A privacy-preserving computation-based indoor robot collision prevention method

Abstract

This invention discloses an indoor robot collision prevention method based on privacy computing, comprising: S1, real-time acquisition of environmental visual data to generate semantic obstacle features; S2, construction of a privacy-aware field vector; S3, prediction of dynamic obstacle intention trajectories and generation of a repulsive potential energy field based on an improved ConvLSTM model using an interactive potential energy factor; S4, fusion of the potential energy field to generate and execute obstacle avoidance control commands; S5, differential privacy processing of the execution field and results to generate decision gene fragments and upload them to the cloud; S6, introduction of a scene novelty factor through an improved MAP-Elites algorithm to aggregate fragments and generate a new generation of global obstacle avoidance strategy model; S7, distribution of the strategy model to the robot. This invention achieves end-to-end privacy-protected swarm intelligence collaborative evolution, solving the problems of high privacy leakage risk, weak swarm collaboration ability, and poor strategy generalization of traditional methods.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security and privacy computing, and in particular to an indoor robot collision prevention method based on privacy computing. Background Technology

[0002] Indoor robot collision avoidance methods based on privacy computing have been widely applied in recent years in fields such as smart healthcare, smart warehousing, and commercial services because they can achieve collaborative evolution of collective intelligence while protecting user privacy. This has become an important development direction for realizing safe collaboration among large-scale indoor robots. However, in practical applications, indoor robot obstacle avoidance scenarios face many challenges, including dynamically changing environments, strict privacy regulations, and homogeneous group strategies. The deployment effectiveness of existing technologies remains constrained by various factors.

[0003] Currently, most indoor robot avoidance and control systems rely on the local perception and decision-making of a single indoor robot, making it difficult to utilize the valuable experience accumulated by other indoor robots in different scenarios. This results in poor model generalization ability and insufficient adaptability when facing unknown environments. Although some systems attempt to share data in the cloud, they often directly upload raw sensor data or decision context, which seriously exposes users' personal privacy and environmental layout, posing a huge risk of data leakage and failing to meet increasingly stringent privacy protection requirements.

[0004] Furthermore, most existing cloud-based strategy optimization methods only pursue single performance indicators such as obstacle avoidance success rate, neglecting the diversity of decision-making behaviors and scenario coverage. This can easily lead to all indoor robots learning similar strategies, resulting in collective failures in specific complex scenarios. At the same time, the strategy optimization process lacks effective incentives for novel behaviors, making it difficult to explore and discover optimal solutions for rare but high-risk scenarios. This limits the improvement of the overall robustness of the system and seriously affects the practical value and stability of the model in real-world complex environments.

[0005] Therefore, how to provide a privacy-preserving computation-based indoor robot collision prevention method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a collision avoidance method for indoor robots based on privacy computing. This invention fully integrates key steps such as homomorphic encrypted feature extraction, privacy-aware field construction, improved ConvLSTM model trajectory prediction, differential privacy experience uploading, and improved MAP-Elites algorithm cloud evolution. It constructs a closed-loop optimization process for obstacle avoidance strategies that is edge-cloud collaborative, data-driven, and maintains full privacy, achieving continuous evolution of the swarm intelligence of indoor robots and dynamic improvement of collision avoidance capabilities while protecting user privacy. This invention introduces an interaction potential energy factor into the improved ConvLSTM model, achieving a fusion prediction of physical priors and data-driven approaches, improving the accuracy and physical rationality of trajectory prediction. By introducing a scene novelty factor into the improved MAP-Elites algorithm, it incentivizes exploration of diverse scenarios, avoiding strategy homogenization. This invention possesses advantages such as high privacy protection strength, accurate prediction decisions, and strong policy evolution adaptability, significantly improving the obstacle avoidance success rate and overall robustness of indoor robots in complex dynamic environments, thereby effectively solving problems such as high privacy leakage risk, weak swarm collaboration ability, and poor policy generalization in existing technologies.

[0007] An indoor robot collision prevention method based on privacy computing according to an embodiment of the present invention includes the following steps: S1. Real-time acquisition of environmental visual data through homomorphic encryption preprocessing to generate semantic obstacle features; S2. Construct a privacy-aware field using semantic obstacle features and generate a privacy-aware field vector; S3. Based on the privacy-aware field, an interactive potential energy factor is introduced. Through an improved ConvLSTM model, the field vector labeled as dynamic obstacles is continuously tracked and its changes are analyzed to predict the future intention trajectory of the dynamic obstacles and generate a repulsive potential energy field. S4. The privacy-aware field and the repulsive potential energy field are integrated to form a comprehensive potential energy field. Based on the comprehensive potential energy field, obstacle avoidance control commands are generated and executed by finding the path with the fastest potential energy gradient descent. S5. Record the execution start and end times and execution results of obstacle avoidance control commands, perform differential privacy processing, generate decision gene fragments, and upload them to the cloud server; S6. In the cloud server, decision gene fragments from multiple indoor robots are collected. A behavior-performance map is constructed using the improved MAP-Elites algorithm. A scene novelty factor is introduced to quantify the difference between the decision gene fragments and the behavior-performance map. High-performance decision gene combinations covering multiple scenes are selected to generate a new generation of global obstacle avoidance strategy models. S7. Distribute the new generation of global obstacle avoidance strategy model to the indoor robot to continuously generate and execute obstacle avoidance control commands.

[0008] Optionally, S1 specifically includes: S11. On the indoor robot end, each pixel in the environmental visual data collected by the homomorphic encryption preprocessing integrated into the image sensor is calculated in the encrypted domain with the Hamming distance of the preset obstacle semantic label template. Pixel areas with Hamming distance less than the set distance threshold are marked as obstacles and assigned corresponding semantic labels to generate semantic labels for obstacles. S12. Within the encrypted domain, a stereo vision matching algorithm within the encrypted domain is used to perform pixel-level matching of the left and right eye images in the acquired environmental visual data, calculate the disparity value of each obstacle pixel, convert the disparity value into a depth value based on the camera intrinsic parameters, and calculate the relative spatial position of the obstacle by combining the pixel coordinates. S13. Within the encrypted domain, for the same obstacle marked in two consecutive frames of environmental visual data, the pixel displacement of the obstacle on the image plane is calculated using the optical flow method. Combined with the relative spatial position of the obstacle, the pixel displacement is converted into the actual spatial displacement and divided by the time interval between the two frames to obtain the velocity vector of the obstacle. S14. Combine the semantic labels, relative spatial positions, and velocity vectors of obstacles in sequence to generate semantic obstacle features.

[0009] Optionally, S2 specifically includes: S21. Map the relative spatial position of semantic obstacle features to a two-dimensional grid map of a preset size. Each grid cell stores the semantic labels of all obstacles within the coverage area, generating a static privacy-aware field containing location and semantic information. S22. Fill the velocity vector in the semantic obstacle features into the corresponding grid cell in the static privacy-aware field. Each grid cell contains position, semantic and velocity information, generating a dynamic privacy-aware field. S23. Arrange the semantic tags, relative spatial positions, and velocity vectors stored in each grid cell of the privacy-aware field into a one-dimensional vector in a preset order to generate a privacy-aware field vector.

[0010] Optionally, S3 specifically includes: S31. Based on the privacy-aware field, filter out grid cells with semantic tags as dynamic obstacles, read the relative spatial position and velocity vector of the dynamic obstacle in each grid cell, calculate the Euclidean distance between each dynamic obstacle and the current position of the indoor robot, and divide the preset attenuation constant by the Euclidean distance to obtain a distance quotient. S32. Using the natural constant e as the base and the opposite of the distance quotient as the exponent, perform an exponential operation to obtain an intermediate value that decreases as the distance increases. Multiply the intermediate value by a preset intensity constant to obtain a final scalar repulsive force value, which is the interaction potential energy factor. S33. Using the improved ConvLSTM model, the privacy-aware field vector of the current time step is reshaped into two-dimensional grid data. The data is input into a convolutional encoder with a pre-set kernel size. The feature map is generated by the sliding window convolution in the first layer. The ReLU activation function is applied to each feature map. Then, the high-dimensional spatial feature map is output by convolution with the same kernel size and the ReLU activation function. S34. The spatial feature map and the interaction potential factor map at time step t are concatenated along the channel dimension to form a concatenated input tensor, which is then input into the input gate, forget gate and output gate respectively. Each gate independently performs a convolution operation with a preset kernel size, and a weight matrix between 0 and 1 is generated by the Sigmoid activation function. S35. Input the concatenated input tensor into the candidate memory layer of the improved ConvLSTM model, perform a convolution operation of a preset size, and generate a candidate value through the Tanh activation function. Multiply the unit state output at time step t-1 element-wise with the weight matrix of the forget gate, and add the result to the element-wise product of the weight matrix of the input gate and the candidate value to update the unit state at time step t. S36. Apply the Tanh activation function to the cell state at time step t, multiply the result element-wise with the weight matrix of the output gate, output the hidden state at time step t, input the hidden state output at the last time step into the convolutional decoder of the improved ConvLSTM model, perform upsampling convolution through the transposed convolutional kernel of the first layer with the same kernel size as the convolutional encoder, and generate a feature map. S37. Apply the ReLU activation function to each feature map, and then perform upsampling convolution through the transposed convolution kernel of the same size to generate a feature map, which is an initial potential energy field containing future trajectory prediction with the same size as the original privacy-aware field. The value of each grid cell in the initial potential energy field is weighted and summed with the interaction potential energy factor at the corresponding position to generate a repulsive potential energy field that evolves over time and contains predicted trajectory information.

[0011] Optionally, S4 specifically includes: S41. The privacy-aware field and the repulsive potential energy field at the current time step are mapped one-to-one according to the grid position, and the values ​​in the same grid cell are added element by element to generate a comprehensive potential energy field containing static environmental information and dynamic hazard prediction. S42. In the comprehensive potential energy field, taking the grid cell where the indoor robot is currently located as the starting point, calculate the potential energy difference between it and the eight adjacent grid cells around it, and determine the direction of the fastest potential energy decrease and the corresponding adjacent grid cells. S43. Move the indoor robot to the adjacent grid cell with the fastest potential energy decrease, and use the grid cell as the new starting point. Repeat the process of calculating and judging the potential energy difference to gradually build a path from the starting point to the target point, which is composed of multiple grid cells. S44. Convert the constructed path into obstacle avoidance control commands for the linear velocity and angular velocity of the indoor robot, and send the obstacle avoidance control commands to the motion controller of the indoor robot to perform predictive obstacle avoidance operations.

[0012] Optionally, S5 specifically includes: S51. Record the comprehensive potential energy field at the moment of executing the obstacle avoidance control command as the execution field, and record the actual motion trajectory of the indoor robot after executing the obstacle avoidance control command as the execution result. S52. Concatenate the execution start and end of the execution process and the execution result in chronological order to form a long data vector. Perform differential privacy processing, initialize a pseudo-random number generator, and use the current time step as the seed to generate a uniformly distributed random number between 0 and 1. Substitute the random number generator into the inverse cumulative distribution function of the Laplace distribution for calculation to obtain a Laplace distribution noise value. S53. Add the Laplace distribution noise value to each value in the long data vector to generate an anonymous data vector after differential privacy processing. Hash the anonymous data vector after differential privacy processing to generate an anonymous decision gene fragment of a set length, and upload it to the cloud server through an encrypted communication channel.

[0013] Optionally, the inverse cumulative distribution function of the Laplace distribution specifically includes: With pre-set scale parameters, at the start of the calculation, it is determined whether the uniformly distributed random number between 0 and 1 is less than 0.5 or greater than or equal to 0.5. If it is greater than or equal to 0.5, the second calculation path is entered, the number 1 is subtracted from the current uniformly distributed random number, the result is multiplied by 2, and a natural logarithm operation is performed to obtain a negative number. Multiplying this negative number by -1 gives a positive number. Multiplying this positive number by a pre-set scaling parameter gives a Laplace distribution noise value. If the initial uniformly distributed random number is less than 0.5, multiply the uniformly distributed random number by 2 and directly perform a natural logarithm operation on the result. Multiplying the logarithm result by a pre-set scaling parameter also gives a Laplace distribution noise value.

[0014] Optionally, S6 specifically includes: S61. In the cloud server, all received anonymous decision gene fragments are collected. Each decision gene fragment is input into a preset three-layer fully connected neural network and calculated with the corresponding preset weight matrix and bias vector. The ReLU activation function is applied element by element to generate a final behavior descriptor that reduces the dimensionality of the intermediate feature vector output by the first layer to one-quarter. A unique behavior descriptor is generated for each decision gene fragment. S62. Initialize a two-dimensional grid structure in memory as a behavior-performance map. Set the horizontal axis of the behavior-performance map to represent the first dimension of the behavior descriptor and the vertical axis to represent the second dimension of the behavior descriptor. Divide the horizontal axis and the vertical axis into a preset number of equally spaced intervals to form a grid. Each grid is a file unit. S63. For each decision gene segment, extract the values ​​of the first and second dimensions from the corresponding behavioral descriptor, and normalize them to an integer range from 0 to a preset number minus one, to obtain a horizontal axis index and a vertical axis index. S64. Mark the file unit corresponding to the coordinate index as the target unit, calculate the obstacle avoidance success rate of the current decision gene segment when it is executed locally as the performance index, store the current decision gene segment and the corresponding performance index together into the target unit. If there are other segments in the target unit, compare the performance index and only keep the one with the higher performance index. S65. For each newly arrived decision gene fragment, obtain the behavior descriptor, traverse all non-empty file units in the behavior-performance map, read the behavior descriptor of the existing fragments in each file unit, calculate the Euclidean distance between the behavior descriptor of the newly arrived decision gene fragment and the behavior descriptor of each existing fragment, and sum all the Euclidean distance values ​​to obtain a total value, which is the scene novelty factor. S66. Set a novelty weight coefficient, add the performance index of each decision gene segment to the scenario novelty factor multiplied by the corresponding novelty weight coefficient, generate a comprehensive fitness score, establish an empty parent candidate pool, sort all decision gene segments according to their corresponding comprehensive fitness scores from high to low, select the top preset number of decision gene segments with the highest comprehensive fitness scores, and put them into the parent candidate pool. S67. From the parent candidate pool, randomly select two parent decision gene segments each time and perform a crossover operation with a preset probability. That is, randomly generate a crossover position and swap the parts of the two segments after the crossover position to generate two new offspring segments. Perform a Gaussian mutation operation with a probability of 1 minus the preset probability. For each value in the decision gene segment, add a noise value randomly drawn from a Gaussian distribution with a mean of 0 and a variance of 1 minus the preset probability to generate a new offspring decision gene segment. S68. For each newly generated offspring decision gene fragment, repeat the steps to calculate the new behavioral descriptor, performance index, and overall fitness score. Based on the new behavioral descriptor, determine the target unit in the behavior-performance map. If the overall fitness score of the offspring decision gene fragment is higher than the overall fitness score of the existing decision gene fragment in the target unit, replace the existing decision gene fragment with the offspring decision gene fragment; otherwise, discard the offspring decision gene fragment. S69. When the average comprehensive fitness score of all file units in the behavior-performance map does not increase for a preset number of consecutive iterations, the evolution process is terminated. Each file unit is traversed, the decision gene fragment with the highest comprehensive fitness score is read, and all the read decision gene fragments are spliced ​​together according to the coordinate order of the file unit to form a complete set of strategies, which is the new generation of global obstacle avoidance strategy model.

[0015] Optionally, S7 specifically includes: S71. On the indoor robot end, a new generation of global obstacle avoidance strategy model is received from the cloud server through an encrypted communication channel and written into the local memory to complete the update of the global obstacle avoidance strategy model. S72. During the operation of the indoor robot, predictive obstacle avoidance control commands are continuously generated and output. The obstacle avoidance control commands are sent to the motion controller of the indoor robot, and the motion controller drives the indoor robot to execute the obstacle avoidance control commands.

[0016] The beneficial effects of this invention are: First, by integrating homomorphic encryption preprocessing into the indoor robot end, semantic obstacle features are extracted directly within the encrypted domain, constructing a dynamic privacy-aware field. This fundamentally avoids the risk of leakage of the original environmental images and provides a highly privacy-secure data foundation for subsequent intelligent decision-making.

[0017] Secondly, the intention trajectory prediction of dynamic obstacles is realized based on the improved ConvLSTM model. This model strengthens the judgment of the physical rationality of the obstacle movement trend by introducing the interactive potential energy factor, which significantly improves the accuracy and robustness of trajectory prediction in complex dynamic environments and provides a reliable basis for generating predictive obstacle avoidance commands.

[0018] Furthermore, in the cloud server, a behavior-performance map is constructed using the improved MAP-Elites algorithm. This algorithm introduces a scenario novelty factor, which not only considers the performance indicators of decision gene fragments when evaluating them, but also quantifies the difference between their behavior and existing strategies. This incentivizes the exploration and coverage of diverse scenarios, effectively avoids the homogenization of group strategies, and achieves the diversity and high-performance evolution of global obstacle avoidance strategies.

[0019] In summary, this invention constructs a closed-loop optimization system for edge-cloud collaboration by integrating edge-side privacy computing, improved ConvLSTM model prediction, and improved MAP-Elites algorithm evolution. This significantly improves the privacy security, prediction accuracy, and swarm intelligence level of indoor robot collision prevention and control, and enhances the practical value and stability of the system in real and complex scenarios. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an indoor robot collision prevention method based on privacy computing proposed in this invention; Figure 2 This is a flowchart of the dynamic obstacle trajectory prediction process using the improved ConvLSTM model based on interactive potential energy factors proposed in this invention. Figure 3 This is a flowchart of the global policy evolution of the improved MAP-Elites algorithm based on scene novelty factor proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A collision prevention method for indoor robots based on privacy computing includes the following steps: S1. On the indoor robot end, environmental visual data is collected in real time through homomorphic encryption preprocessing integrated into the image sensor. Without generating a recognizable original image, the semantic labels, relative spatial positions and velocity vectors of obstacles are directly extracted in the encrypted domain to generate semantic obstacle features. S2. Construct a dynamic, non-visual privacy-aware field using semantic obstacle features and generate a privacy-aware field vector. S3. Based on the privacy-aware field, an interactive potential energy factor is introduced. Through the improved ConvLSTM model, the field vector labeled as dynamic obstacles is continuously tracked and its changes are analyzed to predict the future intention trajectory of the dynamic obstacles and generate a repulsive potential energy field that evolves over time and contains the predicted trajectory information. S4. The privacy-aware field and the repulsive potential energy field are integrated to form a comprehensive potential energy field for path planning. Based on the comprehensive potential energy field, the indoor robot generates and executes predictive obstacle avoidance control commands by finding the path with the fastest potential energy gradient descent. S5. Record the execution start and end times and execution results of obstacle avoidance control commands. After differential privacy processing of the execution start and end times and execution results, generate anonymous decision gene fragments and upload the decision gene fragments to the cloud server. S6. In the cloud server, decision gene fragments from multiple indoor robots are collected. A behavior-performance map is constructed using the improved MAP-Elites algorithm. A scene novelty factor is introduced to quantify the difference between the behavioral characteristics of the decision gene fragments and the existing features in the behavior-performance map. High-performance decision gene combinations covering multiple scenarios are selected from these to generate a new generation of global obstacle avoidance strategy models. S7. Distribute the new generation of global obstacle avoidance strategy model to the indoor robot to continuously generate and execute obstacle avoidance control commands.

[0023] This implementation significantly improves the safety and swarm intelligence of indoor robot collision avoidance. By employing edge-side homomorphic encryption and privacy-aware field construction, feature extraction is achieved while ensuring the original visual data remains undisclosed, completely eliminating privacy risks. An improved ConvLSTM model incorporating interactive potential factors, combined with physical prior prediction of dynamic obstacle intention trajectories, achieves high-precision predictive obstacle avoidance. In the cloud, an improved MAP-Elites algorithm and scene novelty factor are used to aggregate anonymized multi-robot experiences for collaborative evolution, avoiding strategy homogenization. This edge-cloud collaborative mechanism effectively solves the challenge of swarm collaboration in complex dynamic environments, significantly improving the robustness, adaptability, and overall operational efficiency of obstacle avoidance strategies.

[0024] In this embodiment, S1 specifically includes: S11. On the indoor robot end, each pixel in the environmental visual data collected by the homomorphic encryption preprocessing integrated into the image sensor is calculated in the encrypted domain with the Hamming distance of the preset obstacle semantic label template. Pixel areas with Hamming distance less than the set distance threshold are marked as obstacles and assigned corresponding semantic labels to generate semantic labels for obstacles. S12. Within the encrypted domain, a stereo vision matching algorithm within the encrypted domain is used to perform pixel-level matching of the left and right eye images in the acquired environmental visual data, calculate the disparity value of each obstacle pixel, convert the disparity value into a depth value based on the camera intrinsic parameters, and calculate the relative spatial position of the obstacle by combining the pixel coordinates. S13. Within the encrypted domain, for the same obstacle marked in two consecutive frames of environmental visual data, the pixel displacement of the obstacle on the image plane is calculated using the optical flow method. Combined with the relative spatial position of the obstacle, the pixel displacement is converted into the actual spatial displacement and divided by the time interval between the two frames to obtain the velocity vector of the obstacle. S14. Combine the semantic labels, relative spatial positions, and velocity vectors of obstacles in sequence to generate semantic obstacle features.

[0025] In this embodiment, S2 specifically includes: S21. Map the relative spatial position of semantic obstacle features to a two-dimensional grid map of a preset size. Each grid cell stores the semantic labels of all obstacles within the coverage area, generating a static privacy-aware field containing location and semantic information. S22. Fill the velocity vector in the semantic obstacle features into the corresponding grid cell in the static privacy-aware field. Each grid cell contains position, semantic and velocity information, generating a dynamic privacy-aware field. S23. Arrange the semantic tags, relative spatial positions, and velocity vectors stored in each grid cell of the privacy-aware field into a one-dimensional vector in a preset order to generate a privacy-aware field vector.

[0026] In this embodiment, S3 specifically includes: S31. Based on the privacy-aware field, filter out grid cells with semantic tags as dynamic obstacles, read the relative spatial position and velocity vector of the dynamic obstacle in each grid cell, calculate the Euclidean distance between each dynamic obstacle and the current position of the indoor robot, and divide the preset attenuation constant by the Euclidean distance to obtain a distance quotient. S32. Using the natural constant e as the base and the opposite of the distance quotient as the exponent, perform an exponential operation to obtain an intermediate value that decreases as the distance increases. Multiply the intermediate value by a preset intensity constant to obtain a final scalar repulsive force value, which is the interaction potential energy factor. S33. Using the improved ConvLSTM model, the privacy-aware field vector of the current time step is reshaped into two-dimensional grid data. The data is input into a convolutional encoder with a pre-set kernel size. The feature map is generated by the sliding window convolution in the first layer. The ReLU activation function is applied to each feature map. Then, the high-dimensional spatial feature map is output by convolution with the same kernel size and the ReLU activation function. S34. The spatial feature map and the interaction potential factor map at time step t are concatenated along the channel dimension to form a concatenated input tensor, which is then input into the input gate, forget gate and output gate respectively. Each gate independently performs a convolution operation with a preset kernel size, and a weight matrix between 0 and 1 is generated by the Sigmoid activation function. S35. Input the concatenated input tensor into the candidate memory layer of the improved ConvLSTM model, perform a convolution operation of a preset size, and generate a candidate value through the Tanh activation function. Multiply the unit state output at time step t-1 element-wise with the weight matrix of the forget gate, and add the result to the element-wise product of the weight matrix of the input gate and the candidate value to update the unit state at time step t. S36. Apply the Tanh activation function to the cell state at time step t, multiply the result element-wise with the weight matrix of the output gate, output the hidden state at time step t, input the hidden state output at the last time step into the convolutional decoder of the improved ConvLSTM model, perform upsampling convolution through the transposed convolutional kernel of the first layer with the same kernel size as the convolutional encoder, and generate a feature map. S37. Apply the ReLU activation function to each feature map, and then perform upsampling convolution through the transposed convolution kernel of the same size to generate a feature map, which is an initial potential energy field containing future trajectory prediction with the same size as the original privacy-aware field. The value of each grid cell in the initial potential energy field is weighted and summed with the interaction potential energy factor at the corresponding position to generate a repulsive potential energy field that evolves over time and contains predicted trajectory information.

[0027] This implementation achieves physically interpretable intention trajectory prediction for dynamic obstacles by introducing an improved ConvLSTM model and an interactive potential factor. First, the interactive potential factor, which decays spatially based on Euclidean distance, is calculated to inject prior physical knowledge into the model, strengthening the constraint on the obstacle's motion trend. A convolutional encoder extracts high-dimensional spatial features of the privacy-aware field, and these features are fused with the potential factor and input into the ConvLSTM. Through gating mechanisms and memory units, temporal dependencies are accurately captured, effectively predicting future trajectories. Finally, a transposed convolutional decoder and potential-weighted generation of a repulsive field are combined. This invention significantly improves the accuracy and continuity of trajectory prediction in complex dynamic environments, enhancing the physical rationality and robustness of obstacle avoidance strategies.

[0028] In this embodiment, S4 specifically includes: S41. The privacy-aware field and the repulsive potential energy field at the current time step are mapped one-to-one according to the grid position, and the values ​​in the same grid cell are added element by element to generate a comprehensive potential energy field containing static environmental information and dynamic hazard prediction. S42. In the comprehensive potential energy field, taking the grid cell where the indoor robot is currently located as the starting point, calculate the potential energy difference between it and the eight adjacent grid cells around it, and determine the direction of the fastest potential energy decrease and the corresponding adjacent grid cells. S43. Move the indoor robot to the adjacent grid cell with the fastest potential energy decrease, and use the grid cell as the new starting point. Repeat the process of calculating and judging the potential energy difference to gradually build a path from the starting point to the target point, which is composed of multiple grid cells. S44. Convert the constructed path into obstacle avoidance control commands for the linear velocity and angular velocity of the indoor robot, and send the obstacle avoidance control commands to the motion controller of the indoor robot to perform predictive obstacle avoidance operations.

[0029] In this embodiment, S5 specifically includes: S51. Record the comprehensive potential energy field at the moment of executing the obstacle avoidance control command as the execution field, and record the actual motion trajectory of the indoor robot after executing the obstacle avoidance control command as the execution result. S52. Concatenate the execution start and end of the execution process and the execution result in chronological order to form a long data vector. Perform differential privacy processing, initialize a pseudo-random number generator, and use the current time step as the seed to generate a uniformly distributed random number between 0 and 1. Substitute the random number generator into the inverse cumulative distribution function of the Laplace distribution for calculation to obtain a Laplace distribution noise value. S53. Add the Laplace distribution noise value to each value in the long data vector to generate an anonymous data vector after differential privacy processing. Hash the anonymous data vector after differential privacy processing to generate an anonymous decision gene fragment of a set length, and upload it to the cloud server through an encrypted communication channel.

[0030] In this embodiment, the inverse cumulative distribution function of the Laplace distribution specifically includes: A pre-set scale parameter determines the range of noise fluctuations and is directly related to the strength of privacy protection. At the start of the calculation, it is determined whether the uniformly distributed random number between 0 and 1 is less than 0.5 or greater than or equal to 0.5. If it is greater than or equal to 0.5, the second calculation path is entered. The number 1 is subtracted from the current uniformly distributed random number, and the result is multiplied by 2. Then, a natural logarithm operation is performed to obtain a negative number. Multiplying this negative number by -1 gives a positive number. Multiplying this positive number by a pre-set scaling parameter gives a Laplace distribution noise value. If the initial uniformly distributed random number is less than 0.5, multiply the uniformly distributed random number by 2 and directly perform a natural logarithm operation on the result. Multiplying the logarithm result by a pre-set scaling parameter also gives a Laplace distribution noise value.

[0031] In this embodiment, S6 specifically includes: S61. In the cloud server, all received anonymous decision gene fragments are collected. Each decision gene fragment is input into a preset three-layer fully connected neural network and calculated with the corresponding preset weight matrix and bias vector. The ReLU activation function is applied element by element to generate a final behavior descriptor that reduces the dimensionality of the intermediate feature vector output by the first layer to one-quarter. A unique behavior descriptor is generated for each decision gene fragment. S62. Initialize a two-dimensional grid structure in memory as a behavior-performance map. Set the horizontal axis of the behavior-performance map to represent the first dimension of the behavior descriptor and the vertical axis to represent the second dimension of the behavior descriptor. Divide the horizontal axis and the vertical axis into a preset number of equally spaced intervals to form a grid. Each grid is a file unit. S63. For each decision gene segment, extract the values ​​of the first and second dimensions from the corresponding behavioral descriptor, and normalize them to an integer range from 0 to a preset number minus one, to obtain a horizontal axis index and a vertical axis index. S64. Mark the file unit corresponding to the coordinate index as the target unit, calculate the obstacle avoidance success rate of the current decision gene segment when it is executed locally as the performance index, store the current decision gene segment and the corresponding performance index together into the target unit. If there are other segments in the target unit, compare the performance index and only keep the one with the higher performance index. S65. For each newly arrived decision gene fragment, obtain the behavior descriptor, traverse all non-empty file units in the behavior-performance map, read the behavior descriptor of the existing fragments in each file unit, calculate the Euclidean distance between the behavior descriptor of the newly arrived decision gene fragment and the behavior descriptor of each existing fragment, and sum all the Euclidean distance values ​​to obtain a total value, which is the scene novelty factor. S66. Set a novelty weight coefficient, add the performance index of each decision gene segment to the scenario novelty factor multiplied by the corresponding novelty weight coefficient, generate a comprehensive fitness score, establish an empty parent candidate pool, sort all decision gene segments according to their corresponding comprehensive fitness scores from high to low, select the top preset number of decision gene segments with the highest comprehensive fitness scores, and put them into the parent candidate pool. S67. From the parent candidate pool, randomly select two parent decision gene segments each time and perform a crossover operation with a preset probability. That is, randomly generate a crossover position and swap the parts of the two segments after the crossover position to generate two new offspring segments. Perform a Gaussian mutation operation with a probability of 1 minus the preset probability. For each value in the decision gene segment, add a noise value randomly drawn from a Gaussian distribution with a mean of 0 and a variance of 1 minus the preset probability to generate a new offspring decision gene segment. S68. For each newly generated offspring decision gene fragment, repeat the steps to calculate the new behavioral descriptor, performance index, and overall fitness score. Based on the new behavioral descriptor, determine the target unit in the behavior-performance map. If the overall fitness score of the offspring decision gene fragment is higher than the overall fitness score of the existing decision gene fragment in the target unit, replace the existing decision gene fragment with the offspring decision gene fragment; otherwise, discard the offspring decision gene fragment. S69. When the average comprehensive fitness score of all file units in the behavior-performance map does not increase for a preset number of consecutive iterations, the evolution process is terminated. Each file unit is traversed, the decision gene fragment with the highest comprehensive fitness score is read, and all the read decision gene fragments are spliced ​​together according to the coordinate order of the file unit to form a complete set of strategies, which is the new generation of global obstacle avoidance strategy model.

[0032] This implementation achieves efficient co-evolution of obstacle avoidance strategies for robot swarms by introducing an improved MAP-Elites algorithm and a scene novelty factor. A fully connected neural network is used to map decision gene fragments to behavioral descriptors, constructing a multi-dimensional behavior-performance map. The scene novelty factor is quantified using Euclidean distance, and a comprehensive fitness score is generated. The mechanism encourages the exploration of diverse behaviors, avoids strategy homogenization, and generates high-performance decision gene combinations covering multiple scenarios through cross-mutation. Finally, these combinations are assembled to generate a new generation of global obstacle avoidance strategy models. This invention significantly improves the generalization ability and adaptability of the strategy, ensuring high obstacle avoidance success rate and system robustness in complex dynamic environments.

[0033] In this embodiment, S7 specifically includes: S71. On the indoor robot end, a new generation of global obstacle avoidance strategy model is received from the cloud server through an encrypted communication channel and written into the local memory to complete the update of the global obstacle avoidance strategy model. S72. During the operation of the indoor robot, predictive obstacle avoidance control commands are continuously generated and output. The obstacle avoidance control commands are sent to the motion controller of the indoor robot, and the motion controller drives the indoor robot to execute the obstacle avoidance control commands, thereby realizing collision prevention and control based on a new generation of global obstacle avoidance strategy.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a leading smart logistics unmanned warehouse project in a province. This unmanned warehouse covers an area of ​​over 100,000 square meters and deploys more than 500 identical autonomous mobile indoor robots responsible for transporting goods from shelves to workstations. The warehouse environment is complex, containing over 10,000 dynamically changing storage locations, hundreds of mobile workstations, and frequently accessed mixed pedestrian and vehicle passageways. The indoor robots need to operate 24 / 7, handling over 200,000 transport tasks daily. Their operational efficiency and safety directly affect the throughput capacity and operating costs of the entire warehousing and logistics system.

[0035] Prior to the deployment of this invention, the unmanned warehouse employed an obstacle avoidance scheme based on traditional SLAM and a static rule base. While this scheme performed adequately with static obstacles, it revealed numerous problems when dealing with dynamic environments. Furthermore, to optimize the algorithm, the old system periodically uploaded raw point cloud data and local maps of the area surrounding the indoor robot. This not only placed enormous pressure on data transmission and storage but also raised serious concerns among customers regarding the potential leakage of trade secrets such as the warehouse's internal cargo layout and stored product categories. The traditional method achieved an average obstacle avoidance success rate of 95.8%, with an average task completion time 18% longer than the theoretical optimum. Moreover, during peak hours, the damage rate of goods caused by emergency braking of the indoor robot reached as high as 0.5%.

[0036] In this embodiment, the present invention is fully deployed in the indoor robot operating system and cloud scheduling platform of the unmanned warehouse. On the indoor robot end-user, each AMR's image sensor integrates a homomorphic encryption preprocessing module. During the robot's movement, the environmental visual data collected in real-time undergoes Hamming distance calculation with a preset obstacle semantic label template within an encrypted domain before entering the main processor. This directly generates semantic obstacle features containing obstacle semantic labels, relative spatial positions, and velocity vectors in an encrypted state. These features are used to construct a dynamic privacy-aware field, whose data does not contain any identifiable original image information, fundamentally eliminating the risk of visual privacy leakage within the warehouse.

[0037] Based on the privacy-aware field, the improved ConvLSTM model begins to play a crucial role. The model introduces an interaction potential factor that is dynamically calculated based on the distance and type of obstacles. This allows indoor robots not only to see the positions of other indoor robots and people, but also to predict their intended trajectories in the next few seconds by combining physical laws.

[0038] After the indoor robot executes a series of obstacle avoidance control commands, the system records the combined potential energy field at the moment of execution and the subsequent actual motion trajectory. These two data points are concatenated into a long vector, and then processed for differential privacy by adding Laplace noise, before being hashed and encoded into an anonymous decision gene fragment. This fragment acts like a desensitized summary of experience, containing only abstract information about how a certain decision led to a certain result in a similar situation, completely excluding specific warehouse maps, cargo locations, or personnel identities. These anonymous decision gene fragments are encrypted and uploaded to a cloud server. In the cloud, after aggregating massive decision gene fragments from 500 indoor robots across the warehouse, the improved MAP-Elites algorithm begins its evolutionary process of collective intelligence. The algorithm first generates a behavior descriptor for each fragment using a three-layer fully connected neural network and places it into a 50x50 behavior-performance map. Unlike traditional algorithms that only pursue obstacle avoidance success rates, the scenario novelty factor introduced in this invention gives higher rewards to novel behaviors appearing in "sparse areas" of the map. To quantify the beneficial effects of this invention, we tracked and compared key performance indicators before and after deployment for several months; the specific data is shown in the table below. Table 1. Performance Comparison Data of the Invention and Traditional Methods in Smart Logistics Unmanned Warehouse Scenarios

[0039] Based on the comparative data shown in Table 1 above, the revolutionary improvements brought about by this invention are clearly evident. The obstacle avoidance success rate has increased from 95.8% to 99.7%, meaning that task failures and equipment damage caused by collisions are virtually eliminated. Most notably, traffic jams have been drastically reduced from 15.2 per shift to 0.3, essentially resolving the long-standing efficiency bottleneck plaguing the warehouse. Due to the smoother indoor robot movement trajectory, the damage rate of goods caused by emergency braking has decreased by 92%, directly preventing significant economic losses.

[0040] In terms of data security and efficiency, the daily uploaded data volume plummeted from 52.3TB to 1.8TB, a reduction of 96.6%. This not only represents a significant saving in network bandwidth and storage costs, but more importantly, all uploaded data consists of anonymized fragments processed with both homomorphic encryption and differential privacy, completely eliminating customer concerns about the leakage of trade secrets. The evolution cycle of cloud strategies has also been shortened from 90 days of manual version iteration to 14 days of fully automated evolution, enabling the entire indoor robot fleet to quickly adapt to dynamic needs such as warehouse layout adjustments and business process changes.

[0041] Overall, in this implementation case of a smart logistics unmanned warehouse, the present invention successfully addresses the core pain points of traditional indoor robot systems in dynamic and complex environments—such as low efficiency, poor coordination, and high risk of privacy leakage—through the collaboration of edge-side privacy computing and cloud-based swarm intelligence. Its superior performance across multiple dimensions, including obstacle avoidance success rate, operational efficiency, data security, and system self-evolution capabilities, fully demonstrates the invention's high practical value and broad industry application prospects.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A collision prevention method for indoor robots based on privacy computing, characterized in that, Includes the following steps: S1. Real-time acquisition of environmental visual data through homomorphic encryption preprocessing to generate semantic obstacle features; S2. Construct a privacy-aware field using semantic obstacle features and generate a privacy-aware field vector; S3. Based on the privacy-aware field, an interactive potential energy factor is introduced. Through an improved ConvLSTM model, the field vector labeled as dynamic obstacles is continuously tracked and its changes are analyzed to predict the future intention trajectory of the dynamic obstacles and generate a repulsive potential energy field. S4. The privacy-aware field and the repulsive potential energy field are integrated to form a comprehensive potential energy field. Based on the comprehensive potential energy field, obstacle avoidance control commands are generated and executed by finding the path with the fastest potential energy gradient descent. S5. Record the execution start and end times and execution results of obstacle avoidance control commands, perform differential privacy processing, generate decision gene fragments, and upload them to the cloud server; S6. In the cloud server, decision gene fragments from multiple indoor robots are collected. A behavior-performance map is constructed using the improved MAP-Elites algorithm. A scene novelty factor is introduced to quantify the difference between the decision gene fragments and the behavior-performance map. High-performance decision gene combinations covering multiple scenes are selected to generate a new generation of global obstacle avoidance strategy models. S7. Distribute the new generation of global obstacle avoidance strategy model to the indoor robot to continuously generate and execute obstacle avoidance control commands.

2. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, S1 specifically includes: S11. On the indoor robot end, each pixel in the environmental visual data collected by the homomorphic encryption preprocessing integrated into the image sensor is calculated in the encrypted domain with the Hamming distance of the preset obstacle semantic label template. Pixel areas with Hamming distance less than the set distance threshold are marked as obstacles and assigned corresponding semantic labels to generate semantic labels for obstacles. S12. Within the encrypted domain, a stereo vision matching algorithm within the encrypted domain is used to perform pixel-level matching of the left and right eye images in the acquired environmental visual data, calculate the disparity value of each obstacle pixel, convert the disparity value into a depth value based on the camera intrinsic parameters, and calculate the relative spatial position of the obstacle by combining the pixel coordinates. S13. Within the encrypted domain, for the same obstacle marked in two consecutive frames of environmental visual data, the pixel displacement of the obstacle on the image plane is calculated using the optical flow method. Combined with the relative spatial position of the obstacle, the pixel displacement is converted into the actual spatial displacement and divided by the time interval between the two frames to obtain the velocity vector of the obstacle. S14. Combine the semantic labels, relative spatial positions, and velocity vectors of obstacles in sequence to generate semantic obstacle features.

3. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, S2 specifically includes: S21. Map the relative spatial position of semantic obstacle features to a two-dimensional grid map of a preset size. Each grid cell stores the semantic labels of all obstacles within the coverage area, generating a static privacy-aware field containing location and semantic information. S22. Fill the velocity vector in the semantic obstacle features into the corresponding grid cell in the static privacy-aware field. Each grid cell contains position, semantic and velocity information, generating a dynamic privacy-aware field. S23. Arrange the semantic tags, relative spatial positions, and velocity vectors stored in each grid cell of the privacy-aware field into a one-dimensional vector in a preset order to generate a privacy-aware field vector.

4. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, S3 specifically includes: S31. Based on the privacy-aware field, filter out grid cells with semantic tags as dynamic obstacles, read the relative spatial position and velocity vector of the dynamic obstacle in each grid cell, calculate the Euclidean distance between each dynamic obstacle and the current position of the indoor robot, and divide the preset attenuation constant by the Euclidean distance to obtain a distance quotient. S32. Using the natural constant e as the base and the opposite of the distance quotient as the exponent, perform an exponential operation to obtain an intermediate value that decreases as the distance increases. Multiply the intermediate value by a preset intensity constant to obtain a final scalar repulsive force value, which is the interaction potential energy factor. S33. Using the improved ConvLSTM model, the privacy-aware field vector of the current time step is reshaped into two-dimensional grid data. The data is input into a convolutional encoder with a pre-set kernel size. The feature map is generated by the sliding window convolution in the first layer. The ReLU activation function is applied to each feature map. Then, the high-dimensional spatial feature map is output by convolution with the same kernel size and the ReLU activation function. S34. The spatial feature map and the interaction potential factor map at time step t are concatenated along the channel dimension to form a concatenated input tensor, which is then input into the input gate, forget gate and output gate respectively. Each gate independently performs a convolution operation with a preset kernel size, and a weight matrix between 0 and 1 is generated by the Sigmoid activation function. S35. Input the concatenated input tensor into the candidate memory layer of the improved ConvLSTM model, perform a convolution operation of a preset size, and generate a candidate value through the Tanh activation function. Multiply the unit state output at time step t-1 element-wise with the weight matrix of the forget gate, and add the result to the element-wise product of the weight matrix of the input gate and the candidate value to update the unit state at time step t. S36. Apply the Tanh activation function to the cell state at time step t, multiply the result element-wise with the weight matrix of the output gate, output the hidden state at time step t, input the hidden state output at the last time step into the convolutional decoder of the improved ConvLSTM model, perform upsampling convolution through the transposed convolutional kernel of the first layer with the same kernel size as the convolutional encoder, and generate a feature map. S37. Apply the ReLU activation function to each feature map, and then perform upsampling convolution through the transposed convolution kernel of the same size to generate a feature map, which is an initial potential energy field containing future trajectory prediction with the same size as the original privacy-aware field. The value of each grid cell in the initial potential energy field is weighted and summed with the interaction potential energy factor at the corresponding position to generate a repulsive potential energy field that evolves over time and contains predicted trajectory information.

5. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, S4 specifically includes: S41. The privacy-aware field and the repulsive potential energy field at the current time step are mapped one-to-one according to the grid position, and the values ​​in the same grid cell are added element by element to generate a comprehensive potential energy field containing static environmental information and dynamic hazard prediction. S42. In the comprehensive potential energy field, taking the grid cell where the indoor robot is currently located as the starting point, calculate the potential energy difference between it and the eight adjacent grid cells around it, and determine the direction of the fastest potential energy decrease and the corresponding adjacent grid cells. S43. Move the indoor robot to the adjacent grid cell with the fastest potential energy decrease, and use the grid cell as the new starting point. Repeat the process of calculating and judging the potential energy difference to gradually build a path from the starting point to the target point, which is composed of multiple grid cells. S44. Convert the constructed path into obstacle avoidance control commands for the linear velocity and angular velocity of the indoor robot, and send the obstacle avoidance control commands to the motion controller of the indoor robot to perform predictive obstacle avoidance operations.

6. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, S5 specifically includes: S51. Record the comprehensive potential energy field at the moment of executing the obstacle avoidance control command as the execution field, and record the actual motion trajectory of the indoor robot after executing the obstacle avoidance control command as the execution result. S52. Concatenate the execution start and end of the execution process and the execution result in chronological order to form a long data vector. Perform differential privacy processing, initialize a pseudo-random number generator, and use the current time step as the seed to generate a uniformly distributed random number between 0 and 1. Substitute the random number generator into the inverse cumulative distribution function of the Laplace distribution for calculation to obtain a Laplace distribution noise value. S53. Add the Laplace distribution noise value to each value in the long data vector to generate an anonymous data vector after differential privacy processing. Hash the anonymous data vector after differential privacy processing to generate an anonymous decision gene fragment of a set length, and upload it to the cloud server through an encrypted communication channel.

7. The indoor robot collision prevention method based on privacy computing according to claim 6, characterized in that, The inverse cumulative distribution function of the Laplace distribution specifically includes: With pre-set scale parameters, at the start of the calculation, it is determined whether the uniformly distributed random number between 0 and 1 is less than 0.5 or greater than or equal to 0.

5. If it is greater than or equal to 0.5, the second calculation path is entered, the number 1 is subtracted from the current uniformly distributed random number, the result is multiplied by 2, and a natural logarithm operation is performed to obtain a negative number. Multiplying this negative number by -1 gives a positive number. Multiplying this positive number by a pre-set scaling parameter gives a Laplace distribution noise value. If the initial uniformly distributed random number is less than 0.5, multiply the uniformly distributed random number by 2 and directly perform a natural logarithm operation on the result. Multiplying the logarithm result by a pre-set scaling parameter also gives a Laplace distribution noise value.

8. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, S6 specifically includes: S61. In the cloud server, all received anonymous decision gene fragments are collected. Each decision gene fragment is input into a preset three-layer fully connected neural network and calculated with the corresponding preset weight matrix and bias vector. The ReLU activation function is applied element by element to generate a final behavior descriptor that reduces the dimensionality of the intermediate feature vector output by the first layer to one-quarter. A unique behavior descriptor is generated for each decision gene fragment. S62. Initialize a two-dimensional grid structure in memory as a behavior-performance map. Set the horizontal axis of the behavior-performance map to represent the first dimension of the behavior descriptor and the vertical axis to represent the second dimension of the behavior descriptor. Divide the horizontal axis and the vertical axis into a preset number of equally spaced intervals to form a grid. Each grid is a file unit. S63. For each decision gene segment, extract the values ​​of the first and second dimensions from the corresponding behavioral descriptor, and normalize them to an integer range from 0 to a preset number minus one, to obtain a horizontal axis index and a vertical axis index. S64. Mark the file unit corresponding to the coordinate index as the target unit, calculate the obstacle avoidance success rate of the current decision gene segment when it is executed locally as the performance index, store the current decision gene segment and the corresponding performance index together into the target unit. If there are other segments in the target unit, compare the performance index and only keep the one with the higher performance index. S65. For each newly arrived decision gene fragment, obtain the behavior descriptor, traverse all non-empty file units in the behavior-performance map, read the behavior descriptor of the existing fragments in each file unit, calculate the Euclidean distance between the behavior descriptor of the newly arrived decision gene fragment and the behavior descriptor of each existing fragment, and sum all the Euclidean distance values ​​to obtain a total value, which is the scene novelty factor. S66. Set a novelty weight coefficient, add the performance index of each decision gene segment to the scenario novelty factor multiplied by the corresponding novelty weight coefficient, generate a comprehensive fitness score, establish an empty parent candidate pool, sort all decision gene segments according to their corresponding comprehensive fitness scores from high to low, select the top preset number of decision gene segments with the highest comprehensive fitness scores, and put them into the parent candidate pool. S67. From the parent candidate pool, randomly select two parent decision gene segments each time and perform a crossover operation with a preset probability. That is, randomly generate a crossover position and swap the parts of the two segments after the crossover position to generate two new offspring segments. Perform a Gaussian mutation operation with a probability of 1 minus the preset probability. For each value in the decision gene segment, add a noise value randomly drawn from a Gaussian distribution with a mean of 0 and a variance of 1 minus the preset probability to generate a new offspring decision gene segment. S68. For each newly generated offspring decision gene fragment, repeat the steps to calculate the new behavioral descriptor, performance index, and overall fitness score. Based on the new behavioral descriptor, determine the target unit in the behavior-performance map. If the overall fitness score of the offspring decision gene fragment is higher than the overall fitness score of the existing decision gene fragment in the target unit, replace the existing decision gene fragment with the offspring decision gene fragment; otherwise, discard the offspring decision gene fragment. S69. When the average comprehensive fitness score of all file units in the behavior-performance map does not increase for a preset number of consecutive iterations, the evolution process is terminated. Each file unit is traversed, the decision gene fragment with the highest comprehensive fitness score is read, and all the read decision gene fragments are spliced ​​together according to the coordinate order of the file unit to form a complete set of strategies, which is the new generation of global obstacle avoidance strategy model.

9. The indoor robot collision prevention method based on privacy computing according to claim 1, characterized in that, Specifically, S7 includes: S71. On the indoor robot end, a new generation of global obstacle avoidance strategy model is received from the cloud server through an encrypted communication channel and written into the local memory to complete the update of the global obstacle avoidance strategy model. S72. During the operation of the indoor robot, predictive obstacle avoidance control commands are continuously generated and output. The obstacle avoidance control commands are sent to the motion controller of the indoor robot, and the motion controller drives the indoor robot to execute the obstacle avoidance control commands.

Images