Task adaptive dynamic computing power allocation method and system for heterogeneous robots

Abstract

This invention relates to the field of heterogeneous robot technology, specifically to a method and system for adaptive dynamic computing power allocation for heterogeneous robots. The method includes: collecting typical task information of heterogeneous robots; classifying tasks into different computing power requirement levels based on task complexity, real-time requirements, and data processing volume; extracting key information to construct a preset task computing power library; receiving robot task instructions, parsing the tasks, setting task real-time weights, collecting robot hardware status parameters in real time, and calculating the appropriate computing power; calculating the matching degree between the appropriate computing power and the computing power currently provided by the hardware; when the matching degree is lower than a preset threshold, performing structured channel pruning and fine-tuning the model through knowledge distillation; when the computing power requirement still cannot be met after pruning, identifying non-core tasks and initiating a cloud-based task allocation process. This technical solution can improve the task execution efficiency and hardware adaptability of heterogeneous robots.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous robot technology, and more specifically to a task-adaptive dynamic computing power allocation method and system for heterogeneous robots. Background Technology

[0002] Heterogeneous robots (robots equipped with processors, sensors, and actuators of different architectures) are widely used in complex scenarios such as industrial manufacturing, emergency rescue, and intelligent inspection. In these scenarios, robots need to handle multiple types of tasks simultaneously, including visual recognition, path planning, and decision control, and the operating environment is highly dynamic and uncertain.

[0003] In existing robot computing power management technologies, there are several mainstream modes: one is the fixed configuration mode, which is to preset the computing power allocation scheme according to the robot's hardware model at the factory, and the proportion of computing power resources occupied by each task remains unchanged during operation; the other is a single-dimensional scheduling strategy, such as making a rough allocation of computing power based only on the task type (such as "navigation" or "grasping"), or making a static allocation of incremental computing power based only on the hardware's access status (such as whether it is connected to a Type-C expansion dock or network cable).

[0004] However, as the complexity of robotic tasks increases and the operating environment becomes more varied, existing technologies are gradually revealing certain limitations: First, existing scheduling algorithms are mostly designed based on ideal hardware conditions. For example, when the robot's motherboard temperature rises and triggers frequency reduction protection, or when the battery power drops and power supply becomes insufficient, the actual computing power capacity of the hardware will be significantly reduced. If high-load tasks are still assigned according to a fixed scheme, it is very easy to cause system lag, task execution failure, or even hardware damage; conversely, if low-computing-power tasks are assigned when the hardware is in a high-performance state, it will result in a huge waste of resources.

[0005] While reference document CN114700957B proposes to reduce local computing power requirements through collaborative allocation of computing power among local area network devices, this solution only focuses on the sharing of computing power between devices and fails to accurately match the real-time weight of specific tasks with the instantaneous load state of the hardware, resulting in insufficient adaptability in heterogeneous robot mixed operation scenarios. Similarly, although reference document CN120630679A uses model compression strategies such as pruning and quantization to optimize performance, its compression process lacks a design that is related to the current task requirements. This can easily lead to a decrease in the accuracy of critical tasks due to over-compression, or an inability to meet real-time requirements due to insufficient compression.

[0006] Furthermore, existing technologies typically lack effective solutions when local computing power pools cannot meet sudden surges in computing power demand. While some solutions mention utilizing the cloud or external computing centers, they generally lack mechanisms for identifying and categorizing "core tasks" and "non-core tasks." Once local computing power bottlenecks occur, it is impossible to tap into potential through real-time adjustments to the model structure (such as pruning and distillation), nor can non-core tasks be seamlessly offloaded to the cloud or nearby robots intelligently, leading to the failure of core tasks due to resource congestion.

[0007] Existing AI models on robots are mostly deployed with a fixed structure, meaning that the model complexity remains unchanged after deployment. This implies that when there is an imbalance between computing power supply and demand, the system cannot maintain task continuity by dynamically adjusting model complexity (such as finding a balance between accuracy and speed). This static resource management approach results in low overall system energy efficiency, slow response, and difficulty in achieving globally optimal allocation of computing resources in heterogeneous robot clusters. Summary of the Invention

[0008] The purpose of this invention is to propose a task-adaptive dynamic computing power allocation method and system for heterogeneous robots. This technical solution can improve the task execution efficiency and hardware adaptability of heterogeneous robots.

[0009] To achieve the above objectives, in a first aspect, the present invention proposes a task-adaptive dynamic computing power allocation method for heterogeneous robots, comprising: Collect typical task information of heterogeneous robots, classify tasks into different computing power requirement levels according to task complexity, real-time requirements and data processing volume, and extract key information to build a preset task computing power library. Receive robot task instructions, parse the task, set the task real-time weight, collect the robot's hardware status parameters in real time, and calculate the appropriate computing power. Calculate the degree of matching between the computing power of the target system and the computing power that the current hardware can provide; When the matching degree is lower than the preset threshold, structured channel pruning is performed, and the model is fine-tuned through knowledge distillation; If the computing power demand still cannot be met after pruning, identify non-core tasks and initiate the cloud-based task offloading process.

[0010] Beneficial effects of the basic solution: By collecting typical task information of heterogeneous robots, classifying computing power demand levels according to task complexity, real-time requirements, and data processing volume, and building a preset task computing power library, the solution achieves standardized and quantitative management of task computing power demand, changes the traditional fixed and extensive mode of computing power allocation, and provides a data foundation for subsequent precise matching.

[0011] By combining task analysis, real-time weight setting, and robot hardware status acquisition, a three-dimensional matching model of task type, hardware status, and computing power requirements is constructed. This model can dynamically calculate the appropriate computing power and evaluate the matching degree, achieving precise matching between computing power requirements and robot hardware computing power, thus solving the problem of poor adaptability in traditional solutions.

[0012] When computing power is insufficient, model fine-tuning is performed through structured channel pruning and knowledge distillation. This reduces the model's computing power usage without significantly sacrificing core task performance, maximizes the use of the robot's local hardware computing power, reduces unnecessary cloud dependencies, and improves computing power utilization.

[0013] When the local model optimization still cannot meet the computing power requirements, the non-core tasks are accurately identified and cloud-based traffic distribution is initiated to prioritize the real-time performance and execution effect of core tasks, avoiding task lag and failure due to insufficient computing power, and significantly improving the completion rate of complex tasks.

[0014] By coordinating local computing power optimization and cloud task offloading, long-term high-load operation of hardware is avoided, effectively reducing energy consumption and hardware wear and tear. While improving task execution efficiency, it significantly improves the endurance performance and operational stability of heterogeneous robots.

[0015] This solution is compatible with heterogeneous robots with different computing power and hardware configurations, and can be flexibly adapted to various types of robots without the need for customized modifications for a single model. It has good versatility and scenario expansion capabilities.

[0016] As a feasible and preferred solution, the key information extracted includes task type, core computing power requirements, task priority, and associated AI model type.

[0017] As a feasible and preferred solution, the task computing power library supports offline batch updates and online real-time updates. Offline batch updates retrain the computing power demand prediction model based on historical task data of the robot cluster and correct the computing power threshold. Online real-time updates collect computing power consumption data and add it to the computing power library after review and confirmation when executing new task types.

[0018] As a feasible and preferred solution, task parsing includes retrieving baseline computing power requirements, task priorities, and associated AI model information from a pre-defined task computing power library; Hardware status parameters include motherboard temperature, battery level, processor load rate, and memory usage, and each parameter is standardized.

[0019] As a feasible and preferred solution, a weighted matching model is constructed to calculate the appropriate computing power F. adapt The formula is:

[0020] in, This is a hardware status correction factor.

[0021] As a feasible and preferred solution, the matching degree calculation formula is as follows:

[0022] in, To adapt to computing power, If the matching degree is greater than or equal to the first preset threshold, the task is executed directly based on the computing power provided by the current hardware; if the matching degree is less than the second preset threshold, the computing power optimization process is triggered.

[0023] As a feasible and preferred solution, when the matching degree is lower than a preset threshold, the AI ​​model pruning optimization process is initiated, with different degrees of pruning initiated when the matching degree is in different ranges; when the adapted computing power is less than the computing power that the current hardware can provide and the matching degree is between 60% and 85%, light pruning is initiated, with a pruning ratio of 10% to 30%; when the matching degree is less than 60%, deep pruning is initiated, with a pruning ratio of 30% to 50%.

[0024] As a feasible and preferred approach, structured pruning includes: calculating importance scores based on the L1 norm of the absolute values ​​of the weights of each layer of the model; calculating importance scores based on the weights of each layer of the model; retaining a predetermined proportion of channels before scoring; and prioritizing the retention of convolutional layers and fully connected layers that are related to the core features of the task. After pruning, the model is fine-tuned through knowledge distillation.

[0025] As a feasible and preferred solution, non-core tasks are determined by task priority, real-time weight, and data correlation; an encrypted communication link is established to compress non-core task data and transmit it to the cloud server for processing, and the cloud processing results are returned through an asynchronous feedback mechanism; network latency is monitored in real time, and when the latency exceeds a preset threshold, the traffic splitting is paused and the task is temporarily stored locally, while core tasks are always kept locally for processing.

[0026] Secondly, the present invention also provides a task adaptive dynamic computing power allocation system for heterogeneous robots, which utilizes the above-mentioned task adaptive dynamic computing power allocation method for heterogeneous robots, including a preset task computing power library, a three-dimensional matching algorithm module, an AI model dynamic pruning optimization module, and a task intelligent diversion module. Attached Figure Description

[0027] Figure 1 This is a logical diagram of a task-adaptive dynamic computing power allocation method for heterogeneous robots. Detailed Implementation

[0028] To make the technical solution and advantages of this application clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only some embodiments of the present invention, and are only used to explain this application, not to limit it. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated; they can be combined with each other to achieve better technical effects. The same reference numerals appearing in the accompanying drawings of the following embodiments represent the same features or components, and can be applied to different embodiments.

[0029] Furthermore, unless otherwise defined, the technical or scientific terms used in this invention description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.

[0030] The present invention will now be described in further detail with reference to the accompanying drawings.

[0031] A variety of heterogeneous robots were selected as experimental subjects, including but not limited to wheeled robots, humanoid robots, and quadruped robots. Each robot was equipped with processors and sensors with different performance (such as lidar, cameras, IMU, etc.).

[0032] Deploy ROS (Robot Operating System) as the operating system on the robot, and install the necessary drivers and middleware. Deploy the task adaptive dynamic computing power allocation system of this invention on the external brain (which can be a high-performance server or a cloud platform), including a preset task computing power library, a three-dimensional matching algorithm module, an AI model dynamic pruning and optimization module, and a task intelligent distribution module.

[0033] Ensure a stable network connection between the robot and the external brain, which can be a wired network (such as Ethernet) or a wireless network (such as Wi-Fi, 4G / 5G, etc.).

[0034] This disclosure provides a task-adaptive dynamic computing power allocation method for heterogeneous robots, referring to... Figure 1 ,include: Step S100, Preset task computing power library construction, including: Step S101, Data Collection and Classification: Based on a large amount of experimental data and actual application scenarios, collect relevant information on typical tasks of various heterogeneous robots, including task complexity (high / medium / low), real-time requirements (milliseconds / seconds / minutes), data processing volume (GB / MB / KB) and the proportion of computationally intensive operations.

[0035] Based on the above characteristics, tasks are divided into different levels of computing power requirements: Tasks with high complexity, high real-time requirements, and large data volume are classified as high computing power requirement levels, such as path planning and 3D environmental modeling. Tasks with medium complexity, medium real-time requirements, and medium data volume are classified as medium computing power requirements, such as target tracking and routine inspection. Tasks with low complexity, low real-time requirements, and small data volume are classified as low computing power requirement levels, such as simple attitude control and signal feedback.

[0036] Step S102: Extract key information. For each task, extract its key information, including task type, core computing power requirement indicators, task priority, and associated AI model type.

[0037] For example, tasks with high computing power requirements include: robot dog climbing hills (core requirements are path planning and power control, associated with a path planning AI model, computing power requirement threshold ≥ 80 GFLOPS); tasks with medium computing power requirements include: inspection robot environmental monitoring (core requirements are image recognition, associated with a lightweight CNN model, computing power requirement threshold 30-60 GFLOPS); and tasks with low computing power requirements include: humanoid robot handshake (core requirements are posture control and force feedback, associated with a posture adjustment AI model, computing power requirement threshold ≤ 20 GFLOPS).

[0038] Step S103, Database Construction and Update: Construct a preset task computing power library and store the above key information in the database.

[0039] The preset task computing power library supports a dual-mode dynamic update mechanism: Offline batch update: Based on historical task data of the robot cluster (task completion rate, computing power consumption deviation, model accuracy change), the computing power demand prediction model is retrained weekly using the gradient boosting tree algorithm to correct the computing power threshold of each task. Online real-time updates: When the robot performs a new task type, a real-time learning process is triggered to collect computing power consumption data during task execution. After manual review and confirmation, the data is added to the computing power database. The updated data must pass a consistency check (computing power deviation from similar tasks ≤10%) before it becomes effective.

[0040] Step S200, establish a matching algorithm, including: Step S201, Task Analysis: After receiving the robot's task instructions, the external brain analyzes the task type and retrieves the corresponding baseline computing power requirement (F) from the preset task computing power library. base The system sets task priorities (P, core tasks P=1.0, non-core tasks P=0.3-0.7) and associated AI model information, while also setting task real-time weights (W, real-time tasks W=0.6, non-real-time tasks W=0.2).

[0041] Step S202, hardware status monitoring: Through the sensors integrated on the robot motherboard and the monitoring module of the external brain, hardware status parameters are collected in real time, including motherboard temperature (T, monitoring accuracy ±0.5℃), battery power (B, accuracy ±1%), CPU / GPU load rate (L, sampling frequency 10Hz), and memory usage (M), and each parameter is standardized to the [0,1] range.

[0042] Step S203: Dynamically assess computing power requirements and construct a weighted matching model to calculate the appropriate computing power F. adapt The formula is:

[0043] in, The matching threshold rules are set as follows for the hardware status correction coefficient: When T>75℃ or B<15%, K=0.5-0.7 (dynamically adjusted according to the degree of parameter deviation); When 60℃≤T≤75℃ and 15%≤B≤30%, K=0.8-0.9; When T < 60℃ and B > 30%, K = 1.0 - 1.1.

[0044] Step S204, Matching degree judgment and feedback: Based on the matching degree result, decide whether to trigger the computing power optimization process or directly execute the task. The calculation formula is as follows:

[0045] If the matching degree is greater than or equal to the first threshold (in this embodiment, the first threshold is 85%), the task is executed directly; if the matching degree is less than the second threshold (in this embodiment, the second threshold is 60%), the computing power optimization process is triggered; the matching result is fed back to the external brain control module to provide a decision basis for subsequent AI model compression or task diversion.

[0046] Step S300, Dynamic pruning and optimization of the model, using a structured channel pruning algorithm to achieve dynamic compression of the AI ​​model, including; Step S301: Determine the pruning trigger condition and, based on the matching degree result, determine whether the AI ​​model pruning optimization process needs to be initiated. Specifically, when... < When the matching degree is 60%-85%, start light pruning (pruning ratio 10%-30%); when the matching degree is <60%, start deep pruning (pruning ratio 30%-50%).

[0047] Step S302: Adjust the pruning ratio and calculate the importance score based on the L1 norm of the absolute value of the weights of each layer in the model, retaining the channels before the score (1 - pruning ratio). For example, when pruning a 100-layer neural network to 50 layers, prioritize retaining convolutional layers and fully connected layers that are relevant to the core features of the task.

[0048] Step S303, accuracy assurance mechanism: After pruning, the model is fine-tuned through knowledge distillation to ensure that the model accuracy decreases by ≤5%. (Experimental verification: After pruning the robot dog path planning model by 40%, the path planning error increased from 0.12m to 0.15m, but still meets practical requirements). The pruned model parameters are stored in local cache in real time and can be directly called in the same task scenario.

[0049] Step S400, the intelligent task routing mechanism, when deep pruning still cannot meet the F requirement (F < 0.8 × F), initiates the cloud-based routing process, including: Step S401, Non-core task identification: Non-core tasks are identified based on three conditions: task priority P≤0.5, real-time weight W≤0.3, and data correlation (data interaction frequency with core tasks <5 times / minute). Examples of non-core tasks include environmental data storage and historical task log statistics.

[0050] Step S402: Establish an encrypted communication link (using the TLS1.3 protocol), compress non-core task data and transmit it to the cloud server, and send back the cloud processing results through an asynchronous feedback mechanism.

[0051] Step S403: If the task is non-core and local computing power is insufficient, initiate the cloud-based task splitting process; otherwise, execute the task locally. During the splitting process, network latency is monitored in real time, and the task execution location is switched based on the latency. Specifically, network latency is monitored in real time (threshold set to 100ms). When the latency > 100ms, the splitting is paused, and non-core tasks are temporarily stored locally until the network recovers and then re-uploaded. Core tasks (such as obstacle avoidance and power control) are always processed locally to ensure a response latency ≤ 20ms.

[0052] This technical solution proposes a global reward model for optimizing computing power utilization. It primarily includes rewards for GPU utilization, throughput, and latency ratio. For core task assurance, it uses priority weighting and SLA breach penalties as constraints. Additionally, it rewards computing power savings through optimization of computation length and dynamic computation. Finally, it optimizes core tasks and cost-sensitive aspects through multi-objective comprehensive rewards, achieving optimized computing power scheduling in task modes. This model permeates the entire computing power scheduling process. Through reinforcement learning and optimization using multi-objective reward functions, the robot's brain can dynamically calculate, schedule, and prioritize based on utilization, throughput, computing power cost weighting, and core task priority processing, ensuring elastic resource allocation. Ultimately, this achieves the goal of enabling the same computing power to handle increasingly more tasks, with more stable core tasks and gradually lower costs.

[0053] The above content is merely an embodiment of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can improve and implement this solution based on the guidance provided in this application and their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A task-adaptive dynamic computing power allocation method for heterogeneous robots, characterized in that, include: Collect typical task information of heterogeneous robots, classify tasks into different computing power requirement levels according to task complexity, real-time requirements and data processing volume, and extract key information to build a preset task computing power library. Receive robot task instructions, parse the task, set the task real-time weight, collect the robot's hardware status parameters in real time, and calculate the appropriate computing power. Calculate the degree of matching between the computing power of the target system and the computing power that the current hardware can provide; When the matching degree is lower than the preset threshold, structured channel pruning is performed, and the model is fine-tuned through knowledge distillation; If the computing power demand still cannot be met after pruning, identify non-core tasks and initiate the cloud-based task offloading process.

2. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 1, characterized in that, The key information extracted includes task type, core computing power requirements, task priority, and associated AI model type.

3. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 1, characterized in that, The task computing power library supports offline batch updates and online real-time updates. Offline batch updates retrain the computing power demand prediction model based on historical task data of the robot cluster and correct the computing power threshold. Online real-time updates collect computing power consumption data when executing new task types, and then supplement the computing power library after review and confirmation.

4. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 1, characterized in that, Task analysis includes retrieving baseline computing power requirements, task priorities, and associated AI model information from a pre-defined task computing power library; Hardware status parameters include motherboard temperature, battery level, processor load rate, and memory usage, and each parameter is standardized.

5. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 4, characterized in that, Construct a weighted matching model to calculate the appropriate computing power F adapt The formula is: in, This is a hardware status correction factor.

6. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 5, characterized in that, The formula for calculating the matching degree is as follows: in, To adapt to computing power, If the matching degree is greater than or equal to the first preset threshold, the task is executed directly based on the computing power provided by the current hardware; if the matching degree is less than the second preset threshold, the computing power optimization process is triggered.

7. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 1, characterized in that, When the matching degree is lower than the preset threshold, the AI ​​model pruning optimization process is initiated, with different degrees of pruning initiated when the matching degree is in different ranges; when the adapted computing power is less than the computing power that the current hardware can provide and the matching degree is between 60% and 85%, light pruning is initiated, with a pruning ratio of 10% to 30%; when the matching degree is less than 60%, deep pruning is initiated, with a pruning ratio of 30% to 50%.

8. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 1, characterized in that, The structured pruning process includes: calculating importance scores based on the L1 norm of the absolute values ​​of the weights of each layer of the model; retaining a predetermined proportion of channels before scoring; and prioritizing the retention of convolutional and fully connected layers that are relevant to the core features of the task. After pruning, the model is fine-tuned through knowledge distillation.

9. The task-adaptive dynamic computing power allocation method for heterogeneous robots according to claim 1, characterized in that, Non-core tasks are determined by task priority, real-time weight, and data correlation; an encrypted communication link is established, and non-core task data is compressed and transmitted to the cloud server for processing. The cloud processing results are returned through an asynchronous feedback mechanism. The system monitors network latency in real time. When the latency exceeds a preset threshold, it pauses traffic splitting and temporarily stores tasks locally, while core tasks are always processed locally.

10. A task-adaptive dynamic computing power allocation system for heterogeneous robots, characterized in that, The method for adaptive dynamic computing power allocation for heterogeneous robots as described in any one of claims 1-9 is employed, including a preset task computing power library, a three-dimensional matching algorithm module, an AI model dynamic pruning and optimization module, and a task intelligent diversion module.

Images