Body intelligent robot skill runtime data analysis and scheduling method

Abstract

The present application relates to embodied intelligent robot technical field, specifically to embodied intelligent robot skill runtime data analysis and scheduling method. The method comprises: a skill data receiving module receives skill data packets of external data sources and completes format verification and integrity check; a multi-level semantic analysis engine carries out three-level progressive analysis of syntax layer, semantic layer and intention layer on the skill data packets, and generates executable action sequences; a real-time scheduling controller combines the current state of the robot and environmental information to prioritize and time slice allocate the action sequences, and generates a scheduling scheme; an adaptive optimization module continuously iterates scheduling strategy parameters based on historical execution data and real-time feedback. Through the technical framework of multi-level semantic analysis and dynamic scheduling cooperation, the present application solves the problems of information loss, response delay and insufficient adaptability in skill transfer process, and significantly improves the precision and task completion efficiency of robot skill execution in complex dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of embodied intelligent robot technology, and specifically relates to a method for analyzing and scheduling data during the operation of embodied intelligent robot skills. Background Technology

[0002] Embodied intelligence, as a core direction for the application of artificial intelligence, is driving the evolution of robotics from simple perception to a fully intelligent "perception-cognition-decision-execution" chain. Currently, the mainstream path for robot skill learning involves human demonstration data collection combined with imitation learning, but traditional skill data processing methods have systemic limitations: Significant information loss: Existing digital records can only capture surface features such as the pose changes of the end effector, losing implicit information such as fine-tuning of force, timing control, situation judgment, and real-time response to emergencies. For example, in industrial precision assembly scenarios, the gradient changes in the fingertip force when a worker tightens a screw cannot be captured by traditional systems, resulting in a robot's ability to reproduce assembly actions with a yield rate of only about 92%.

[0003] Excessive execution latency: After receiving skill data, preprocessing such as format conversion, coordinate transformation, and motion planning must be completed sequentially. The latency of a single step can reach tens of milliseconds, and the end-to-end latency generally exceeds 100ms. In real-time scenarios such as medical surgery assistance and hazardous environment operations, such latency can easily lead to task failure or even safety accidents.

[0004] Insufficient adaptability: Traditional systems employ static scheduling strategies, strictly adhering to preset action sequences and timings. When environmental changes occur, such as load fluctuations, obstacle intrusions, or target displacement, the strategy cannot be adjusted in real time. Some systems incorporating online learning have extremely high computational complexity, making it difficult to meet real-time requirements. In dynamic environments, task completion rates can fluctuate by more than 40%.

[0005] Inefficient resource utilization: Complex tasks contain multiple sub-tasks, each with significantly different priorities, time requirements, and resource needs. However, traditional systems lack a unified resource scheduling framework, often resulting in resource congestion for some tasks and idle resources for others. In multi-task concurrent scenarios, system resource utilization is generally below 60%.

[0006] Core defects of existing technology

[0007] To address the aforementioned pain points, the industry has proposed three types of improvement solutions, but significant shortcomings still exist: Summary of the Invention

[0008] In view of this, the present invention provides a method for analyzing and scheduling runtime data of embodied intelligent robot skills, in order to solve or alleviate one of the technical problems existing in the prior art, and at least provide a beneficial option.

[0009] The technical solution of this invention is implemented as follows: a method for analyzing and scheduling runtime data of embodied intelligent robot skills. Includes the following steps: Step S1: Receive skill data packets from an external data source through the skill data receiving module. Perform format compliance verification and data integrity verification on the skill data packets in sequence. The format compliance verification verifies whether the data packets conform to the preset skill data structure standard. The data integrity verification verifies that the data packets are not lost or damaged through a hash algorithm. Step S2 involves performing a three-level progressive semantic analysis on the verified skill data package using a multi-level semantic analysis engine. The three-level progressive semantic analysis includes: syntax layer analysis, which extracts action timing information, spatial trajectory, and data frame correlation; semantic layer analysis, which extracts action force range, speed threshold, acceleration limit, and obstacle avoidance constraints; and intent layer analysis, which extracts task objectives, success criteria, emergency switching schemes, and related action dependencies, ultimately generating a structured executable action sequence. Step S3: The robot's current joint angle, end pose, load status, and obstacle distribution, temperature, humidity, and lighting parameters in the operating environment are collected by the real-time scheduling controller. The priority is sorted and time slices are allocated in combination with the action sequence to generate an executable scheduling scheme. The priority sorting is determined by the urgency of the task, the deadline, resource requirements, and the benefit weight. Step S4: Construct a multi-dimensional priority evaluation system through the priority management module. The multi-dimensional priority evaluation system includes task urgency index, resource competition index, task relevance index, and benefit weight index. Each index is weighted and summed through adjustable weight coefficients for the scenario to obtain a comprehensive priority score. When the task status changes significantly, the scheduling queue is automatically recalculated and updated. Step S5: The scheduling problem is modeled as a Markov decision process through the adaptive optimization module. The state space is the robot state, environmental parameters, and task queue, the action space is the scheduling decision, and the reward function is the weighted combination of task completion time, resource utilization, energy consumption, and safety. The scheduling strategy parameters are updated using an online reinforcement learning algorithm. At the same time, the historical scheduling decisions and execution results are stored through an experience playback mechanism, and the model is updated in batches to avoid catastrophic forgetting. Step S6: The scheduling scheme is converted into robot joint control commands by the execution control layer. Commands are issued every 10ms and the execution process is monitored in real time. If an abnormal state is detected, an emergency plan is triggered.

[0010] Furthermore, the multi-level semantic parsing engine in step S2 adopts an action sequence modeling method based on a self-attention Transformer network. The self-attention Transformer network contains 8 attention heads, the input is the temporal feature vector of skill data, and the output is an action graph with attention weights. The attention weights quantify the correlation strength between any two actions, and the correlation strength ranges from 0 to 1. The higher the value, the greater the impact of the action on the current task.

[0011] Furthermore, the multi-level semantic parsing engine in step S2 also implements a dynamic semantic completion mechanism: when there are missing fields or semantic ambiguity in the skill data, the action features of three adjacent time windows in the context are first matched, and then the typical parameter range of the same skill is retrieved from the domain knowledge base to generate at least two candidate completion schemes. The optimal scheme is selected to fill the missing field by confidence scoring. The domain knowledge base contains 1,200 standard skill parameters in three major fields: industrial assembly, medical surgery, and home care.

[0012] Furthermore, the multi-level semantic parsing engine in step S2 supports the fusion parsing of multimodal skill data: it simultaneously receives RGB-D images collected by a visual sensor, contact pressure distribution collected by a tactile sensor, and six-dimensional force data collected by a force sensor. Through feature-level fusion, the three types of modal data are mapped to the same feature space, and cross-modal correlation features are extracted. These cross-modal correlation features are used to correct the errors of single-modal parsing and improve the accuracy of skill representation.

[0013] Furthermore, the real-time scheduling controller in step S3 adopts a two-stage scheduling architecture with prediction-execution separation: in the prediction stage, the LSTM model is used to predict the task arrival rate, CPU / memory / communication bandwidth resource availability, and environmental change trends within the next 5 time windows, generating 3 candidate scheduling schemes; in the execution stage, the robot state and environmental parameters are combined with real-time feedback, and the scheme with the highest comprehensive score is selected from the 3 candidate schemes for execution. The comprehensive score is calculated by the task completion probability, resource consumption, and safety risk.

[0014] Furthermore, the real-time scheduling controller in step S3 adopts a time-slice-based soft real-time scheduling algorithm: the processor time is divided into fixed time slices of 10ms in length, and a maximum of 8 action tasks are scheduled in parallel within each time slice. The time slice allocation ratio is dynamically adjusted according to the deadline requirements of the tasks. Tasks with a deadline of less than 100ms are allocated no less than 60% of the time slice resources, while reserving 10% of the time slices as an emergency buffer to ensure that high-priority tasks are completed first.

[0015] Furthermore, the real-time scheduling controller in step S3 supports distributed collaborative scheduling: when there are multiple robots or multiple processor cores in the system, each node exchanges task status and resource information every 50ms through the message passing interface, and uses the distributed alternating direction multiplier method to calculate the globally optimal scheduling scheme, thereby achieving cross-node task load balancing and resource collaboration. When a single node fails, the task is automatically migrated to other available nodes.

[0016] Furthermore, the priority management module in step S4 also introduces a fairness constraint mechanism: the maximum waiting time threshold for a task is set to 300s. When the waiting time of a task increases by 10s, its comprehensive priority score is automatically increased by 5% to prevent low-priority tasks from being unable to execute due to continuous preemption; at the same time, a maximum priority limit is set to prevent a single task from excessively occupying resources and causing all other tasks to be blocked.

[0017] Furthermore, the adaptive optimization module in step S5 implements Pareto front tracking for multi-objective optimization: simultaneously optimizing four objectives—shortest task completion time, highest resource utilization, lowest energy consumption, and highest safety—and using the NSGA-II multi-objective evolutionary algorithm to solve for the Pareto optimal solution set. The Pareto optimal solution set contains at least 20 non-dominated solutions, and users can select the corresponding solution as the current scheduling strategy according to their preferences for specific application scenarios. The preference parameters can be adjusted through a visual interface.

[0018] Furthermore, it includes: a skill data receiving module, configured to support four access protocols: TCP / IP, UDP, MQTT, and HTTP, to complete the reception and verification of skill data packets; A multi-level semantic parsing engine is configured to perform three-level progressive semantic parsing, action sequence modeling, dynamic semantic completion, and multi-modal fusion parsing. The real-time scheduling controller is configured to perform predictive-execution separation scheduling, time slice allocation, and distributed collaborative scheduling; the priority management module is configured to maintain a multi-dimensional priority evaluation system and a fairness constraint mechanism. The adaptive optimization module is configured to perform online reinforcement learning, experience replay, and multi-objective Pareto optimization. The execution control layer is configured to convert scheduling schemes into control commands and monitor the execution process; the modules are connected through a high-speed PCIe data channel with a data transmission bandwidth of no less than 10Gbps, realizing a closed-loop processing of skill data from reception, parsing, scheduling to execution.

[0019] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: (1) The three-level semantic parsing fully preserves the implicit information of the skills. After testing in the industrial precision assembly scenario, the skill trajectory restoration error was reduced from ±0.15mm in the traditional method to ±0.03mm, and the parsing accuracy was improved by 42.3%, which can meet the needs of ultra-high precision scenarios such as medical surgery and chip packaging.

[0020] (2) The prediction-execution separation architecture decouples preprocessing and scheduling, reducing the end-to-end response latency from 120ms in the traditional method to 28ms, a reduction of 76.7%, which is suitable for strong real-time scenarios such as dynamic target tracking and emergency response in dangerous environments.

[0021] (3) The adaptive optimization mechanism enables the system to maintain a task completion rate of over 95% and a fluctuation range of less than 7.2% under scenarios of load fluctuation and environmental changes, which is far superior to the fluctuation level of over 35% of existing technologies.

[0022] (4) The multi-objective scheduling algorithm realizes global optimization of resource allocation. In the multi-task concurrent scenario, the CPU utilization rate increases from 58% to 89%, the memory usage rate increases from 62% to 85%, and the average task completion time is shortened by 26.8%.

[0023] (5) The unified data interface standard supports access to five mainstream skill data formats, increasing the skill reuse rate from 23% to 78% in the traditional method and reducing system integration costs by 65%, laying the foundation for the construction of an open skill ecosystem.

[0024] (6) Deployment efficiency is greatly improved: the scheduling strategy migration capability shortens the adaptation time for new scenarios from an average of 14 days to 6 days, improving deployment efficiency by 57.1% and significantly reducing application promotion costs.

[0025] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the overall method of the present invention; Figure 2 This is a flowchart of the dynamic semantic completion process of the present invention; Figure 3This is a flowchart of the multimodal skill data fusion and analysis process of the present invention. Detailed Implementation

[0028] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] Example 1: Precision assembly scenario with industrial robots

[0031] Scene Overview

[0032] It is applied to the precision assembly of crankshafts and cylinder blocks in automobile engines, with an assembly tolerance of ±0.05mm. The first-pass yield of traditional methods is only 94.2%, and the assembly cycle time is 120 seconds per piece.

[0033] Implementation process

[0034] Data Acquisition: Assembly demonstration data of 10 skilled workers were collected simultaneously using a 6-dimensional force sensor, industrial camera, and joint encoder. The size of a single skill data package is 2.3GB, containing 1200 data frames, covering three modalities: arm movement trajectory, fingertip force changes, and visual alignment features.

[0035] Data reception and verification: Skill data is transmitted to the system via TCP / IP protocol. The receiving module performs format verification (compliant with ISO 9283 robot performance standard) and SHA-256 integrity verification, with the packet loss rate controlled within 0.01%.

[0036] Multi-level semantic parsing:

[0037] Scheduling and execution: The real-time scheduling controller generates a scheduling plan by combining the robot's current joint angle (deviation ≤0.01°) and cylinder positioning error (≤0.02mm). The prediction phase completes analysis and preparation three time windows in advance, and the scheduling instructions are updated every 10ms during the execution phase. During assembly, the pressing force is monitored in real time; if it exceeds the threshold, the feed speed is automatically adjusted.

[0038] Optimization and iteration: The adaptive optimization module collects execution data from 200 assembly operations, updates scheduling strategy parameters, and gradually optimizes the force control accuracy of the pressing stage from ±1.2N to ±0.3N.

[0039] Implementation effect

[0040] Example 2: Medical robot-assisted laparoscopic surgery scenario

[0041] Scene Overview

[0042] The auxiliary operation used in laparoscopic cholecystectomy should support the accurate reproduction of the doctor's remote demonstration skills, the maximum allowable response delay for the surgical operation should be 50ms, and the incidence of safety events should be controlled below 0.1%.

[0043] Implementation process

[0044] Data Acquisition: Surgical experts completed 50 simulated surgical demonstrations using a force feedback remote control console, collecting three types of multimodal data: surgical instrument movement trajectory (sampling frequency 1kHz), operating force (resolution 0.01N), and surgical field image (4K@60fps). The size of a single skill data package is 18GB.

[0045] Data reception and verification: Data is transmitted via low-latency UDP protocol. The receiving module performs frame sequence number verification and data integrity verification. The transmission delay is ≤8ms and the packet loss rate is ≤0.005%.

[0046] Multi-level semantic parsing:

[0047] Scheduling and Execution: The real-time scheduling controller adopts a prediction-execution separation architecture. During the prediction phase, resource requirements for the next 30 seconds are updated every 5 seconds, and during the execution phase, control commands are issued every 10ms. The priority management module sets hemostasis and tissue protection actions as the highest priority, with a weighting coefficient of 0.8, which is much higher than the 0.3 of normal operations.

[0048] Optimization and iteration: The adaptive optimization module collects execution data from 100 surgeries and updates the scheduling strategy through experience replay, optimizing the response latency for abnormal situations from 35ms to 22ms.

[0049] Implementation effect

[0050] Example 3: Service Robot Home Care Scenario

[0051] Scene Overview

[0052] It is applied to two types of nursing tasks for disabled elderly people: turning over assistance and feeding assistance. It needs to be adapted to the differences in weight, body position and tolerance of different elderly people. The success rate of nursing actions needs to be ≥95% and the user satisfaction score needs to be ≥90 points.

[0053] Implementation process

[0054] Data collection: Nursing experts demonstrated nursing actions to 30 elderly people through remote teaching, collecting three types of data: robot joint angles, contact pressure, and images of the elderly people's body positions. The size of a single skill data package is 3.2GB.

[0055] Data reception and verification: Access to the home LAN via the MQTT IoT protocol. The receiving module performs JSON format verification and CRC32 integrity verification, with a data transmission delay of ≤15ms.

[0056] Multi-level semantic parsing:

[0057] Scheduling and execution: The real-time scheduling controller generates a scheduling plan by combining the elderly person's current position (identified by depth camera), the robot's remaining battery power (≥30%), and the distribution of obstacles in the room (detected by LiDAR). The priority is updated every 2 seconds: if the elderly person moves, the priority of the turning task is automatically increased by 20%.

[0058] Optimization and iteration: The adaptive optimization module collects execution data from 500 nursing tasks and optimizes the balance between movement comfort and energy consumption through multi-objective optimization, reducing the force fluctuation of the turning movement from ±3N to ±0.8N.

[0059] Implementation effect

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for analyzing and scheduling runtime data of embodied intelligent robot skills, characterized in that: Includes the following steps: Step S1: Receive skill data packets from an external data source through the skill data receiving module. Perform format compliance verification and data integrity verification on the skill data packets in sequence. The format compliance verification verifies whether the data packets conform to the preset skill data structure standard. The data integrity verification verifies that the data packets are not lost or damaged through a hash algorithm. Step S2 involves performing a three-level progressive semantic analysis on the verified skill data package using a multi-level semantic analysis engine. The three-level progressive semantic analysis includes: syntax layer analysis, which extracts action timing information, spatial trajectory, and data frame correlation; semantic layer analysis, which extracts action force range, speed threshold, acceleration limit, and obstacle avoidance constraints; and intent layer analysis, which extracts task objectives, success criteria, emergency switching schemes, and related action dependencies, ultimately generating a structured executable action sequence. Step S3: The robot's current joint angle, end pose, load status, and obstacle distribution, temperature, humidity, and lighting parameters in the operating environment are collected by the real-time scheduling controller. The priority is sorted and time slices are allocated in combination with the action sequence to generate an executable scheduling scheme. The priority sorting is determined by the urgency of the task, the deadline, resource requirements, and the benefit weight. Step S4: Construct a multi-dimensional priority evaluation system through the priority management module. The multi-dimensional priority evaluation system includes task urgency index, resource competition index, task relevance index, and benefit weight index. Each index is weighted and summed through adjustable weight coefficients for the scenario to obtain a comprehensive priority score. When the task status changes significantly, the scheduling queue is automatically recalculated and updated. Step S5: The scheduling problem is modeled as a Markov decision process through the adaptive optimization module. The state space is the robot state, environmental parameters, and task queue, the action space is the scheduling decision, and the reward function is the weighted combination of task completion time, resource utilization, energy consumption, and safety. The scheduling strategy parameters are updated using an online reinforcement learning algorithm. At the same time, the historical scheduling decisions and execution results are stored through an experience playback mechanism, and the model is updated in batches to avoid catastrophic forgetting. Step S6: The scheduling scheme is converted into robot joint control commands by the execution control layer. Commands are issued every 10ms and the execution process is monitored in real time. If an abnormal state is detected, an emergency plan is triggered.

2. The somatically intelligent robotic skill runtime data parsing and scheduling method of claim 1, wherein: The multi-level semantic parsing engine in step S2 adopts an action sequence modeling method based on a self-attention Transformer network. The self-attention Transformer network contains 8 attention heads. The input is the temporal feature vector of skill data, and the output is an action graph with attention weights. The attention weights quantify the correlation strength between any two actions. The correlation strength ranges from 0 to 1. The higher the value, the greater the impact of the action on the current task.

3. The somatically intelligent robotic skill runtime data parsing and scheduling method of claim 1, wherein: The multi-level semantic parsing engine in step S2 also implements a dynamic semantic completion mechanism: when there are missing fields or semantic ambiguity in the skill data, the action features of three adjacent time windows in the context are first matched, and then the typical parameter range of the same skill is retrieved from the domain knowledge base to generate at least two candidate completion schemes. The optimal scheme is selected to fill the missing field by confidence scoring. The domain knowledge base contains 1,200 standard skill parameters in three major fields: industrial assembly, medical surgery, and home care.

4. The method for analyzing and scheduling runtime data of embodied intelligent robot skills according to claim 1, characterized in that: The multi-level semantic parsing engine in step S2 supports the fusion parsing of multimodal skill data: it simultaneously receives RGB-D images collected by a visual sensor, contact pressure distribution collected by a tactile sensor, and six-dimensional force data collected by a force sensor. Through feature-level fusion, the three types of modal data are mapped to the same feature space, and cross-modal correlation features are extracted. The cross-modal correlation features are used to correct the error of single-modal parsing and improve the accuracy of skill representation.

5. The method for parsing and scheduling runtime data of embodied intelligent robot skills according to claim 1, characterized in that: The real-time scheduling controller in step S3 adopts a two-stage scheduling architecture with prediction-execution separation: the prediction stage uses an LSTM model to predict the task arrival rate, CPU / memory / communication bandwidth resource availability, and environmental change trends within the next 5 time windows, generating 3 candidate scheduling schemes; the execution stage combines the real-time feedback of robot status and environmental parameters, and selects the scheme with the highest comprehensive score from the 3 candidate schemes for execution, the comprehensive score being calculated by the task completion probability, resource consumption, and safety risk.

6. The method for parsing and scheduling runtime data of embodied intelligent robot skills according to claim 1, characterized in that: The real-time scheduling controller in step S3 adopts a time-slice-based soft real-time scheduling algorithm: the processor time is divided into fixed time slices of 10ms in length, and a maximum of 8 action tasks are scheduled in parallel within each time slice. The time slice allocation ratio is dynamically adjusted according to the deadline requirements of the tasks. Tasks with a deadline of less than 100ms are allocated no less than 60% of the time slice resources, while reserving 10% of the time slices as an emergency buffer to ensure that high-priority tasks are completed first.

7. The method for parsing and scheduling runtime data of embodied intelligent robot skills according to claim 1, characterized in that: The real-time scheduling controller in step S3 supports distributed collaborative scheduling: when there are multiple robots or multiple processor cores in the system, each node exchanges task status and resource information every 50ms through the message passing interface, and uses the distributed alternating direction multiplier method to calculate the global optimal scheduling scheme, so as to realize cross-node task load balancing and resource collaboration. When a single node fails, the task is automatically migrated to other available nodes.

8. The method for parsing and scheduling runtime data of embodied intelligent robot skills according to claim 1, characterized in that: The priority management module in step S4 also introduces a fairness constraint mechanism: the maximum waiting time threshold for a task is set to 300s. When the waiting time of a task increases by 10s, its comprehensive priority score is automatically increased by 5% to prevent low-priority tasks from being unable to execute due to continuous preemption. At the same time, a maximum priority limit is set to prevent a single task from excessively occupying resources and causing all other tasks to be blocked.

9. The method for parsing and scheduling runtime data of embodied intelligent robot skills according to claim 1, characterized in that: The adaptive optimization module in step S5 implements Pareto front tracking for multi-objective optimization: simultaneously optimizing four objectives: shortest task completion time, highest resource utilization, lowest energy consumption, and highest safety. The NSGA-II multi-objective evolutionary algorithm is used to solve for the Pareto optimal solution set, which contains at least 20 non-dominated solutions. Users can select the corresponding solution as the current scheduling strategy according to their preferences for specific application scenarios. The preference parameters can be adjusted through a visual interface.

10. The method for parsing and scheduling runtime data of embodied intelligent robot skills according to any one of claims 1 to 9, characterized in that: include: The skill data receiving module is configured to support four access protocols: TCP / IP, UDP, MQTT, and HTTP, to complete the reception and verification of skill data packets; A multi-level semantic parsing engine is configured to perform three-level progressive semantic parsing, action sequence modeling, dynamic semantic completion, and multi-modal fusion parsing. The real-time scheduling controller is configured to perform predictive-execution separation scheduling, time slice allocation, and distributed collaborative scheduling; the priority management module is configured to maintain a multi-dimensional priority evaluation system and a fairness constraint mechanism. The adaptive optimization module is configured to perform online reinforcement learning, experience replay, and multi-objective Pareto optimization. The execution control layer is configured to convert scheduling schemes into control commands and monitor the execution process; the modules are connected through a high-speed PCIe data channel with a data transmission bandwidth of no less than 10Gbps, realizing a closed-loop processing of skill data from reception, parsing, scheduling to execution.

Images