End-to-end trajectory planning algorithm for magnetic adsorption cleaning robot

Abstract

The invention discloses an end-to-end trajectory planning algorithm for a magnetic adsorption cleaning robot, and aims to improve the autonomous navigation and cleaning efficiency of the robot in a complex environment. According to the algorithm, firstly, through a multi-mode perception backbone network, visual, ultrasonic and laser radar data are fused by using a Transform model, and environment features are extracted; and then, based on Decoder design of Transform, robot control variables (linear velocity, angular velocity and electromagnet opening) are embedded into a decoding process, and accurate trajectory planning and electromagnet control are realized. In order to ensure the stability and performance of the model, a step-by-step training strategy is adopted, a sensing network parameter training decoder is firstly fixed, and then end-to-end joint training is carried out. By introducing position coding, a regularization method, an early stop strategy and a gradient cutting technology, the generalization ability and training efficiency of the model are improved. The method has remarkable advantages in the aspects of multi-modal data fusion, environment feature representation, trajectory planning and control and the like, and is suitable for cleaning operation in various complex environments.

Description

Technical Field

[0001] The invention belongs to the field of intelligent robot trajectory planning, and in particular relates to an end-to-end trajectory planning algorithm for a magnetic adsorption cleaning robot. Background Art

[0002] Intelligent cleaning robots using magnetic adsorption technology perform well in maintenance and cleaning tasks on vertical or inclined metal surfaces, and can significantly reduce the cost of cleaning large storage tank equipment. The robot's ability to clean autonomously is directly related to the adaptability of the trajectory planning algorithm. Implementing path planning inside the oil sump of a large oil storage tank is significantly different from ideal cylindrical navigation. The internal structure of large oil storage facilities is complex, and there are obstacles such as valves, sensors, and reinforcements that affect the robot's movement; during the cleaning process, the environment will change dynamically due to changes in the medium on the inner surface of the tank and the dripping of uncleaned oil. The above factors have challenged the trajectory planning algorithm of the intelligent agent, and have placed higher requirements on the cumulative error handling capability, real-time response capability, and optimization of cleaning task performance. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide an end-to-end trajectory planning algorithm for a magnetic adsorption cleaning robot. The algorithm uses artificial intelligence end-to-end model technology to directly output the planned trajectory and the control instructions of the magnetic adsorption electromagnet after the multimodal sensor input is processed by the model. In this process, the perception results are transmitted in the form of features, and the model memorizes, learns and predicts the future environment and behavior. The final planned trajectory and electromagnet control instructions come from the direct output after decoding the predicted information.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: An end-to-end trajectory planning algorithm for a magnetic adsorption cleaning robot, the steps are: Step 1: Train the multimodal perception backbone network, use the multimodal sensor data collected by the magnetic adsorption cleaning robot, and use the Transformer model to extract environmental features. , the Transformer model is a deep learning model; Step 2: Design a Transformer-based Decoder network to embed the robot control quantity into the existing encoding information and generate a decoder suitable for the current environment characteristics. The decoder generates a new output sequence based on the previously generated output sequence (i.e., the control instructions of the previous time step); Step 3: Design a Transformer-based regression output header network, which is responsible for mapping the last layer of features in the Decoder sequence output into the target trajectory of the agent and the target strength of the magnetic adsorption electromagnet; Step 4: Use a step-by-step training strategy to implement the training process and obtain the optimal end-to-end trajectory of the magnetic adsorption cleaning robot. Preferably, in step 1, the multimodal sensor data collected by the magnetic adsorption cleaning robot is used to extract environmental features using the Transformer model. ,The multimodal perception backbone network training process is as follows: The original input of the multimodal sensor is a collection of time series data, representing visual ,ultrasound and LiDAR data ; The data in each set are at different time steps The dimensions of each sensor data depend on the characteristics of its type. Depends on a Tensor, representing the height, width and number of channels of the image (RGB is 3); for ultrasonic data is a dimension of A vector representing the reading value of the ultrasonic sensor; for the lidar data is a The matrix of Represents the number of laser radar points, Represents the characteristic dimension of each point (spatial coordinates and reflection intensity, usually 4 dimensions) ; ; ; Different networks are used to extract the temporal features of the time series data sets of each modality mentioned above. Features below , each set of features has a different dimension For example, vision uses convolutional neural networks (CNNs) and uses As the parameter set of CNN, we finally get Similarly, for the ultrasonic distance field, a fully connected network FC is used and As a parameter set, the lidar point cloud uses the more mature pre-trained point cloud processing network PointNet. As parameters: ; ; ; The characteristics of each mode Mapped to a space of the same dimension, assuming the target dimension is , each modality has a feature map matrix and the bias vector , and can be combined to form a sequence : ; ; ; Constructing queries for multi-head attention fusion ,key ,value matrices, each of which consists of features Multiply by their respective weight matrices , and Get, respectively, through , and The projected query, key, and value matrices: ; ; ; Constructing multi-head attention calculation equation , which represents Attention Head Output: The similarity between the query and the key is reflected by calculating the dot product of the query and the key, divided by To avoid gradient instability, The function converts the similarity scores into probability distributions, which are then compared with the value matrix Multiply them together to get the weighted output. Represents the The query, key, and value matrix of each head, , is the number of attention heads.

[0005] ; Each attention head can be further Output connections And multiply by an output projection matrix , get the multi-head attention output , that is, the fused features ; ; ; Get fusion features After that, multiple residual connections are performed to convert the original features Add to fusion features On top, we further implement layer normalization To alleviate the vanishing gradient, a feedforward neural network is applied to the position of each modality in the process Further processing and fusion are performed to enhance nonlinear expression capabilities and average pooling is performed on the modal dimension To obtain a unified feature representation: ; ; ; ; Add position encoding to features To preserve the time series information, and further combine the embedding vectors of all time steps into a complete time series (input sequence) : ; ; Embed the sequence enter Get environmental characteristics , that is, the output feature of the Encoder, and for the supervisory signal (Expected label) Design loss function , This function is used to measure the degree of match between the model's output and the target, and supervise the training process through cross entropy loss. ; ; .

[0006] Preferably, in step 2, the obtained environmental characteristics are expressed as Used as the input sequence of the decoder part of , providing contextual information of the environment. The Decoder input sequence includes the output sequence of the previous time step , and pass it through embedding and position encoding (PositionEmbed) , mapped to dimensional vector space: ; is the position encoder, using the following design to convert the robot's control volume , which includes the linear velocity , facing angle and solenoid opening Trajectories are embedded in sequences to enhance the correlation between trajectory planning and solenoid control: ; ; ; ; Then the update of the position encoding and the final embedding equation are expressed as. represents the traditional sine-cosine position encoding; ; ; ; ; For each layer of the decoder based on the Transformer architecture, the following modules are included: Masked multi-head self-attention layer , encoder-decoder attention layer (This layer uses the ), feed-forward neural network layer , residual connections, and layer normalization (Here and are the standard deviation and variance, respectively. and are learnable parameters); by stacking multiple such layers, the decoder can effectively process the input sequence and generate output: ; ; ; ; ; ; ; .

[0007] Preferably, the sub-steps of step 3 are: First, define the trajectory information including the orientation angle , Linear speed , solenoid control signal , then obviously the dimension of the output layer is ; Use a linear layer to map the decoder output to the desired output dimension:

[0008]

[0009]

[0010]

[0011]

[0012] in refers to the output vector of the decoder at the tth time step; is the weight matrix of the output layer, is the bias vector, is the output without activation function; Since the value range and properties of each output are different, it is necessary to select a suitable activation function for each output. represent : ; ; ; Then the final control vector is: ; Calculate the loss function, respectively, angle loss , use the angle difference calculation to ensure the continuity of the angle error, and use the mean square error to calculate the angle loss , speed loss and electromagnet strength loss : ; ; ; ; The output loss function of trajectory and electromagnet control is obtained, and the optimal weight term is comprehensively measured by multiple experiments: .

[0013] Preferably, in step 4, the process of the step-by-step training strategy is as follows: (1) Based on freezing the trained Transformer perception network parameters, the decoder is trained using the extracted environmental features, decoder, and output layer in combination with the defined loss function; (2) At each time step t, the frozen perception network is used to extract environmental features as the input of the decoder. At the same time, the decoder receives the control information of the previous time step. The control information includes linear velocity, angular velocity and electromagnet opening, which are embedded and fused into the decoder input. The control information is mapped to the same dimension as the position encoding through linear transformation, and then weighted addition is performed with the position encoding to form a new position encoding. (3) During the training process, a comprehensive loss function is defined to measure the difference between the control instructions predicted by the model and the actual value, including the errors in linear velocity, angular velocity, and electromagnet opening. A weight coefficient is set for each loss item to balance the contribution of each part to the total loss. By minimizing the total loss function, the parameters of the decoder and output layer are optimized so that the model can accurately generate control instructions. (4) Unfreeze the perception network and perform end-to-end joint training with the decoder, while optimizing the parameters of the perception network and the decoder to minimize the total loss function.

[0014] Preferably, when the step-by-step training strategy is adopted, the teacher forcing strategy is adopted in the entire training process. When the training sequence generation model is used, the real target sequence is used as the input of the next time step to accelerate the convergence of the model and improve the training efficiency. At the same time, the gradient clipping method is used to ensure the stability of the model training.

[0015] The present invention can achieve the following beneficial effects: 1. The present invention adopts a multimodal perception backbone network, which can effectively integrate multiple sensor data such as vision, ultrasound and lidar, and improve the cleaning robot's ability to understand and adapt to the environment. By using different networks (such as convolutional neural networks, fully connected networks, point cloud processing networks) to extract features from data of different modalities, and deeply integrating them through a multi-head attention mechanism, it ensures that the model can fully understand the surrounding environment from multiple perspectives.

[0016] 2. Through the use of the Transformer model, the present invention can more efficiently capture long-term dependencies in time series data, which is crucial for path planning of cleaning robots. In particular, by adding position encoding, the time series information is maintained, which helps the robot better understand the changes in the dynamic environment.

[0017] 3. The Transformer-based decoder design allows the robot's control variables (such as linear velocity, angular velocity, and electromagnet opening) to be embedded in the decoding process, enhancing the correlation between trajectory planning and electromagnet control. This design not only improves the accuracy of trajectory planning, but also ensures that the robot can respond quickly to environmental changes, improving the safety and efficiency of cleaning operations.

[0018] 4. The step-by-step training method is adopted. The decoder is trained by fixing the perception network parameters first, and then the end-to-end joint training is performed, which effectively solves the instability problem that may be encountered in the early stage of model training. This method helps to gradually improve the overall performance of the model, and it is also easy to adjust and optimize each component, avoiding the occurrence of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a schematic diagram of a cleaning robot and a multimodal data source of the present invention; Figure 2 The end-to-end deep neural network model architecture of the present invention; Figure 3 A schematic diagram of the result (simulation environment) of the robot trajectory planning of the present invention. DETAILED DESCRIPTION

[0020] The preferred solution is Figures 1 to 3 As shown, an end-to-end trajectory planning algorithm for a magnetic adsorption cleaning robot is designed to improve the robot's autonomous navigation and cleaning efficiency in complex environments. The algorithm first uses a multimodal perception backbone network to fuse visual, ultrasonic and lidar data using a Transformer model to extract environmental features. Then, based on the Transformer-based Decoder design, the robot control quantities (linear velocity, angular velocity and electromagnet opening) are embedded in the decoding process to achieve accurate trajectory planning and electromagnet control. To ensure the stability and performance of the model, a step-by-step training strategy is adopted. The decoder is trained by fixing the perception network parameters first, and then end-to-end joint training is performed. By introducing position encoding, regularization methods, early stopping strategies and gradient clipping techniques, the generalization ability and training efficiency of the model are improved. The present invention shows significant advantages in multimodal data fusion, environmental feature representation, trajectory planning and control, and is suitable for cleaning operations in various complex environments. The specific method is as follows: Step 1: Train the multimodal perception backbone network. Use the multimodal sensor data collected by the magnetic adsorption cleaning robot and the Transformer model to extract environmental features. ,The mathematical description of the multimodal perception backbone network is as follows.

[0021] The original input of the multimodal sensor is a collection of time series data, representing visual, ultrasonic and lidar data. The data in each collection is at different time steps. The following records are recorded, and the dimensions of each sensor data depend on the characteristics of its type: ; ; ; Different networks are used to extract the temporal features of the time series data sets of each modality mentioned above. Features below , each set of features has a different dimension For example, vision uses convolutional neural networks (CNNs) and uses As the parameter set of CNN, we finally get Similarly, for the ultrasonic distance field, a fully connected network FC is used and As a parameter set, the lidar point cloud uses the more mature pre-trained point cloud processing network PointNet. As parameters: ; ; ; The characteristics of each mode Mapped to a space of the same dimension, assuming the target dimension is , each modality has a feature map matrix and the bias vector , and can be combined to form a sequence : ; ; ; Constructing queries for multi-head attention fusion ,key ,value matrices, each of which consists of features Multiply by their respective weight matrices , and Get, respectively, through , and The projected query, key, and value matrices: ; ; ; Constructing multi-head attention calculation equation , which represents Attention Head Output: The similarity between the query and the key is reflected by calculating the dot product of the query and the key, divided by To avoid gradient instability, The function converts the similarity scores into probability distributions, which are then compared with the value matrix Multiply them together to get the weighted output. Represents the The query, key, and value matrix of each head, , is the number of attention heads.

[0022] ; Each attention head can be further Output connections And multiply by an output projection matrix , get the multi-head attention output , that is, the fused features ; ; ; Get fusion features After that, multiple residual connections are performed to convert the original features Add to fusion features On top, we further implement layer normalization To alleviate the vanishing gradient, a feedforward neural network is applied to the position of each modality in the process Further processing and fusion are performed to enhance nonlinear expression capabilities and average pooling is performed on the modal dimension To obtain a unified feature representation: ; ; ; ; Add position encoding to features To preserve the time series information, and further combine the embedding vectors of all time steps into a complete time series (input sequence) : ; ; Embed the sequence enter Get environmental characteristics , that is, the output feature of the Encoder, and for the supervisory signal (Expected label) Design loss function , This function is used to measure the degree of match between the model's output and the target, and supervise the training process through cross entropy loss. ; ; .

[0023] Step 2: Design a Transformer-based Decoder to embed the robot control quantity and encode the environmental characteristics and the previous sequence output to generate a new output sequence.

[0024] The obtained environmental characteristics represent Used as the input sequence of the decoder part of , providing contextual information of the environment. The Decoder input sequence includes the output sequence of the previous time step , and pass it through embedding and position encoding (PositionEmbed) , mapped to dimensional vector space: ; is the position encoder, using the following design to convert the robot's control volume , which includes the linear velocity , facing angle and solenoid opening Trajectories are embedded in sequences to enhance the correlation between trajectory planning and solenoid control: ; ; ; ; Then the update of the position encoding and the final embedding equation are expressed as. represents the traditional sine-cosine position encoding; ; ; ; ; For each layer of the decoder based on the Transformer architecture, the following modules are included: Masked multi-head self-attention layer , encoder-decoder attention layer (This layer uses the ), feed-forward neural network layer , residual connections, and layer normalization (Here and are the standard deviation and variance, respectively. and are learnable parameters); by stacking multiple such layers, the decoder can effectively process the input sequence and generate output: ; ; ; ; ; ; ; .

[0025] Step 3: Design a Transformer-based Header to convert the sequence output of the Decoder into the target trajectory of the agent and the target strength of the magnetic adsorption electromagnet.

[0026] First, define the trajectory information including the orientation angle , Linear speed , solenoid control signal , then obviously the dimension of the output layer is ; Use a linear layer to map the decoder output to the desired output dimension:

[0027]

[0028]

[0029]

[0030]

[0031] in refers to the output vector of the decoder at the tth time step; is the weight matrix of the output layer, is the bias vector, is the output without activation function; Since the value range and properties of each output are different, it is necessary to select a suitable activation function for each output. represent : ; ; ; Then the final control vector is: ; Calculate the loss function, respectively is the angle loss , use the angle difference calculation to ensure the continuity of the angle error, and use the mean square error to calculate the angle loss , speed loss and electromagnet strength loss : ; ; ; ; The output loss function of trajectory and electromagnet control is obtained, and the optimal weight term is measured comprehensively through multiple experiments: .

[0032] Step 4: Implement the training process using a step-by-step training strategy.

[0033] In order to ensure the stability of the model and improve performance, we adopted a step-by-step training strategy. First, on the basis of freezing the trained Transformer perception network parameters, we used the environmental features extracted in the previous steps, the decoder and the output layer, combined with the defined loss function, to train the decoder.

[0034] At each time step t, a frozen perception network is used to extract environmental features as the input of the decoder. At the same time, the decoder receives the control quantity information (linear velocity, angular velocity, and electromagnet opening) of the previous time step, embeds it and fuses it into the decoder input. Specifically, the control quantity is mapped to the same dimension as the position encoding through linear transformation, and then weightedly added to the position encoding to form a new position encoding. This allows the model to not only use time position information during the decoding process, but also combine the previous control quantity information to improve the modeling ability of the control instruction sequence.

[0035] During the training process, a comprehensive loss function is defined to measure the difference between the control instructions predicted by the model and the actual value, including the errors of linear velocity, angular velocity and electromagnet opening, and a weight coefficient is set for each loss item to balance the contribution of each part to the total loss. By minimizing the total loss function and optimizing the parameters of the decoder and output layer, the model can accurately generate control instructions.

[0036] Next, we unfreeze the perception network and perform end-to-end joint training with the decoder, while optimizing the parameters of the perception network and the decoder to minimize the total loss function. In this way, the various parts of the model can work together to further improve the overall performance. To prevent overfitting, we use an early stopping strategy during training. When the performance of the validation set no longer improves, we stop training early, and use regularization methods such as weight decay and dropout to enhance the generalization ability of the model.

[0037] The entire training process adopts the teacher forcing strategy. When training the sequence generation model, the real target sequence is used as the input of the next time step to accelerate the convergence of the model and improve the training efficiency. At the same time, attention is paid to the problem of gradient disappearance or explosion, and gradient clipping and other technologies are used to ensure the stability of model training.

[0038] Receive multimodal sensor input: The algorithm receives multimodal input data from different sensors, including camera image input, inertial navigation data, near-field ultrasonic data, lidar data, etc. These inputs implicitly provide contextual information about the robot and its environment; Feature extraction network: Input the data of each modality into the corresponding feature extraction network to extract their respective high-dimensional features; Environmental feature BEV representation: The feature output from step 2 is reorganized into a BEV-based environmental representation Z through a Transformer Encoder module using a multi-head attention mechanism. This representation provides the context of an environment; Trajectory and Environment Transformer Modules: At this point, the Trajectory Transformer and Environment Transformer modules receive the encoded environment representation Z. The Environment Transformer module extracts key spatial features from Z, while the Trajectory Transformer focuses on path planning information through the Transformer’s self-attention mechanism. These modules are able to capture the complex relationship between the environment and the path.

[0039] Control Transformer Module: The control transformer receives the output features from the trajectory transformer and the environment transformer, and combines the robot's current action query to generate control predictions containing timing information.

[0040] Control sequence decoder: The output of the control transformer is passed to the control sequence decoder head to generate a time-step control sequence. These sequences represent the specific action instructions that the robot should take at the next moment, including the movement direction, speed, and adjustment of the magnetic adsorption strength.

[0041] Robot control output: Ultimately, the output of the control sequence decoder is converted into actual robot control signals that drive the robot to perform cleaning tasks in the environment. Through end-to-end planning and control, the robot is able to achieve autonomous path planning based on sensor data, avoid obstacles and clean specific areas in magnetic adsorption mode.

[0042] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. An end-to-end trajectory planning algorithm for a magnetic adsorption cleaning robot, characterized in that The following steps are involved: Step 1: Train the multimodal perception backbone network, use the multimodal sensor data collected by the magnetic adsorption cleaning robot, and use the Transformer model to extract environmental features. , the Transformer model is a deep learning model; Step 2: Design a Transformer-based Decoder network to embed the robot control quantity into the existing encoding information and generate a decoder suitable for the current environment characteristics. The decoder will generate a new output sequence based on the previously generated output sequence. Step 3: Design a Transformer-based regression output header network, which is responsible for mapping the last layer of features in the Decoder sequence output into the target trajectory of the agent and the target strength of the magnetic adsorption electromagnet; Step 4: Use a step-by-step training strategy to implement the training process and obtain the optimal end-to-end trajectory of the magnetic adsorption cleaning robot.

2. The end-to-end trajectory planning algorithm of a magnetic adsorption cleaning robot according to claim 1, characterized in that: In step 1, the multimodal sensor data collected by the magnetic adsorption cleaning robot is used to extract environmental features using the Transformer model. ,The multimodal perception backbone network training process is as follows: The original input of the multimodal sensor is a collection of time series data, representing visual ,ultrasound and LiDAR data ; The data in each set are at different time steps The dimensions of each sensor data depend on the characteristics of its type; for image data Depends on a The tensor represents the height, width and number of channels of the image, that is, RGB is 3; for ultrasonic data is a dimension of A vector representing the reading value of the ultrasonic sensor; for the lidar data is a The matrix of Represents the number of laser radar points, Represents the feature dimension of each point: ； ； ； Different networks are used to extract the temporal features of the time series data sets of each modality mentioned above. Features below , each set of features has a different dimension , visual using convolutional neural network CNN and using As the parameter set of CNN, we finally get feature set; for the ultrasonic distance field, a fully connected network FC is used and As a parameter set, the lidar point cloud uses the pre-trained point cloud processing network PointNet. As parameters: ； ； ； The characteristics of each mode Mapped to a space of the same dimension, assuming the target dimension is , each modality has a feature map matrix and the bias vector , and can be combined to form a sequence : ； ； ； Constructing queries for multi-head attention fusion ,key ,value matrices, each of which consists of features Multiply by their respective weight matrices , and Get, respectively, through , and The projected query, key, and value matrices: ； ； ； Constructing multi-head attention calculation equation , which represents Attention Head Output: The similarity between the query and the key is reflected by calculating the dot product of the query and the key, divided by To avoid gradient instability, The function converts the similarity scores into probability distributions, which are then compared with the value matrix Multiply them together to get the weighted output; Represents the The query, key, and value matrix of each head, , is the number of attention heads: ； Each attention head Output connections And multiply by an output projection matrix , get the multi-head attention output , that is, the fused features ; ； ； Get fusion features After that, multiple residual connections are performed to convert the original features Add to fusion features On top, implement layer normalization To alleviate the vanishing gradient, a feedforward neural network is applied to the position of each modality in the process Further processing and fusion are performed to enhance nonlinear expression capabilities and average pooling is performed on the modal dimension To obtain a unified feature representation: ； ； ； ； Add position encoding to features To preserve the time series information, and further combine the embedding vectors of all time steps into a complete time series : ； ； Will enter Get environmental characteristics , that is, the output feature of the Encoder, and for the supervisory signal Design loss function , This function is used to measure the degree of match between the model output and the target, and supervise the training process through cross entropy loss; ； 。 3. The end-to-end trajectory planning algorithm of a magnetic adsorption cleaning robot according to claim 1 is characterized in that: In step 2, the obtained environmental features are expressed as Used as the input sequence of the decoder part, providing contextual information of the environment; the Decoder input sequence includes the output sequence of the previous time step , and encode it through embedding and position encoding , mapped to dimensional vector space: ； is the position encoder, using the following design to convert the robot's control volume , which includes the linear velocity , facing angle and solenoid opening Trajectories are embedded in sequences to enhance the correlation between trajectory planning and solenoid control: ； ； ； ； Then the update of the position encoding and the final embedding equation are expressed as; represents the traditional sine-cosine position encoding; ； ； ； ； For each layer of the decoder based on the Transformer architecture, the following modules are included: Masked multi-head self-attention layer , encoder-decoder attention layer , feed-forward neural network layer , residual connections, and layer normalization ,in, and are the standard deviation and variance, respectively. and are learnable parameters; by stacking multiple such layers, the decoder can effectively process the input sequence and generate output: ； ； ； ； ； ； ； 。 4. The end-to-end trajectory planning algorithm of a magnetic adsorption cleaning robot according to claim 1, characterized in that: The sub-steps of step 3 are: First, define the trajectory information including the orientation angle , Linear speed , solenoid control signal , then obviously the dimension of the output layer is ; Use a linear layer to map the decoder output to the desired output dimension: ； ； ； ； ； in refers to the output vector of the decoder at the tth time step; is the weight matrix of the output layer, is the bias vector, is the output without activation function; Since the value range and properties of each output are different, it is necessary to select a suitable activation function for each output. represent : ； ； ； Then the final control vector is: ； Calculate the loss function, respectively is the angle loss , use the angle difference calculation to ensure the continuity of the angle error, and use the mean square error to calculate the angle loss , speed loss and electromagnet strength loss : ； ； ； ； The output loss function of trajectory and electromagnet control is obtained, and the optimal weight term is measured comprehensively through multiple experiments: 。 5. The end-to-end trajectory planning algorithm of a magnetic adsorption cleaning robot according to claim 1, characterized in that: In step 4, the process of step-by-step training strategy is as follows: (1) Based on freezing the trained Transformer perception network parameters, the decoder is trained using the extracted environmental features, decoder, and output layer in combination with the defined loss function; (2) At each time step t, the frozen perception network is used to extract environmental features as the input of the decoder. At the same time, the decoder receives the control information of the previous time step; the control information includes linear velocity, angular velocity and electromagnet opening, which is embedded and fused into the decoder input. The control amount is mapped to the same dimension as the position code through linear transformation, and then weightedly added to the position code to form a new position code; (3) During the training process, a comprehensive loss function is defined to measure the difference between the control instructions predicted by the model and the actual value, including the errors in linear velocity, angular velocity, and electromagnet opening. A weight coefficient is set for each loss item to balance the contribution of each part to the total loss. By minimizing the total loss function, the parameters of the decoder and output layer are optimized so that the model can accurately generate control instructions. (4) Unfreeze the perception network and perform end-to-end joint training with the decoder, while optimizing the parameters of the perception network and the decoder to minimize the total loss function.

6. The end-to-end trajectory planning algorithm of a magnetic adsorption cleaning robot according to claim 5, characterized in that: When adopting the step-by-step training strategy, the teacher forcing strategy is adopted in the whole training process. When training the sequence generation model, the real target sequence is used as the input of the next time step to accelerate the convergence of the model and improve the training efficiency. At the same time, the gradient clipping method is used to ensure the stability of the model training.

Images