are then jointly constructed as a robot cell operation model, and the safe action constraints of the robot cell operation model are constructed according to each discrete potential code. The representation of the teaching skill information is quantified by calculating the logarithmic likelihood of the potential discrete code, so as to extract the safe action constraints; the robot cell operation model is trained through the video frames of the teaching video and the action trajectory of the robot cell micro-operation in the teaching video until the loss function of the robot cell operation model converges, and the trained robot cell operation model is obtained; the safe action constraints are as follows:

[0047]

[0048] Among them, lnp θ () represents log-likelihood; represents the discrete latent code z at time t+1 t+1 The predicted value, z t-k:t represents the discrete potential code from time tk to time t, k represents the frame length in the subsequence; λ s represents the threshold parameter of the safety constraint; η t+1 represents the safety constraint value at time t+1 during the training process; M represents the total number of samples of potential discrete coding of the teaching video; lnp θ,m () represents the log-likelihood value of the potential discrete code, Given the potential discrete code corresponding to the subsequence The log-likelihood value, lnp θ,m (z t+1 |z t-k:t ) represents the potential discrete code corresponding to the subsequence. t+1 The log-likelihood value of .

[0049] When the robot cell operation model is trained, the video frames of the teaching video and the motion trajectory of the robot cell micro-operation in the teaching video are input into the robot cell operation model. The motion trajectory of the robot cell micro-operation in the teaching video is the motion trajectory of the robot's mechanical arm end and the micro-gripper. The video frames of the teaching video are sequentially input into the encoders of the trained image generation network and the attention network. After processing, the results are fused with the motion trajectory of the robot cell micro-operation in the teaching video, and then the fusion results are input into the multi-layer perceptron MLP for training. During training, the network parameters of the encoder of the attention network are frozen, and the image generation network and the multi-layer perceptron MLP are trained.

[0050] The robot cell operation model is trained, i.e., behavior cloning is performed. The teaching video frames and motion trajectories are used as input. The teaching video frames are compressed into discrete latent codes through the VQ-GAN encoder. The discrete latent code sequence is used as input. The autoregressive Transformer is used to extract features, and the features are fused with the action trajectory of the previous time step. The multi-layer perceptron MLP is used to predict the action at the next moment. The VQ-GAN encoder network parameters are frozen, the autoregressive Transformer network and the multi-layer perceptron MLP network are trained, and the network parameters are saved.

[0051] Loss function of the robot cell manipulation model The details are as follows:

[0052]

[0053] Where MSE() represents mean square error; a t and They represent the real action in the action trajectory of the robot cell micro-manipulation at time t and the predicted action output by the robot cell manipulation model; and λ skill They represent the reconstruction and skill information distribution loss functions of the autoregressive Transformer network and the balance weight parameters of its skill information loss, respectively.

[0054] Finally, when the robot performs autonomous cell micro-operation, the video frame of the robot's autonomous cell micro-operation is obtained in real time and input into the trained robot cell operation model together with the robot action at the previous moment. The robot action at the next moment is output while satisfying the safety action constraints. The input is sent to the robot controller, and the controller will determine the robot's next action. The trajectory smoothing method is used for smoothing, and then the robot is controlled to execute the smoothed action to realize the robot cell micro-manipulation.

[0055] The trajectory smoothing method used when performing actions is as follows:

[0056]

[0057] Among them, a t and a t+1 They represent the smoothed robot actions at the current and next moments respectively; γ represents the smoothing parameter.

[0058] like Figure 3As shown, the hardware structure of the robot system used in the specific implementation of the present invention is mainly composed of an inverted microscope, a microscope operating robot arm, a camera, a force feedback device and a micro-gripper. Among them, the inverted microscope is IXplore Standard, equipped with an XY electric stage, and its maximum magnification can reach 400x through the eyepiece and the objective lens. The robot arm uses TransferMan, which has three degrees of freedom and a motion resolution of 0.05μm. The camera W160 is installed on the inverted microscope to capture image information at a rate of 60 frames per second. Two Phantom Touch devices are used to realize force feedback of the end effector. Two structurally symmetrical micro-grippers are designed for dexterous cell manipulation and integrated into the left and right robot arms, such as Figure 3 As shown, the micro-gripper includes a micro-claw 1, two Bragg gratings 2, an optical fiber 3, a moving sleeve 4, a rotating shaft 5, a first DC servo motor 6, a second DC servo motor 7, a photoelectric gate 8, a cylindrical advancing cam 9, a rotating gear 10, a sleeve push rod 11, a rotating shaft rotating gear 12 and a base 13. The photoelectric gate 8, the cylindrical advancing cam 9, the rotating gear 10, the sleeve push rod 11 and the rotating shaft rotating gear 12 are mounted on one side of the base 13, and the rotating shaft 5, the first DC servo motor 6 and the second DC servo motor 7 are mounted on the other side of the base 13. The output shaft of the first DC servo motor 6 is synchronously connected to the central axis of the rotating gear 10, one end of the rotating shaft 5 is synchronously connected to the central end of the rotating shaft rotating gear 12, and the sleeve push rod 1 1 is synchronously connected to the other end of the center of the rotating shaft rotating gear 12, the rotating gear 10 is meshed with the rotating shaft rotating gear 12, the cylindrical advancing cam 9 is pressed on the sliding sleeve push rod 11, the central axis of the cylindrical advancing cam 9 is synchronously connected to the output shaft of the second DC servo motor 7, the photoelectric gate 8 is installed on the side close to the cylindrical advancing cam 9 and is used for the reset operation of the cylindrical advancing cam, the cylindrical advancing cam 9, the rotating gear 10 and the rotating shaft rotating gear 12 are arranged in sequence, one end of the moving sleeve 4 is synchronously connected and installed at the center of the other end of the rotating shaft 5, two Bragg gratings 2 and optical fibers 3 are arranged on the moving sleeve 4, and the micro claw 1 is installed at the other end of the moving sleeve 4; the micro gripper has two degrees of freedom, namely, grasping action and wrist-like rotational movement. The axial rotation is realized by the first DC servo motor 6 driving a group of rotating gears 10, and the grasping action is completed by the second DC servo motor 7 pushing the moving sleeve 4 through the cylindrical advancing cam 9. Three Bragg gratings FBG (Fiber Bragg Grating) 2 are integrated on the moving sleeve 4 to sense bending deformation, provide force sensing and collision detection to prevent damage to the micro gripper. In terms of software control, the software and control parts of the system are integrated into the Robot Operating System (ROS) Melodic running on the Ubuntu 18.04 operating system, and the computer configuration is Intel Core i9-10980Xe CPU and NVIDIA TITAN Xp.

[0059] In the specific implementation of the present invention, zebrafish embryonic cells are used as experimental objects. These cells are placed in a culture dish with added nutrient solution to maintain their activity, thereby ensuring that the cells can maintain a normal physiological state during the experimental operation, facilitating precise cell micro-operation experimental research.

[0060] The entire cell micromanipulation task mainly includes three subtasks: pushing cells, grabbing cells, and tearing cell membranes. In the cell pushing subtask, a single end effector pushes the cell, which requires relatively low dexterity and collaboration. The main purpose is to move the cell to a suitable operating position, such as the center of the field of view. The cell grabbing subtask requires the use of dual end effectors to grab the cell membrane. This process requires high dexterity and collaboration, and it is necessary to ensure that only the cell membrane is grabbed to avoid damaging the inside of the cell to ensure the activity of the cell. The cell membrane tearing subtask is to use the left and right microgrippers to slowly separate and tear the cell membrane to release the embryo. The tearing and squeezing of the membrane in this process requires the highest dexterity and collaboration, and the interaction process between the robot and the cell is also the most complex.

[0061] During the teaching operation data collection stage, the cell operation teaching was performed by controlling the robot through a remote operation system. The teaching used force feedback equipment to accurately manipulate the robotic arm and micro-gripper to complete the above three subtasks based on the cell image under the microscope. Throughout the process, the Robot Operating System (ROS) platform recorded a variety of data including video information, end effector trajectory, axial rotation angle, grasping action, and multi-joint robotic arm trajectory. Finally, 175 valid demonstration samples were collected, with a total duration of 186.1 minutes.

[0062] The loss function of the network during training changes as follows Figure 4 (a) Figure 4 (b) Figure 4 (c) and Figure 4 As shown in (d). Figure 4 As shown in (a), the loss function of the VQ-GAN network converges, indicating that the network can effectively compress image data into discrete potential codes z. Figure 4 As shown in (b), the reconstruction and skill information distribution loss function of the Transformer network Convergence. Figure 4 As shown in (d), it is proved that in the cell operation teaching sample data, the log-likelihood data gradually increases as the operation proceeds, further proving that the network effectively represents the operation skill information and proving the rationality of setting the skill information as a safe operation constraint. Figure 4 (c) shows the loss function of the robot cell manipulation model The effective convergence proves that the network can predict the correct cell operation actions.

[0063] In order to verify the effectiveness of the VQ-GAN encoder, it was replaced with the encoders of the residual network ResNet18 (ResidualNetwork 18), the convolutional neural network MobileNetV2 (Inverted Residuals and LinearBottlenecks) and the convolutional neural network VGG16 (Visual Geometry Group 16-layer network) network, and the normalized log-likelihood values ​​were calculated for the subtasks in the cell membrane tearing process (pushing cells, grabbing cells, and cell membrane tearing). The experimental results show that the VQ-GAN encoder has shown significant superiority in all subtasks. Specifically, the normalized log-likelihood value of VQ-GAN in the cell-pushing task is 0.74±0.07, which is significantly higher than 0.39±0.11 of ResNet18, 0.27±0.06 of MobileNetV2, and 0.44±0.15 of VGG16. In the cell-grasping task, the normalized log-likelihood value of VQ-GAN is 0.86±0.10, which is much higher than other encoders (ResNet18: 0 .21±0.08, MobileNetV2: 0.19±0.13, VGG16: 0.28±0.09; in the cell tearing task, the normalized log-likelihood value of VQ-GAN is 0.82±0.12, which is also significantly higher than other encoders (ResNet18: 0.33±0.16, MobileNetV2: 0.25±0.07, VGG16: 0.26±0.12). These data fully demonstrate the excellent performance of the VQ-GAN encoder in the cell tearing task, and its efficient feature extraction and representation capabilities give it a significant advantage in handling complex tasks.

[0064] Finally, when the robot operates autonomously, it acquires video frames in real time and repeats the above encoding, feature extraction, action prediction and constraint verification process, and performs operations according to the safety action constraints to ensure the safety and accuracy of the robot operation. s is set to 0.8, and the smoothing parameter γ is set to 0.2. Figure 5 As shown, the robot controls the end effector to complete four zebrafish embryo cell membrane tearing processes. The robot successfully completes four zebrafish embryo cell membrane tearing operation tests. The embryos are intact after membrane tearing, proving the effectiveness of the robot cell micro-manipulation method based on imitation learning of the safety constraint mechanism.

[0065] In order to verify the superiority of the proposed method, it was compared with existing imitation learning methods, including behavioral cloning BC (Behavioral Cloning), action chunking ACT (Action Chunking with Transformers), visual imitation learning VINN (Visual Imitation Learning with Neural Networks), and diffusion models Diffusion. Each method was tested 30 times, and the success rate of the subtask stage, average success rate, and final success rate of the cell membrane tearing task were counted. The specific data are as follows: the success rate of the BC method in the task of pushing cells was 93.3% (28 / 30), the success rate of grabbing cells was 32.1% (9 / 28), the success rate of cell membrane tearing was 66.7% (6 / 9), the average success rate was 64.0%, and the final success rate was 20.0% (6 / 30); the success rate of the ACT method in the task of pushing cells was 96.7% (29 / 30), the success rate of grabbing cells was 58.6% (17 / 29), the success rate of cell membrane tearing was 76.5% (13 / 17), the average success rate was 77.2%, and the final success rate was 43.3% (13 / 30); the success rate of the VINN method in the task of pushing cells was 96.7% (29 / 30), the success rate of grabbing cells was 31.0% ( 9 / 29), the success rate of cell membrane tearing was 22.2% (2 / 9), the average success rate was 50.0%, and the final success rate was 6.67% (2 / 30); the success rate of the Diffusion method in the task of pushing cells was 76.7% (23 / 30), the success rate of grabbing cells was 26.1% (6 / 23), the success rate of cell membrane tearing was 0 (0 / 6), the average success rate was 34.3%, and the final success rate was 0 (0 / 30); the success rate of the proposed method in the task of pushing cells was 96.7% (29 / 30), the success rate of grabbing cells was 82.8% (24 / 29), the success rate of cell membrane tearing was 79.2% (19 / 24), the average success rate was 86.2%, and the final success rate was 63.3% (19 / 30). These data show that the proposed method is significantly better than existing methods in all key indicators. Its success rate in the cell pushing task is comparable to that of ACT and VINN, but its success rate in the cell grabbing and cell membrane tearing tasks is significantly higher than other methods. Both the average success rate and the final success rate are much higher than other methods, which fully demonstrates its excellent performance, high efficiency and robustness in the cell membrane tearing task, and shows its significant technical advantages in complex task processing.

[0066] Finally, it should be noted that the above embodiments and explanations are only used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope disclosed in the technical solution of the present invention, which should be included in the scope of protection of the claims of the present invention.

Claims

1. A robot cell micro-manipulation method based on imitation learning of safety constraint mechanism, characterized in that: include: Step 1: Obtain the teaching video of the robot cell micro-manipulation and extract the video frames as the training set to train the image generation network and the attention network containing the reconstruction and skill information distribution loss functions, obtain the encoders of the trained image generation network and attention network, and extract the discrete potential code of each teaching video frame; Step 2: The trained image generation network, the encoder of the attention network, and the multi-layer perceptron MLP are jointly constructed into a robot cell operation model, and the safe action constraints of the robot cell operation model are constructed according to each discrete potential code; the robot cell operation model is trained through the video frames of the teaching video and the action trajectory of the robot cell micro-operation in the teaching video until the loss function of the robot cell operation model converges, and the trained robot cell operation model is obtained; Step 3: When the robot performs autonomous cell micro-operation, the video frames of the robot's autonomous cell micro-operation are obtained in real time and input into the trained robot cell operation model together with the robot action at the previous moment. The robot action at the next moment is output while satisfying the safety action constraints, and then the robot is controlled to perform the action to realize robot cell micro-operation.

2. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 1 is characterized in that: In the step 1, the teaching video of the robot cell micro-operation is a video of the robot performing micro-operation on the cell according to the preset action; the image generation network and the attention network respectively use vector quantization and generative adversarial network VQ-GAN and autoregressive network Transformer, and the training set is input into the vector quantization and generative adversarial network VQ-GAN for training to obtain the trained vector quantization and generative adversarial network VQ-GAN, and at the same time, the encoder of the vector quantization and generative adversarial network VQ-GAN compresses each teaching video frame to obtain the discrete potential code z of each teaching video frame; each discrete potential code z is input into the autoregressive network Transformer for training, and the likelihood value of each discrete potential code z is maximized until the reconstruction and skill information distribution loss function converges to obtain the trained autoregressive network Transformer.

3. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 2 is characterized in that: The autoregressive network Transformer optimizes the distribution of discrete potential codes z based on maximum likelihood estimation. The optimization objective function of the autoregressive network Transformer is as follows: Where θ represents the model parameters of the autoregressive network Transformer; T represents the number of time steps of the teaching video; Z represents the number of dimensions of the discrete potential code; logp θ () represents log-likelihood; represents the encoding value at the i-th position in the discrete potential code at the t-th time, represents the code value at the 1st to i-1th position in the discrete potential code at the tth time; z max(1,t-k):t-1 represents the potential discrete coding from max(1,tk) to t-1 subsequence, k represents the frame length in the subsequence.

4. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 3 is characterized in that: The reconstruction and skill information distribution loss function of the autoregressive Transformer network as follows: Among them, Cross_entropy() represents the cross entropy loss between the predicted value and the true value, that is, the reconstruction loss; t and Represent the true value and predicted value of the discrete latent code at time t respectively; represents the predicted value of the code value at the i-th position in the discrete potential code at the t-th time, Represents the predicted value of the code value at the 1st to i-1th positions in the discrete potential code at time t.

5. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 1 is characterized in that: In the step 2, the safety action constraints are as follows: Among them, lnp θ () represents log-likelihood; represents the discrete latent code z at time t+1 t+1 The predicted value, z t-k:t represents the discrete potential code from time tj to time t, k represents the frame length; λ s represents the threshold parameter of the safety constraint; η t+1 represents the safety constraint value at time t+1 during the training process; M represents the total number of samples of potential discrete coding of the teaching video; lnp θ,m () represents the log-likelihood value of the potential discrete code.

6. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 1 is characterized in that: In the step 2, when the robot cell operation model is trained, the video frames of the teaching video and the motion trajectory of the robot cell micro-operation in the teaching video are input into the robot cell operation model, and the video frames of the teaching video are sequentially input into the encoders of the trained image generation network and the attention network. After processing, the results are fused with the motion trajectory of the robot cell micro-operation in the teaching video, and then the fusion results are input into the multi-layer perceptron MLP for training. During training, the network parameters of the encoder of the attention network are frozen, and the image generation network and the multi-layer perceptron MLP are trained.

7. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 1 is characterized in that: In step 2, the loss function of the robot cell operation model is The details are as follows: Where MSE() represents mean square error; a t and They represent the real action in the action trajectory of the robot cell micro-manipulation at time t and the predicted action output by the robot cell manipulation model; and λ skill They represent the reconstruction and skill information distribution loss functions of the autoregressive Transformer network and the balance weight parameters of its skill information loss, respectively.

8. The robot cell micro-manipulation method based on imitation learning of safety constraint mechanism according to claim 1 is characterized in that: In step 3, the robot action at the next moment currently output by the trained robot cell operation model is The input is sent to the robot controller, and the controller will determine the robot's next action. The trajectory smoothing method is used for smoothing, and then the robot is controlled to perform the smoothed action.

9. An electronic device, characterized in that: include: A memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having program data stored thereon, characterized in that: When the program data is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images