Robot AI model dynamic compression method and control system

Abstract

The invention discloses a robot AI model dynamic compression method and control system, and the method comprises the steps: obtaining the hardware state data and task type of a robot, and carrying out the normalization processing and storage of the hardware state data; based on the hardware state data and / or the task type, selecting a corresponding compression strategy in real time to compress the AI model so as to obtain a lightweight model, the compression strategy comprises at least one of compression of the AI model based on the hardware state data, compression of the AI model based on the hardware state data in combination with the reward function and the task type, and compression of the AI model based on the task type and / or the hardware state data. According to the method, the computing power of the robot control system, the complexity of the AI model and the computing speed can be balanced, the problem of how to effectively operate the AI small model in the mobile robot control system is solved, and the balance between the robot control precision and the speed is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile robots, and in particular to a robot AI model dynamic compression method and control system. Background Art

[0002] The application of mobile robots in manufacturing scenarios has enhanced the intelligence and flexibility of automated production lines, and their adoption is becoming increasingly widespread. When performing mobile tasks in manufacturing scenarios, mobile robots require complex signal processing, data processing, and control processes, including environmental perception, task instruction parsing, automatic obstacle avoidance, and autonomous decision-making. Currently, many mobile robots rely on traditional digital processing systems. Even though they utilize industrial computer systems with high computing speeds and bandwidth as the robot's controller, this has significantly improved the robot's computing capabilities and can run AI neural network algorithm models to enhance the robot's intelligence.

[0003] However, with the increasing demands for high-speed cycles and dynamically changing complex scenarios in manufacturing, mobile robots face increasing challenges in real-time computing speed, control response time, and autonomous decision-making performance during operation. In most cases, robots are required to have both high-speed response performance and meet higher standards in navigation accuracy, perception accuracy, and autonomous decision-making performance. However, many existing robots struggle to meet the growing computing demands due to limited controller hardware resources. This is especially true when running large-scale and complex AI neural network algorithm models. Due to the limitations of the platform's computing power and model size, achieving good computing results is difficult, affecting the robot's overall performance.

[0004] Therefore, how to dynamically compress the AI ​​model so that it can run effectively based on the robot controller hardware platform and achieve better performance has become an issue that urgently needs to be studied and solved. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides a robot AI model dynamic compression method and control system to at least solve the problem that existing robots cannot achieve good computing results after deploying AI models.

[0006] The embodiments of the present invention provide the following technical solutions:

[0007] An embodiment of the present invention provides a method for dynamic compression of a robot AI model, which is applied to a robot deployed with an AI model, including:

[0008] Acquiring hardware status data and a task type of the robot, normalizing and storing the hardware status data, wherein the hardware status data includes at least one of CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate;

[0009] Based on the hardware status data and / or the task type, a corresponding compression strategy is selected in real time to compress the AI ​​model to obtain a lightweight model. The compression strategy includes compressing the AI ​​model based on the hardware status data, compressing the AI ​​model based on the hardware status data combined with the reward function and the task type, and compressing the AI ​​model based on the task type and / or the hardware status data. Compressing the AI ​​model includes at least one of pruning, quantizing, and skipping the AI ​​model.

[0010] Furthermore, compressing the AI ​​model based on the hardware status data includes:

[0011] A maximum parameter index of the AI ​​model is set based on the hardware status data to adjust the maximum parameter index of the AI ​​model according to different hardware status data, where the maximum parameter index includes at least one of a maximum parameter amount and a maximum FPS.

[0012] Furthermore, a preset threshold is set for the hardware status data, and the preset threshold includes at least one of a CPU utilization threshold, a CPU temperature threshold, a remaining memory value threshold, and a communication channel blocking rate threshold.

[0013] Furthermore, the hardware status data is combined with the reward function and the task type to compress the AI ​​model, including:

[0014] Setting a reward function, wherein the reward function takes the model parameter amount of the AI ​​model, the maximum FPS of the model, the maximum error corresponding to the robot parameters, and the time delay of a certain performance of the AI ​​model as parameters, and setting weight factors for the model parameter amount, the maximum FPS of the model, the maximum error corresponding to the robot parameters, and the time delay of a certain performance of the AI ​​model respectively;

[0015] After obtaining the model parameter amount and FPS parameter of the AI ​​model based on the reward function, the AI ​​model is compressed in combination with the task type.

[0016] Furthermore, corresponding compression weights are set for pruning, quantization, and skipping according to different task types, and the AI ​​model is compressed according to the compression weights;

[0017] Wherein, when the task type is mobile navigation and obstacle avoidance, the compression weight level of pruning is set to low, the compression weight level of quantization is set to medium, and the skip weight is set to high; and / or

[0018] When the task type is a terminal operation, the compression weight level of pruning is set to medium, the compression weight level of quantization is set to low, and the skip weight is set to none; and / or

[0019] When the task type is scene perception, setting the compression weight level of pruning to high, the compression weight level of quantization to medium, and the skip weight to low; and / or

[0020] When the task type is autonomous planning, the compression weight level of pruning is set to high, the compression weight level of quantization is set to high, and the skip weight is set to medium; and / or

[0021] When the task type is sudden obstacle avoidance, the compression weight level of pruning is set to low, and the compression weight level and skip weight of quantization are both set to high; and / or

[0022] When the task type is a high-precision operation, the compression weight level of pruning and the compression weight level of quantization are both set to medium, and the skip weight is set to low.

[0023] Furthermore, compressing the AI ​​model based on the task type and / or the hardware status data includes:

[0024] According to the working scene, task nature and complexity, the robot's task types are divided into navigation task type, obstacle avoidance behavior type, operation task type and data task type. The navigation task type includes long-distance sparse scene navigation, few-obstacle, static environment navigation, conventional navigation and dynamically changing scene navigation; the obstacle avoidance behavior type includes conventional automatic obstacle avoidance and sudden / emergency obstacle avoidance; the operation task type includes material grasping operation task, material transfer task, movement + operation collaborative task, high-speed beat movement operation, dynamic mutation operation, conventional precision operation task, multi-task and human-machine collaboration task; the data task includes battery power consumption data, dense data flow occupancy task, real-time perception data task and low-precision computing task;

[0025] The AI ​​model is compressed according to the task type of the robot.

[0026] Furthermore, the CPU utilization, CPU temperature, remaining memory value and communication channel blocking rate are monitored, and when each parameter reaches a preset threshold, the AI ​​model is compressed.

[0027] Furthermore, when there are multiple areas in the same work scene, if the multiple areas are divided into multiple task types, the robot executes corresponding compression methods according to different task types in the multiple areas.

[0028] Furthermore, the method further comprises:

[0029] Running the lightweight model and extracting model performance parameters, wherein the model performance parameters include accuracy, delay, and response time;

[0030] When the model performance parameters do not meet the standards, the AI ​​model is rolled back to the original configuration state.

[0031] An embodiment of the present invention further provides a mobile robot AI model dynamic compression control system, comprising:

[0032] a hardware status monitoring unit, configured to obtain hardware status data and task type of the robot, normalize and store the hardware status data, wherein the hardware status data includes at least one of CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate;

[0033] a compression strategy selection unit, configured to select a corresponding compression strategy in real time based on the hardware status data and / or the task type, wherein the compression strategy comprises at least one of compressing the AI ​​model based on the hardware status data, compressing the AI ​​model based on the hardware status data combined with a reward function and the task type, and compressing the AI ​​model based on the task type and / or the hardware status data;

[0034] A model dynamic adjustment unit is used to compress the AI ​​model, where compressing the AI ​​model includes at least one of pruning, quantizing, and skipping the AI ​​model.

[0035] Compared with the prior art, the at least one technical solution adopted in the embodiment of the present invention can achieve the following beneficial effects:

[0036] The present invention provides a dynamic compression method for a robot AI model. By detecting the real-time hardware status data of the robot controller and dynamically adjusting the AI ​​model structure, it achieves an effective balance between control accuracy and inference speed, ensures the real-time calculation of high-priority tasks, and solves the problem of effectively running small AI models in the control system of mobile robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0038] Figure 1 This is a flowchart of a method for dynamic compression of a robot AI model according to an embodiment of the present invention;

[0039] Figure 2 This is a flowchart of a method for dynamic compression of a robot AI model according to an embodiment of the present invention;

[0040] Figure 3 Schematic diagram of a compression strategy according to an embodiment of the present invention (1);

[0041] Figure 4 Schematic diagram of the compression strategy according to an embodiment of the present invention (II);

[0042] Figure 5 This is a structural block diagram of a dynamic compression control system of a mobile robot AI model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0044] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the features in the following embodiments and embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative work are within the scope of protection of this application.

[0045] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0046] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0047] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples, however, one skilled in the art will appreciate that the examples can be practiced without these specific details.

[0048] The application of mobile robots in manufacturing scenarios has enhanced the intelligence and flexibility of automated production lines, and their applications are becoming more and more widespread.

[0049] However, when performing mobile tasks in manufacturing scenarios, mobile robots face complex signal processing, data processing, and control processes, including environmental perception, task instruction parsing, automatic obstacle avoidance, and autonomous decision-making. During operation, factors such as the high cycle times and dynamic, complex environments often place higher demands and challenges on the robots' real-time computing speed, control response time, and autonomous decision-making performance. In many cases, robots are required to possess both high-speed response performance and meet requirements in navigation precision, perception accuracy, and autonomous decision-making performance, which is undoubtedly a difficult task for existing robots.

[0050] Many existing robots rely on traditional digital processing systems. For example, even if industrial computer systems with high computing speeds and bandwidth are used as robot controllers, they still struggle to cope with the increasing computing demands. Furthermore, intelligent robot control systems must also run AI neural network algorithm models to enhance the robot's intelligence level.

[0051] The size of the AI ​​neural network algorithm model places extremely high demands on the robot controller, including the hardware resources of the computing platform and the size and complexity of the model itself. Therefore, it is necessary to study how to enable the AI ​​model to achieve better computing results on a limited robot controller, thereby enhancing the performance of the robot; in particular, how to dynamically compress the AI ​​model so that the model can effectively run based on the robot controller hardware platform to increase the performance of the robot.

[0052] In view of this, the inventors conducted in-depth research and improvements on the robot's hardware status data, the robot's task type, the AI ​​model performance data, and the AI ​​model compression method, and found that: by obtaining the robot's hardware status data and task type, and forming compression strategies for multiple AI models based on the hardware status data and task type, and selecting corresponding compression strategies according to actual needs, it is possible to achieve a balance among computing power, AI model complexity, and computing speed, so that the multi-task processing requirements of the mobile robot control system can achieve better performance when the computing power of the mobile robot is limited.

[0053] Based on this, this specification proposes a solution: Figure 1As shown, a dynamic compression method for a robot AI model of the present invention proposes a selection strategy for dynamically compressing the mobile robot AI model based on the robot's hardware status data and task type. That is, the task type is subdivided in combination with the different types of control tasks of the mobile robot, and each type of subdivided task is associated and set with the specific compression method (pruning, quantization, skipping) of the AI ​​small model. Through this correlation / mapping relationship, the AI ​​model is dynamically compressed according to the different working conditions of the robot, reducing the difficulty of dynamically compressing the AI ​​model in the actual system.

[0054] The following describes the technical solutions provided by various embodiments of the present application in conjunction with the accompanying drawings.

[0055] like Figures 1 to 4 As shown, an embodiment of the present invention provides a robot AI model dynamic compression method, which is applied to a robot deployed with an AI model, including:

[0056] Step S102: Obtain hardware status data and task type of the robot, normalize and store the hardware status data, where the hardware status data includes at least one of CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate;

[0057] Step S104: Based on the hardware status data and / or task type, a corresponding compression strategy is selected in real time to compress the AI ​​model to obtain a lightweight model. The compression strategy includes at least one of compressing the AI ​​model based on the hardware status data, compressing the AI ​​model based on the hardware status data combined with the reward function and the task type, and compressing the AI ​​model based on the task type and / or hardware status data; compressing the AI ​​model includes at least one of pruning, quantizing, and skipping the AI ​​model.

[0058] Among them, pruning mainly reduces model parameters. It is suitable for tasks with large computational complexity and not very sensitive to the number of model parameters, such as navigation. Reducing a certain number of model parameters will not have a big impact on navigation accuracy, but will improve computing speed.

[0059] Among them, quantization is to reduce the precision of numerical values, such as reducing them from floating-point numbers to integer numbers. While saving computing hardware resources, the reduction of numerical precision can increase computing speed and reduce energy consumption. It is suitable for scenarios that are sensitive to memory and energy consumption.

[0060] Among them, skipping is to dynamically skip the calculation of certain layers in the model, which is more suitable for situations that require fast response and sudden high load, such as high-speed dynamic emergency obstacle avoidance operations.

[0061] Among them, the present invention provides at least three compression strategies, which enable the robot to reasonably select the corresponding compression strategy according to different hardware status data and task types, so that the AI ​​model can be reasonably compressed according to its own computing power and working conditions to obtain a lightweight model.

[0062] Among them, selecting the optimal compression strategy based on the real-time hardware status and task type can enable the AI ​​model to dynamically adapt to robots in different working conditions and hardware states, avoiding resource overload.

[0063] Among them, by normalizing the hardware status data, the hardware status data can be stored in a standardized manner to facilitate unified threshold judgment and historical data analysis.

[0064] Among them, when compressing the AI ​​model, it can support compression methods such as pruning, quantization, and skipping, and can reasonably compress the AI ​​model under different hardware states and task types, so that the AI ​​model can be compatible with robots with different computing power or different states.

[0065] The present invention provides a dynamic compression method for a robot AI model, proposes a core framework for a dynamic compression AI model, and adjusts the compression strategy based on hardware status and task type, solving the problem that traditional robot controllers have difficulty balancing computing power, model complexity, and real-time requirements, resulting in insufficient performance of AI models in highly dynamic scenarios.

[0066] In some embodiments, compressing the AI ​​model based on hardware status data (the first compression strategy provided by the present invention) includes:

[0067] Step S104a: Setting a maximum parameter index of the AI ​​model based on the hardware status data to adjust the maximum parameter index of the AI ​​model according to different hardware status data, where the maximum parameter index includes at least one of a maximum parameter amount and a maximum FPS.

[0068] Among them, by dynamically setting the maximum parameter indicators and maximum FPS of the AI ​​model according to the hardware status data, the robot can avoid overloading the AI ​​model.

[0069] Among them, by dynamically adjusting the parameters and performance indicators of the AI ​​model according to the hardware status data, the robot can also adaptively adjust / compress the AI ​​model.

[0070] For example, when the CPU temperature rises, the model complexity is reduced to prioritize stable system operation.

[0071] Furthermore, a preset threshold is set for the hardware status data, and the preset threshold includes at least one of a CPU utilization threshold, a CPU temperature threshold, a remaining memory value threshold, and a communication channel blocking rate threshold.

[0072] Among them, by triggering the compression mechanism through preset hardware thresholds, it is possible to actively protect hardware resources, prevent the AI ​​model from occupying too much hardware computing resources, causing the robot system to be unable to run stably, and avoid hardware overload.

[0073] Specifically, the model's maximum parameter metrics, such as maximum parameter count and maximum FPS, are set based on f (CPU utilization, remaining memory, CPU temperature, and communication channel data blocking rate). Depending on the specific situation, thresholds for CPU utilization, remaining memory, CPU temperature, and communication channel data blocking rate can be set. Here, f() represents a function, with CPU utilization, remaining memory, and other factors as components of the function, which can have multiple combinations. These combinations can be obtained from actual testing to serve as the basis for strategy selection, or they can be directly presented in a table based on the characteristic attributes of the hardware system.

[0074] Among them, the strategy selection based on the hardware resource status is a basic or framework strategy selection basis, and only provides the scope of compression strategy selection.

[0075] Step S104a constrains the model size through the maximum parameter index to ensure that the compressed model runs within the normal working state of the hardware.

[0076] In some embodiments, hardware status data is combined with reward function and task type to compress the AI ​​model (the second compression strategy provided by the embodiment of the present invention) including:

[0077] Step S104b: Setting a reward function. The reward function uses the model parameter quantity of the AI ​​model, the maximum FPS of the model, the maximum error corresponding to the robot parameters, and the time delay of a certain performance of the AI ​​model as parameters, and sets weight factors for the model parameter quantity, the maximum FPS of the model, the maximum error corresponding to the robot parameters, and a certain performance of the AI ​​model respectively;

[0078] Step S104c: After obtaining the model parameter quantity and FPS parameter of the AI ​​model based on the reward function, compress the AI ​​model in combination with the task type.

[0079] Among them, by setting weight factors for parameters such as the model parameter amount and the maximum FPS of the model, and quantitatively calculating the "optimal solution", multi-objective optimization of the AI ​​model can be achieved, and the AI ​​model can be compressed in combination with the task type. This can make the AI ​​model more reasonably compressed, so that the robot can reasonably compress the AI ​​model under different task types.

[0080] Among them, the reward function can be:

[0081] R=δ1*Npara+δ2*FPS+δ3*ΔE+δ4*Tdelay;

[0082] Among them, δ1, δ2, δ3, and δ4 are all weight factors; FPS is the number of frames per second, which refers to the maximum number of frames that an AI model can process per second under a specific hardware environment. It is a key indicator for measuring the real-time performance of the model; Npara is the model parameter; ΔE is the maximum error (accuracy) corresponding to a certain type of robot parameter, such as mobile navigation error; Tdelay is the time delay of a certain performance indicator of the model.

[0083] Among them, the type and number of parameters in the reward function, as well as the weight of each parameter, can be determined independently based on actual conditions.

[0084] In some of these embodiments, when the AI ​​model is used for a single performance purpose, the reward function may also be composed of one or two parameters that best characterize the performance.

[0085] After the system obtains the appropriate model parameter quantity, FPS and other parameters based on the reinforcement learning results, it uses pruning (by reducing the number of model parameters), quantization, and skipping in the specific model adjustment strategy to dynamically adjust the model size, thereby achieving lightweight model.

[0086] Furthermore, corresponding compression weights can be set for pruning, quantization, and skipping according to different task types, and the AI ​​model can be compressed according to the compression weights.

[0087] Among them, when the task type is mobile navigation and obstacle avoidance, the compression weight level of pruning is set to low, the compression weight level of quantization is set to medium, and the skip weight is set to high. When the task type is terminal operation, the compression weight level of pruning is set to medium, the compression weight level of quantization is set to low, and the skip weight is set to none. When the task type is scene perception, the compression weight level of pruning is set to high, the compression weight level of quantization is set to medium, and the skip weight is set to low. When the task type is autonomous planning, the compression weight level of pruning is set to high, the compression weight level of quantization is set to high, and the skip weight is set to medium. When the task type is sudden obstacle avoidance, the compression weight level of pruning is set to low, and the compression weight level of quantization and the skip weight are both set to high. When the task type is high-precision operation, the compression weight level of pruning and the compression weight level of quantization are both set to medium, and the skip weight is set to low.

[0088] Among them, the higher the compression weight level, the more inclined to execute the corresponding compression method. For example, when the skip weight is high, the skip operation is more inclined to be executed.

[0089] When the task type is a terminal operation, the compression weight level of the skip can be set to high+, where high+ means directly executing the skip and skipping the non-critical semantic layer, while the underlying data is maintained.

[0090] When the compression weight level is zero, the compression method is allowed to be executed. For example, when the skipping compression weight is zero, skipping layers is not allowed.

[0091] Among them, when the task type is autonomous planning, the compression weight level of pruning is high+, which means that the focus is on reducing memory usage.

[0092] Among them, by strongly associating the task type and compression weight, scenario-based compression can be achieved, solving the problem that different tasks have large differences in model performance requirements and require targeted compression of AI models.

[0093] Steps S104b to S104c, by introducing a reward function to optimize the compression strategy, can balance model performance and computing resource consumption, so that the robot can reasonably compress the AI ​​model according to the reward function and task type, making the compression strategy more adaptable to the actual working conditions of the robot, and providing corresponding compression methods for different working conditions of the robot, so that the robot can better balance computing power and AI model complexity under different working conditions or scenarios.

[0094] In some embodiments, compressing the AI ​​model based on task type and / or hardware status data (the third compression strategy provided by the present invention) includes:

[0095] Step S104d: Based on the working scenario, task nature, and complexity, the robot's task types are divided into navigation task types, obstacle avoidance behavior types, operation task types, and data task types. Navigation task types include long-distance sparse scene navigation, navigation in a static environment with few obstacles, conventional navigation, and navigation in a dynamically changing scene. Obstacle avoidance behavior types include conventional automatic obstacle avoidance and sudden / emergency obstacle avoidance. Operation task types include material grasping operation tasks, material transfer tasks, movement + operation collaborative tasks, high-speed beat movement operations, dynamic mutation operations, conventional precision operation tasks, multi-tasking, and human-machine collaboration tasks. Data tasks include battery power consumption data, dense data flow occupancy tasks, real-time perception data tasks, and low-precision computing tasks.

[0096] Step S104f: compress the AI ​​model according to the robot's task type.

[0097] When the robot is running in a scene, navigation, obstacle avoidance, operation, and data services are classified into the above categories and levels according to the specific circumstances of the scene, the nature of the task, and the complexity, so as to better implement the dynamic adjustment strategy (pruning, quantization, skipping) of the subsequent AI model.

[0098] Steps S104d and 104f refine the task classification, enabling the robot to select different compression methods to compress the AI ​​model according to the task classification, so that the robot can reasonably compress the AI ​​model in a variety of complex scenarios.

[0099] Furthermore, compressing the AI ​​model based on task type and / or hardware status data also includes:

[0100] Step S104e: monitor CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate, and compress the AI ​​model when each parameter reaches a preset threshold.

[0101] Step S104e monitors the hardware status data and combines it with task classification, so that the robot can compress the AI ​​model more reasonably based on its own hardware status and task classification, solving the problem of being unable to reasonably compress the AI ​​model by compressing it based on a single element.

[0102] in, Figure 4 The method of compressing AI models based on task type and hardware status data is introduced in detail in the present invention, and only a small explanation is given here.

[0103] For example, the types of tasks that can trigger only pruning include memory occupancy, conventional automatic obstacle avoidance tasks, long-distance sparse scene navigation tasks, and conventional precision operation tasks. Any of the above indicators can trigger pruning execution; the tasks that can trigger pruning and quantization include material grasping operation tasks or material transfer tasks. That is, when the robot performs one or both of the two tasks, the system will perform pruning and quantization on the AI ​​model to ensure that the compressed small model can meet the requirements of the current task and achieve a balance between accuracy (such as grasping accuracy), performance (such as grasping speed) and system computing resources. Others such as Figure 4 As shown in , they are not described here one by one.

[0104] Among them, in different compression methods (pruning, skipping, quantization), manual settings or editable items can be set to manually specify certain task types to execute corresponding compression methods.

[0105] In some embodiments, when there are multiple areas in the robot's working scene, if the multiple areas are divided into multiple task types, the robot executes corresponding compression methods according to different task types in the multiple areas.

[0106] Among them, dynamic strategy switching in multi-region and multi-task scenarios can solve the problem of conflicting task requirements in different regions in the same environment.

[0107] Specifically, in this application, after classifying various types of navigation tasks, obstacle avoidance behaviors, operation tasks, data tasks, etc., different compression methods are used to compress the AI ​​model according to the specific application of the mobile robot.

[0108] Among them, the embodiment of the present invention sets the attribution of various tasks of the robot based on the above classification to facilitate the compressed execution of the model.

[0109] Taking navigation as an example, when a mobile robot is used in a manufacturing scenario, the robot's navigation category in the scenario is defined as conventional navigation based on the judgment of the scenario; or according to different areas in the same scene, it is defined as conventional navigation in area A; entering area B requires switching to dynamically changing scene navigation; thereby optimizing the compression method of the AI ​​model in a targeted manner (pruning is performed in area A, skipping is performed in area B, or two adjustment strategies or all three adjustment strategies are executed in combination with other category attributes).

[0110] In some embodiments, the robot AI model dynamic compression method further includes:

[0111] Step S106: running the lightweight model and extracting model performance parameters, which include accuracy, delay, and response time;

[0112] Step S108: When the model performance parameters do not meet the standards, the AI ​​model is rolled back to the original configuration state.

[0113] Steps S106 and S108 can ensure the reliability of the system through the model performance feedback and rollback mechanism, solve the problem that compression may cause model failure, and improve the safety net capability.

[0114] A robot AI model dynamic compression method of the present invention, based on the compression rules given in this application, can more simply compress the AI ​​model in the robot controller, and through the second compression strategy provided by the present invention, it can first use the reinforcement learning (reward mechanism) method to obtain certain characteristic parameters of the model (such as the number of parameters of the model, the actual delay of a certain performance indicator of the model, etc.) and certain parameter indicators of the robot (such as the maximum error). Under the reference of these parameters, according to Figure 3 The general rules in [1] can determine the specific compression methods and weights, thereby achieving dynamic compression of the model. This set of rules also deterministically reduces the difficulty of selecting different compression methods, helping engineers better design practical systems.

[0115] The embodiments of the present invention solve the problem of balancing real-time performance, accuracy and adaptability to complex scenarios of mobile robots under limited computing power through hardware status monitoring, task type segmentation, multi-strategy compression and dynamic feedback mechanism, and significantly improve the autonomous decision-making efficiency and system robustness of robots in intelligent manufacturing scenarios.

[0116] like Figure 5 As shown, an embodiment of the present invention further provides a dynamic compression control system for an AI model of a mobile robot, comprising a hardware status monitoring unit 10, a compression strategy selection unit 20, and a model dynamic adjustment unit 30. The system is used to obtain the hardware status data and task type of the robot, and to normalize and store the hardware status data, wherein the hardware status data includes at least one of CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate; the system is used to select a corresponding compression strategy in real time based on the hardware status data and / or task type, wherein the compression strategy includes at least one of compressing the AI ​​model based on the hardware status data, compressing the AI ​​model based on the hardware status data combined with the reward function and the task type, and compressing the AI ​​model based on the task type and / or the hardware status data; and the system is used to compress the AI ​​model, wherein compressing the AI ​​model includes at least one of pruning, quantizing, and skipping the AI ​​model.

[0117] The mobile robot AI model dynamic compression control system also includes an inference execution and performance feedback unit 40 and an exception handling and rollback unit 50. The inference execution and performance feedback unit 40 is used to run the lightweight model and extract model performance parameters, including accuracy, latency, and response time. The exception handling and rollback unit 50 rolls back the AI ​​model to its original configuration if the model performance parameters do not meet the standards.

[0118] By running the compressed model, performance parameters such as accuracy, delay time, and response time are extracted. Taking mobile navigation as an example, navigation accuracy parameters, navigation control operation processing delay time value, system response time, etc. are extracted.

[0119] In this specification, references to the same or similar parts between the various embodiments can be made to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the product embodiments described later, since they correspond to the methods, the description is relatively simple, and the relevant parts can be referred to the partial description of the system embodiment.

[0120] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A robot AI model dynamic compression method, applied to a robot deployed with an AI model, characterized in that: include: Obtaining the robot's hardware status data and task type, normalizing and storing the hardware status data, wherein the hardware status data includes CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate; Based on the hardware status data and / or the task type, a corresponding compression strategy is selected in real time to compress the AI ​​model to obtain a lightweight model. The compression strategy includes compressing the AI ​​model based on the hardware status data, compressing the AI ​​model based on the hardware status data combined with the reward function and the task type, and compressing the AI ​​model based on the task type and / or the hardware status data. Compressing the AI ​​model includes at least one of pruning, quantizing, and skipping the AI ​​model.

2. The robot AI model dynamic compression method according to claim 1, characterized in that: The compressing the AI ​​model based on the hardware status data includes: The maximum parameter index of the AI ​​model is set based on the hardware status data, and the maximum parameter index of the AI ​​model is adjusted according to different hardware status data. The maximum parameter index includes at least one of a maximum parameter amount and a maximum FPS.

3. The robot AI model dynamic compression method according to claim 2, characterized in that: Preset thresholds are set for the hardware status data, and the preset thresholds include a CPU utilization threshold, a CPU temperature threshold, a remaining memory value threshold, and a communication channel blocking rate threshold.

4. The robot AI model dynamic compression method according to claim 1, characterized in that: The hardware status data is combined with the reward function and the task type to compress the AI ​​model, including: Setting a reward function, wherein the reward function takes the model parameter amount of the AI ​​model, the maximum FPS of the model, the maximum error corresponding to the robot parameters, and the time delay of a certain performance of the AI ​​model as parameters, and setting weight factors for the model parameter amount, the maximum FPS of the model, the maximum error corresponding to the robot parameters, and the time delay of a certain performance of the AI ​​model respectively; After obtaining the model parameter amount and FPS parameter of the AI ​​model based on the reward function, the AI ​​model is compressed in combination with the task type.

5. The robot AI model dynamic compression method according to claim 4, characterized in that: Setting corresponding compression weights for pruning, quantization, and skipping according to different task types, and compressing the AI ​​model according to the compression weights; Wherein, when the task type is mobile navigation and obstacle avoidance, the compression weight level of pruning is set to low, the compression weight level of quantization is set to medium, and the skip weight is set to high; and / or When the task type is a terminal operation, the compression weight level of pruning is set to medium, the compression weight level of quantization is set to low, and the skip weight is set to none; and / or When the task type is scene perception, setting the compression weight level of pruning to high, the compression weight level of quantization to medium, and the skip weight to low; and / or When the task type is autonomous planning, the compression weight level of pruning is set to high, the compression weight level of quantization is set to high, and the skip weight is set to medium; and / or When the task type is sudden obstacle avoidance, the compression weight level of pruning is set to low, and the compression weight level and skip weight of quantization are both set to high; and / or When the task type is a high-precision operation, the compression weight level of pruning and the compression weight level of quantization are both set to medium, and the skip weight is set to low.

6. The robot AI model dynamic compression method according to claim 1, characterized in that: The compressing the AI ​​model based on the task type and / or the hardware status data includes: According to the working scene, task nature and complexity, the robot's task types are divided into navigation task type, obstacle avoidance behavior type, operation task type and data task type. The navigation task type includes long-distance sparse scene navigation, few-obstacle, static environment navigation, conventional navigation and dynamically changing scene navigation; the obstacle avoidance behavior type includes conventional automatic obstacle avoidance and sudden / emergency obstacle avoidance; the operation task type includes material grasping operation task, material transfer task, movement + operation collaborative task, high-speed beat movement operation, dynamic mutation operation, conventional precision operation task, multi-task and human-machine collaboration task; the data task includes battery power consumption data, dense data flow occupancy task, real-time perception data task and low-precision computing task; The AI ​​model is compressed according to the task type of the robot.

7. The robot AI model dynamic compression method according to claim 6, characterized in that: The CPU utilization, CPU temperature, remaining memory value and communication channel blocking rate are monitored, and when each parameter reaches a preset threshold, the AI ​​model is compressed.

8. The robot AI model dynamic compression method according to claim 6, characterized in that: When there are multiple areas in the same work scene, if the multiple areas are divided into multiple task types, the robot executes corresponding compression methods according to different task types in the multiple areas.

9. The robot AI model dynamic compression method according to claim 1, characterized in that: Also includes: Running the lightweight model and extracting model performance parameters, wherein the model performance parameters include accuracy, delay, and response time; When the model performance parameters do not meet the standards, the AI ​​model is rolled back to the original configuration state.

10. A dynamic compression control system for an AI model of a mobile robot, characterized in that: include: a hardware status monitoring unit, configured to obtain hardware status data and task type of the robot, normalize and store the hardware status data, wherein the hardware status data includes at least one of CPU utilization, CPU temperature, remaining memory value, and communication channel blocking rate; a compression strategy selection unit, configured to select a corresponding compression strategy in real time based on the hardware status data and / or the task type, wherein the compression strategy comprises at least one of compressing the AI ​​model based on the hardware status data, compressing the AI ​​model based on the hardware status data combined with a reward function and the task type, and compressing the AI ​​model based on the task type and / or the hardware status data; A model dynamic adjustment unit is used to compress the AI ​​model, where compressing the AI ​​model includes at least one of pruning, quantizing, and skipping the AI ​​model.

Images