Robot and peripheral real-time interaction system and method based on edge computing

Abstract

The invention discloses a robot and peripheral real-time interaction system and method based on edge computing, and belongs to the technical field of robot control. The system comprises a region edge node which is used for determining task grading information according to semantic feature vectors of global perception data, performing global decision and scheduling shared standby nodes in a region redundant resource pool, region division is determined based on task throughput, data aggregation requirements and energy consumption constraints, and the shared standby nodes in the region redundant resource pool are distributed in the region edge node. The global sensing data comprises scene state data and micro sensing data; the ultra-near-end edge node is deployed based on task time delay limitation and communication reliability requirements, and is used for collecting scene state data and extracting semantic feature vectors from the micro-sensing data and the scene state data; and the robot terminal is used for collecting micro-sensing data and performing task execution based on task grading information. According to the method, the deterministic time delay of the high-critical task is ensured, and the utilization efficiency of the overall resources of the system is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robot control, and particularly relates to a real-time interaction system and method for robots and peripherals based on edge computing. BACKGROUND

[0002] In advanced industrial scenarios such as intelligent manufacturing, flexible automation and human-machine collaboration, robots need to interact with peripherals such as force sensors, vision cameras and electric clamps at a high frequency and with high reliability. Such applications have strict requirements for the real-time performance, determinacy and safety of the system. How to efficiently integrate global perception information and optimize the use of edge resources while ensuring extreme real-time control performance has become a core challenge for current robot intelligent control systems.

[0003] To meet the above requirements, an edge computing architecture is usually used to sink the computing and decision-making capabilities to the device side.

[0004] However, in a complex wireless environment with multiple robots and peripherals, the existing edge computing architecture cannot guarantee the determinacy delay of high-criticality tasks and the overall resource utilization of the system at the same time. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application provides a real-time interaction system and method for robots and peripherals based on edge computing, which ensures the determinacy delay of high-criticality tasks and improves the utilization efficiency of the overall system resources.

[0006] In a first aspect, the present application provides a real-time interaction system for robots and peripherals based on edge computing, which comprises at least one regional edge node, a plurality of super-close edge nodes and a plurality of robot terminals. Each regional edge node is connected to a micro-edge cluster, which comprises a master node and at least one shared standby node. The master node and the shared standby node are divided based on task classification information for the super-close edge nodes. The super-close edge nodes are connected based on the coupling of shared memory and time-gated scheduling with the robot terminals. The regional edge node is deployed based on the regional division of the robot terminal, and is used to determine the task classification information according to the semantic feature vector of global perception data, make global decisions, and schedule the shared standby node in the regional redundant resource pool. The regional division is determined based on task throughput, data aggregation requirements and energy consumption constraints. The global perception data includes scene state data and micro-sensing data. The super-proximal edge node is deployed based on a task latency limit and a communication reliability requirement, is configured to collect the scene state data, and is configured to extract the semantic feature vector from the micro-sensing data and the scene state data. The robot terminal is configured to collect the micro-sensing data and perform a task based on the task hierarchical information.

[0007] According to an embodiment of the present application, a control channel and a data synchronization channel are configured between the robot terminal and the super-proximal edge node. The super-proximal edge node is configured with a gate-shared memory manager, which is configured to, according to a time-gating scheduling mechanism, generate a time slot trigger signal, open access permission to a mapped logical block in an activated time slot, and enable an authorized robot terminal or super-proximal edge node to perform zero-copy read and write on a corresponding logical block in a shared memory region. The time-gating scheduling mechanism is configured to divide a communication period into exclusive time slots, periodic time slots, and dynamic time slots in time sequence, allocate exclusive time slots to the control channel, and allocate corresponding periodic time slots or dynamic time slots to the data synchronization channel, wherein the exclusive time slots have a higher priority than the periodic time slots and the dynamic time slots. The shared memory region is divided into a plurality of logical blocks, and the logical blocks include a control block and at least one data block. The control block is configured to map the control channel, so that the super-proximal edge node writes the task hierarchical information and control instructions through a kernel-mode memory mapping, and the robot terminal reads. The data block is configured to map the data synchronization channel, so that the robot terminal writes the micro-sensing data, and the super-proximal edge node reads.

[0008] According to an embodiment of the present application, the super-proximal edge node is configured to adjust the period and duty cycle of the time gate of the control channel and the data synchronization channel according to the task hierarchical information issued by the regional edge node, issue control instructions containing task hierarchical information to the robot terminal through the control block in the exclusive time slot of the control channel, allocate a data synchronization channel with a fixed position and a redundant period to a data stream with high criticality of the task hierarchical information, and map the data block to be continuous on the physical memory. The robot terminal is configured to receive the control instructions and respond, according to the task hierarchical information, to synchronize the micro-sensing data in the mapped data block through zero-copy in the authorized and activated time slot. The exclusive time slot of the control channel is located at the starting position of each communication period, and the communication period is synchronized with the control decision period of the robot terminal.

[0009] According to one embodiment of the present application, the super-close edge node integrates a joint arbitrator; The joint arbitrator is configured to adjust the suppression intensity of the data transmission bandwidth of the data synchronization channel according to a deviation value between the performance index of the control channel and the service level target corresponding to the task classification information, when the deviation between the end-to-end delay of the control channel and the upper limit of tolerance corresponding to the task classification information is less than a preset safety margin, and the deviation value is negatively correlated with the suppression intensity; The joint arbitrator is configured with a back pressure feedback unit, which is configured to send a flow control signal to the robot terminal to suspend or reduce the rate of writing micro-sensing data to the data block by the robot terminal; The super-close edge node is further configured to release the bandwidth suppression of the data synchronization channel and restore the transmission rate of the data synchronization channel when the end-to-end delay of the control channel continues to be lower than the safety recovery threshold for a preset stable duration.

[0010] According to one embodiment of the present application, the super-close edge node is configured to: The semantic feature vector is differentially encoded online, and a change summary data is generated based on the time sequence difference degree of the semantic feature vector and the historical semantic feature vector; When the time sequence difference degree exceeds a difference degree threshold, the change summary data is triggered to be uploaded to the regional edge node asynchronously, the change summary data is compressed in a differential encoding format, including the semantic dimension that has changed and the incremental value of the semantic dimension, for internal state log recording and anomaly detection; The regional edge node is configured to receive the change summary data to update the global perception data.

[0011] According to one embodiment of the present application, the deployment of the master node and the shared standby node in the micro-edge cluster is determined based on the joint optimization of radio frequency propagation characteristics, network hop count and switching delay; The radio frequency propagation characteristics include path loss, multipath delay spread and channel coherence time, and are used to evaluate the lower limit of the delay jitter of the wireless link; The network hop count is used to represent the number of intermediate forwarding devices from the super-close edge node to the regional edge node; The switching delay is used to represent the transmission delay introduced by the queuing and processing of the intermediate nodes in the network in the edge network path from the super-close edge node to the regional edge node; The regional edge node is configured to construct a deployment cost function based on the radio frequency propagation characteristics, the network hop count, and the switching delay, to determine an optimal deployment coordinate set of the micro-edge cluster.

[0012] According to an embodiment of the present application, the regional edge node is configured to analyze assistance intention information of each robot terminal in the robot queue, and group super-near edge nodes involved in the robot queue and corresponding shared backup nodes thereof into a virtual joint node in a case where it is detected that the robot terminals have a cooperative operation demand, the assistance intention information including task dependency relationship, spatial cooperation demand, and peripheral device sharing request. All super-near edge nodes in the virtual joint node share resource views and scheduling strategies, and are subjected to cooperative task allocation and load balancing by the regional edge node.

[0013] According to an embodiment of the present application, the task classification information includes high-criticality tasks, medium-criticality tasks, and low-criticality tasks. The operating system kernel layer of the super-near edge node is configured to construct mutually isolated hardware resource slices according to the task classification information. A reflection control task in the high-criticality tasks is allocated an exclusive CPU core, a continuous physical memory interval, and a last-level cache partition. A semantic feature extraction or data aggregation task in the medium-criticality tasks and the low-criticality tasks is allocated a limited computing resource slice.

[0014] According to an embodiment of the present application, the super-near edge node is further configured to extract a semantic feature vector through progressive semantic feature coding. The progressive semantic feature coding generates a hierarchical data structure including at least a first coding layer and a second coding layer. The first coding layer is a key state code including a limited state identifier, used for transmission within a first time window, and representing a security level of current operation and an outline state of the system. The second coding layer is a complete semantic feature description, used for transmission within a second time window, and including complete dimension and precision information of the semantic feature vector. The regional edge node sends a feature request instruction to the super-near edge node according to a current network state and an urgency corresponding to the task classification information, to receive the first coding layer, or the first coding layer and the second coding layer.

[0015] In a second aspect, the present application provides a real-time interaction method of robots and peripherals based on edge computing, applied to a real-time interaction system of robots and peripherals based on edge computing as in the first aspect, and the method includes: The task classification information is determined according to a semantic feature vector of global perception data through a regional edge node, global decision is made, and the shared backup node in the regional redundant resource pool is scheduled, wherein the global perception data includes scene state data and micro-sensing data. The scene state data is collected through a super-near edge node, and the semantic feature vector is extracted from the micro-sensing data and the scene state data. The micro-sensing data is collected through a robot terminal, and task execution is performed based on the task classification information.

[0016] Additional aspects and advantages of the present application will be made apparent by the following description and the accompanying drawings.

[0017] The real-time interaction system and method of robots and peripherals based on edge computing provided by the present application have the following beneficial effects compared with the prior art: (1) Through the deep coupling of shared memory zero-copy and time-gated scheduling between the super-near edge node and the robot terminal, the uncertain delay caused by the traditional network protocol stack is avoided, and the deterministic execution of high-critical tasks is ensured; the regional edge node performs task classification and global scheduling based on the semantic feature vector, and uses the shared backup node pool across clusters to realize elastic disaster recovery, which not only avoids resource redundancy waste, but also ensures system high availability; the system can adapt to environmental changes, reduce interaction latency in complex wireless environments, ensure deterministic latency of high-critical tasks, improve the utilization efficiency of overall system resources, and improve the real-time perception and control capability of robots to peripherals.

[0018] (2) Through the double-channel cooperative mechanism in the robot terminal, the control channel completes instruction issuing in a zero-copy manner in exclusive time slots, which can stabilize the end-to-end latency and jitter, and meet the real-time requirements of high-critical tasks of industrial robots; the data synchronization channel uploads micro-sensing data in isolated time slots, avoiding competition with control traffic for bus bandwidth, while the rotation mechanism of logical blocks prevents read-write conflicts, reduces uncertain delay and kernel overhead, improves system response speed, resource utilization and running stability, and provides bottom support for real-time control of high-safety and high-precision robots. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which: Figure 1 is a structural schematic diagram of the real-time interaction system of robots and peripherals based on edge computing provided by the embodiments of the present application; Figure 2This is a flowchart illustrating a real-time interaction method between a robot and peripheral devices based on edge computing, as provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0022] The following description, in conjunction with the accompanying drawings, details the real-time interaction system and method for robot-peripheral based on edge computing provided in this application, through specific embodiments and application scenarios.

[0023] like Figure 1 As shown, the real-time interaction system between the robot and peripherals based on edge computing includes: at least one regional edge node 110, multiple ultra-near-end edge nodes 120 and multiple robot terminals 130. Each regional edge node 110 is connected to a micro-edge cluster. The micro-edge cluster includes a master node and at least one shared backup node. The master node and the shared backup node are divided into ultra-near-end edge nodes 120 based on task classification information. The ultra-near-end edge node 120 and the robot terminal 130 are connected based on the coupling of shared memory and time-gated scheduling; The regional edge node 110 is deployed based on the regional division of the robot terminal 130. It is used to determine the task classification information according to the semantic feature vector of the global perception data, make global decisions, and schedule the shared backup nodes in the regional redundant resource pool. The regional division is determined based on task throughput, data aggregation requirements and energy consumption constraints. The global perception data includes scene status data and micro-sensor data. The ultra-proximal edge node 120 is deployed based on a task latency limit and a communication reliability requirement, and is used to collect the scene state data, and extract the semantic feature vector from the micro-sensing data and the scene state data. The robot terminal 130 is used to collect the micro-sensing data, and perform a task based on the task classification information.

[0024] It can be understood that the regional edge node (REN) is a regional-level high-performance edge server deployed in a workshop or a production line, and is used to manage a physical region, and is responsible for global perception, task classification and resource scheduling across multiple robot units. The regional edge node is configured with a load perception module and a node management module, which are used to issue start, suspend or downgrade instructions according to the business load state of each ultra-proximal edge node, so as to realize dynamic start and stop of the ultra-proximal edge node. The node load table and the business priority table are also maintained, and the load perception module determines whether to perform a start or suspend operation on the ultra-proximal edge node based on the data about CPU occupation, network bandwidth, number of online robots and number of real-time tasks in the node load table.

[0025] The ultra-proximal edge node (UPEN) is a lightweight edge computing unit deployed in close proximity to the physical location of a single or multiple robot terminals, and the distance between the UPEN and the connected robot terminal is within 10 meters. The UPEN is used for local real-time control and semantic feature extraction, for example, an embedded industrial computer or a customized system on chip (SoC). The ultra-proximal edge node also includes a protocol adaptation module, which is used to convert the control message of the robot terminal into a control instruction recognizable by an external device such as an access control, elevator control, AGV scheduling system or camera trigger signal.

[0026] The micro-edge cluster is a logical computing unit composed of a master UPEN and at least one shared standby UPEN, and serves a group of geographically adjacent robot terminals, and is used to provide high-availability local edge services.

[0027] The master node and the shared standby node are role division in the micro-edge cluster, wherein the master node undertakes exclusive processing of high-criticality tasks, and the shared standby node is in a low-power standby state and can be multiplexed by multiple clusters to improve resource efficiency.

[0028] A robot terminal is an intelligent device with built-in sensors and actuators for performing physical actions, such as industrial robot arms, automated guided vehicles (AGVs), collaborative robots (Cobots), etc. A micro-cache and local preprocessing module are also provided in the robot terminal to temporarily store data and extract features from collected data when network anomalies or latency increases, and only send the preprocessing results to the ultra-close edge node or the regional edge node. In addition, the control channel continues to receive control instructions from the ultra-close edge node when the data synchronization channel is interrupted, and the data in the micro-cache is batched and returned after the data synchronization channel is restored.

[0029] Task priority information is a criticality label of a robot task according to real-time performance, safety, and consequence severity, which is divided by a regional edge node according to a semantic feature vector, and is used to guide resource allocation strategies. For example, the criticality label of an emergency stop or force control loop is high criticality, the criticality label of path tracking is medium criticality, and the criticality label of log uploading is low criticality.

[0030] Shared memory is a common memory area established by a UPEN and a robot terminal through a Peripheral Component Interconnect Express (PCIe) or high-speed Ethernet, and both sides directly read and write through memory mapping (mmap) without kernel copying, achieving zero-copy.

[0031] Time-gated scheduling is based on Time-Sensitive Networking (TSN) mechanism, which divides the communication cycle into fixed time slots, and only allows specific traffic to be sent in designated windows.

[0032] Regional division is to divide the physical space into several management regions according to task load, data aggregation demand, and energy consumption constraints, and each region is governed by a REN.

[0033] Task throughput is the number of control or computing tasks that a robot terminal in a region needs to complete per unit time.

[0034] Data aggregation demand is whether the raw data generated by multiple robot terminals or sensors in a region needs to be fused, compressed, or semantically extracted locally before uploading. For example, multiple AGVs need to fuse position data to generate a global path when cooperating to transport, which has a high data aggregation demand and requires a UPEN cluster with strong collaborative perception capability to be deployed within the region.

[0035] Energy consumption constraint is the energy limit that the robot terminal in the region is subjected to when performing tasks, including the maximum power consumption upper limit, the endurance time requirement or the charging cycle limit; the high energy consumption constraint region needs to optimize the task offloading strategy, reduce the communication overhead, adjust the region boundary to shorten the wireless transmission distance, and prolong the device running time.

[0036] Task latency limit is the maximum end-to-end time limit allowed for a certain type of robot task from the issuance of the instruction to the completion of the execution, which is determined by the criticality level of the task. For example, the latency limit of a high criticality task is usually 1-5 milliseconds, the latency limit of a medium criticality task is 10-20 milliseconds, and the latency limit of a low criticality task is 50-100 milliseconds.

[0037] Communication reliability requirement is the requirement of the task for the data transmission success delivery probability and the anti-interference ability, which is quantified by indicators such as packet loss rate (PLR) and bit error ratio (BER), determines the type of physical link to be used between the UPEN and the robot terminal, whether to enable redundant transmission, and whether to need hardware-level error checking, and affects the deployment distance of the UPEN, the network topology and the hardware selection.

[0038] Global perception data covers the environment and device state information of the entire region; in the global perception data, scene state data includes visual, laser radar, infrared and other environmental perception data collected by the UPEN, and micro-sensing data includes joint torque, encoder position, current, temperature and other bottom sensor data collected by the robot terminal.

[0039] Semantic feature vector is a structured high-level semantic representation extracted from raw global perception data by the edge lightweight convolutional neural network (CNN) model built-in the super-proximal edge node, represented as a fixed-dimension floating-point or integer array, for example, [object type = glass, position = (x, y, z), risk level = high].

[0040] In actual execution, REN is installed according to physical layout in the factory workshop, for example, one per 500㎡, and UPEN is deployed in the control cabinet integrated within 1-10 meters near each robot terminal, and the robot terminal is connected to the actuator through the motor driver and sensor bus.

[0041] REN starts to scan all UPENs in the region, and evaluates the performance of each UPEN according to the computing power, network delay and health status of each UPEN. The UPEN with the best performance is designated as the master node, and the remaining UPENs are marked as shared standby and added to the regional redundancy resource pool. The roles of the master node and the shared standby node can be dynamically switched through the Google Remote Procedure Call (gRPC) heartbeat protocol.

[0042] The UPEN and the robot terminal are connected directly through a 10Gbe Tsn network card or a PCIe Gen4, and the operating system of the UPEN is configured with a Portable Operating System Interface X (POSIX) shared memory. At the same time, the 802.1Qbv function of the TSN switch is enabled, and the time synchronization and time slot table are configured.

[0043] The REN collects historical data, and according to the task throughput corresponding to the number of control cycles per unit time, the data aggregation demand corresponding to the multi-machine coordination frequency, and the energy consumption constraint corresponding to the upper limit of battery capacity / power, runs the K-means clustering algorithm to group the robot terminals, and each group forms a management region.

[0044] The UPEN fuses the micro-sensing data and global sensing data in the global perception data, calls the pre-trained MobileNetV3-Tiny model, and performs inference on the fused global perception data to output a 128-dimensional JSON format semantic feature vector for local control or uploading to the REN.

[0045] The REN receives the semantic feature vectors from each UPEN, and inputs the semantic feature vectors into a pre-trained lightweight Long Short-Term Memory (LSTM) classifier to generate task classification information.

[0046] The REN reassigns the tasks of high-load UPENs to idle standby nodes according to the task classification results. In the case of heartbeat timeout of any master UPEN, the optimal standby UPEN is selected from the regional redundancy resource pool to be activated, and the virtual Intellectual Property (IP) drift is used to realize the seamless switching.

[0047] REN decomposes the end-to-end latency limit of the task into specific latency targets of the wireless transmission segment by the latency budget decomposition method, and quantifies the reliability requirement into a minimum received signal strength (RSSI) and a maximum packet loss rate threshold in the case of deploying UPEN; a digital radio frequency environment map integrating field measurement and simulation is constructed based on the field deployment of the radio frequency probe sampling and the ray tracing model simulation, and the topology hop number and exchange delay between the candidate deployment point and the robot terminal are determined in combination with network planning; a multi-objective optimization algorithm is used to solve the cost function integrating the task weight, predicted latency and link quality, and the optimal deployment coordinate set is output in the solution space satisfying all task QoS constraints; the physical deployment of UPEN is completed according to the set, and the dynamic optimization of the deployment position is realized through the continuous performance monitoring and the self-adaptive triggering mechanism to automatically start the re-optimization process when the performance degrades or the task distribution changes. UPEN connects the camera through the GigEVision interface and collects scene state data through the CAN bus; at the same time, the robot terminal receives the micro-sensing data uploaded by the TSN.

[0048] The robot terminal reads the micro-sensing data of the six-dimensional force sensor, encoder and the like, and after receiving the control instruction issued by UPEN, it executes the high-critical task immediately interrupts other operations according to the task classification information, and the low-critical task is queued for execution.

[0049] According to the real-time interaction system between the robot and the peripheral device based on edge computing provided by the embodiments of the application, through the deep coupling of the ultra-close edge node and the robot terminal through shared memory zero-copy and time-gated scheduling, the uncertain delay caused by the traditional network protocol stack is avoided, and the deterministic execution of high-critical tasks is ensured; the regional edge node performs task classification and global scheduling based on the semantic feature vector, and uses the shared standby node pool across clusters to realize elastic disaster recovery, which not only avoids resource redundancy waste, but also ensures system high availability, the system can adapt to environmental changes, reduce interaction delay in complex wireless environment, ensure deterministic latency of high-critical tasks, improve the utilization efficiency of overall system resources, and improve the real-time perception and control ability of the robot to the peripheral device.

[0050] In some embodiments, a control channel and a data synchronization channel are configured between the robot terminal and the ultra-close edge node; A gated shared memory manager is configured on the ultra-close edge node, which is used to generate a time slot trigger signal according to a time-gated scheduling mechanism, and in the activated time slot, the access permission of the mapped logical block is opened to enable the authorized robot terminal or the ultra-close edge node to perform zero-copy read and write on the corresponding logical block in the shared memory area; The time-gated scheduling mechanism is used to divide a communication period into exclusive slots, periodic slots and dynamic slots in time sequence, and the control channel is allocated an exclusive slot, and the data synchronization channel is allocated a corresponding periodic slot or dynamic slot, and the priority of the exclusive slot is higher than that of the periodic slot and the dynamic slot. The shared memory region is divided into a plurality of logical blocks, and the logical blocks include a control block and at least one data block. The control block is used to map the control channel, so that the super-proximal edge node writes the task hierarchical information and control instructions through kernel-mode memory mapping, and the robot terminal reads. The data block is used to map the data synchronization channel, so that the robot terminal writes the micro-sensing data, and the super-proximal edge node reads.

[0051] It can be understood that the control channel is a high-priority communication path for transmitting high-criticality real-time control instructions, and the data content in the control channel includes task hierarchical information, motion control commands, safety emergency stop signals, etc., requiring low end-to-end delay, small jitter and high reliability, and based on the low-latency profile of DDS, a long connection with the super-proximal edge node is maintained.

[0052] The data synchronization channel is a communication path for uploading non-critical or medium-low priority service data, mainly carrying micro-sensing data collected by the robot terminal, such as joint current, encoder value, temperature, etc., as well as uploading task results, alarms, video thumbnails, allowing certain delay and bandwidth fluctuation.

[0053] The gated shared memory manager is a module deployed in the operating system kernel of the super-proximal edge node (UPEN), responsible for dynamically controlling the access rights of the shared memory region within the predefined communication period according to the precise time slot trigger signal generated by the time-gated scheduling mechanism.

[0054] The time-gated scheduling mechanism is implemented based on the IEEE 802.1Qbv standard, and each fixed-length communication period is divided into a plurality of strictly time-sequenced transmission windows, including exclusive slots, periodic slots and dynamic slots, through a time-sensitive network (TSN) switch or an intelligent network card.

[0055] Among them, the exclusive slot is a high-priority time window reserved for the control channel and cannot be preempted, and only control instructions are allowed to be transmitted in this window; the periodic slot is allocated to a regular data stream according to a fixed period, and the dynamic slot is flexibly allocated by a scheduler according to real-time load.

[0056] The shared memory region is a segment of physical memory space established by UPEN and robot terminal through PCIe direct connection, and both sides directly access through memory mapping, avoiding the kernel copy overhead of traditional socket communication.

[0057] The logical block is a logical division unit of the shared memory region, which is isolated from each other to prevent read-write conflict.

[0058] In the logical block, the control block is a logical block dedicated to the control channel, and the task classification information and control instructions are written into it by UPEN through memory mapping in the kernel mode through the UIO (Userspace I / O) driver. The application program or real-time control loop of the robot terminal directly reads the content of the block through the user state pointer after authorization, realizes zero-copy read-write, and forms a spatial and temporal double-isolated deterministic interaction channel. The data block is one or more logical blocks for the data synchronization channel, and the micro-sensing data is written in a structured format by the robot terminal in the specified time slot, and UPEN reads and processes it in the subsequent time slot.

[0059] In actual execution, a connection is established between UPEN and robot terminal through a 10 GbE physical link supporting TSN, and a shared memory region is created using the POSIX shared memory interface; A custom gated shared memory manager driver module is loaded in the UPEN kernel, which subscribes to the periodic clock interrupt provided by the TSN time synchronization protocol, and generates time slot trigger signals according to the preconfigured time slot table; At system startup, the shared memory is divided into a control block and several data blocks, and the access permission is set through memory protection keys (MPK) or page table attributes; The time-gated scheduling mechanism is configured by the TSN scheduler on UPEN, and each communication period is divided to obtain exclusive time slots, periodic time slots and dynamic time slots.

[0060] For example, the communication period is 1 ms, of which 0-50 μs is the exclusive time slot for the control channel, 50-900 μs is the periodic time slot for the regular sensing data of the data synchronization channel, and 900-1000 μs is the dynamic time slot for the burst log of the data synchronization channel.

[0061] When entering the exclusive time slot, the gate shared memory manager automatically removes the write protection of the control block, and the UPEN kernel thread immediately writes the latest control instruction and task classification information into the block. At the same time, the real-time control task of the robot terminal atomically reads the instruction and executes through the pre-mapped user state pointer after detecting the activation of the time slot. In the periodic or dynamic time slot, the robot terminal writes the packaged micro-sensing data into one of the rotating data blocks, and the UPEN completes reading and releases the block for the next round before the end of the time slot; the whole process does not need the central processing unit (CPU) to participate in data copying, nor does it need context switching, and all operations are completed within the pre-allocated time slot window.

[0062] In addition, the robot terminal sets a local 1-3 second micro-cache, when the data synchronization channel is interrupted, the control channel can still receive the access control issued by the ultra-close end, and the robot terminal preferentially completes the action; after the data synchronization channel is restored, the missing state is transmitted in batches.

[0063] In this embodiment, through the dual-channel cooperative mechanism of the robot terminal, the control channel completes the instruction issuance in the exclusive time slot in a zero-copy manner, which can stabilize the end-to-end time delay and jitter, and meet the real-time requirements of high-critical tasks of industrial robots; through the data synchronization channel, the micro-sensing data is uploaded in the isolation time slot, avoiding competition with control traffic for bus bandwidth, and the rotation mechanism of the logical block prevents read-write conflict, reduces uncertain delay and kernel overhead, and improves system response speed, resource utilization and running stability, providing bottom support for high-safety and high-precision robot real-time control.

[0064] In some embodiments, the ultra-close edge node is configured to adjust the period and duty cycle of the time gates of the control channel and the data synchronization channel according to the task classification information issued by the regional edge node, and issue control instructions containing task classification information to the robot terminal through the control block in the exclusive time slot of the control channel. The robot terminal is configured to receive the control instructions and respond, and synchronize the micro-sensing data through zero-copy in the mapped data block in the authorized and activated time slot according to the task classification information. The exclusive time slot of the control channel is located at the starting position of each communication period, and the communication period is synchronized with the control decision period of the robot terminal.

[0065] It can be understood that adjusting the period and duty cycle of the time gate dynamically modifies the repetition interval and effective transmission window length of each communication time slot in the time sensitive network (TSN); The fixed position is that the transmission time slot allocated for the high criticality data stream in the communication cycle defined by the time-gated scheduling mechanism has a determined and unchanged time offset. The relative time interval between the start time of the time slot in each cycle and the start point of the cycle is constant, ensuring that the receiving end can accurately predict the data arrival time and avoiding synchronization failure caused by scheduling jitter.

[0066] The redundancy period is a plurality of repeated transmission opportunities configured for the high criticality data stream to improve reliability. The continuity of the physical memory is used to represent that the data blocks allocated to the high criticality data stream are arranged continuously in the dynamic random access memory (DRAM) address space, avoiding the access delay caused by page table jumping. The control decision period is the time interval for the robot terminal to execute a control algorithm once. The authorized and activated time slot is that the robot terminal can read and write the shared memory only in the time slot explicitly opened by the time-gated mechanism and has access permission. Zero-copy synchronization refers to directly reading and writing data through shared memory without the involvement of CPU for copying or protocol stack processing.

[0067] A shared memory pool is set up in the ultra-near edge node, each data frame is written only once, and multiple consumer ends such as access control driver, video analysis plug-in, and task coordinator only read pointers without copying.

[0068] The robot terminal is configured with a shared memory type data exchange module and a priority queue scheduling module, which are used to locally realize low-latency data interaction between the robot and peripherals; for example, control type messages such as opening the door, elevator to the 3rd floor, and obstacle avoidance pause enter the high priority queue, and video or point cloud type big data frames enter the low priority queue or multicast distribution, without blocking control type messages.

[0069] In actual execution, after completing task classification, a regional edge node (REN) issues task classification information to a corresponding ultra-proximal edge node (UPEN) through a reliable message queue; after receiving the information, the UPEN calls a TSN scheduler interface to dynamically reconfigure a time gating table: if the task is high criticality, the exclusive time slot period of the control channel is set to be consistent with the robot control decision period, and is placed at the start of each period to ensure the first execution, and the periodic time slot of the data synchronization channel is adjusted to a shorter period and two backup time slots of the same position are allocated in the same communication period to increase the redundancy period and prevent single transmission failure; for the data stream corresponding to the high criticality task, the UPEN uses mmap to cooperate with HugeTLB to apply a large page of physical address continuous memory when initializing shared memory, divides it into multiple continuous data blocks, and maps the area to the robot terminal; the robot terminal loads the predefined time slot authorization table when starting, and runs in the control task loop of the real-time kernel to continuously monitor the time slot interrupt signal generated by the TSN hardware; when detecting that the current enters the authorized exclusive time slot, the control instruction containing the task classification information is immediately read from the control block and the corresponding action is executed; in the allocated data synchronization time slot, the robot terminal directly writes the packaged micro-sensing data into the rotating continuous data block, and the UPEN completes reading before the end of the time slot, thereby completing a deterministic interaction period from control to data synchronization.

[0070] It should be noted that in the above process, the hardware timestamp and the TSN gating are accurately synchronized to ensure that the communication period is strictly aligned with the control decision period.

[0071] In this embodiment, by configuring the exclusive time slot for high criticality tasks, the supporting data stream also obtains higher transmission reliability and lower access delay due to the redundancy period and physically continuous memory, the control instruction is issued at the beginning of each period to ensure the timeliness of the control closed loop; zero-copy and accurate time slot cooperation eliminate the queuing, copying and scheduling jitter in traditional communication, while supporting multi-level task on-demand bandwidth allocation.

[0072] In some embodiments, the ultra-proximal edge node integrates a joint arbitrator; The joint arbitrator is configured to, in a case where a deviation between an end-to-end delay of the control channel and a tolerance upper limit corresponding to the task classification information is less than a preset safety margin, adjust a suppression intensity of data transmission bandwidth of the data synchronization channel according to a deviation value between a performance index of the control channel and a service level target corresponding to the task classification information, the deviation value being negatively correlated with the suppression intensity. The joint arbitrator is configured with a back pressure feedback unit, which is used to send a flow control signal to the robot terminal to suspend or reduce the rate of writing micro-sensing data to the data block by the robot terminal; The ultra-proximal edge node is further configured to release the bandwidth suppression on the data synchronization channel and restore the transmission rate of the data synchronization channel after the end-to-end latency of the control channel continuously being lower than the safe recovery threshold for a preset stable duration.

[0073] It can be understood that the joint arbitrator is a hardware-assisted or kernel-level software module integrated in the ultra-proximal edge node (UPEN) for real-time monitoring of control channel performance and dynamic coordination of resource allocation of control and data channels.

[0074] The end-to-end latency of the control channel is used to describe the complete closed-loop time from the generation of the control instruction by the UPEN to the feedback of the execution by the robot terminal, which is calculated by time stamping in the shared memory and by the synchronized clock of both parties.

[0075] The tolerance upper limit corresponding to the task classification information is the maximum latency threshold allowed by the system for each task level, for example, 5 milliseconds for high criticality tasks and 20 milliseconds for medium criticality tasks.

[0076] The preset safety margin is a conservative buffer value used to trigger the protection mechanism in advance to avoid the latency approaching the limit.

[0077] The performance index is a specific measurable parameter used to quantify the current running quality of the control channel, which is collected in real time by hardware time stamping, shared memory marking, or TSN probe, for example, the performance index includes end-to-end latency, latency jitter, and packet loss rate, and the latency jitter is used to represent the standard deviation or peak-to-peak fluctuation of the latency.

[0078] The service level objective (SLO) corresponding to the task classification information is a set of indicators predefined by the system for each level of task, including latency, jitter, reliability, etc.

[0079] The suppression intensity of the data transmission bandwidth is used to represent the degree of limitation of the available bandwidth of the data synchronization channel, which can be realized by reducing the token bucket rate, suspending transmission, reducing the sampling frequency, or compressing the data volume, etc.; the higher the suppression intensity, the less network or computing resources occupied by the data synchronization channel.

[0080] Wherein, the smaller the deviation value, the closer the actual latency to the tolerance upper limit, and the stronger the suppression.

[0081] The flow control signal is an instruction sent through a shared memory flag, a TSN Pause frame, or a custom User Datagram Protocol (UDP) control packet, instructing the robot terminal to slow down or suspend data upload; The safety recovery threshold is a more relaxed recovery criterion than the tolerance upper limit, used to avoid frequent start-stop; The preset stable duration is a time window required for the time delay to be continuously lower than the recovery threshold, to ensure that the system is truly in a stable state before the inhibition is lifted.

[0082] In actual implementation, the UPEN continuously monitors the end-to-end time delay of the control channel at runtime, records a high-precision timestamp before writing an instruction to the control block each time through the Precision Time Protocol (PTP) hardware clock, and writes back a confirmation timestamp after the robot terminal completes execution, and the UPEN calculates the closed-loop time delay after reading it; The joint arbitrator samples the time delay every fixed period and compares it with the tolerance upper limit corresponding to the current task level. If the deviation value is less than the preset safety margin, it is determined that there is a risk. At this time, the joint arbitrator calculates the inhibition intensity through a linear function according to the deviation value, to ensure that the control closed loop of high criticality tasks is not disturbed, and to reduce the token bucket rate of the data synchronization channel. At the same time, the back pressure feedback unit sets a 1-byte flow control flag in the shared memory, and the robot terminal's driver program polls this flow control flag regularly. Once the pause signal is detected, it immediately stops writing new data to the data block or switches to a low-frequency sampling mode. When subsequent monitoring shows that the control channel time delay is continuously lower than the safety recovery threshold for 10 communication periods, the joint arbitrator determines that the system has recovered to stability, clears the flow control flag, and gradually restores the token bucket rate to the original value, completing the lifting of bandwidth inhibition.

[0083] For example, the inhibition intensity based on the nonlinear mapping of the safety margin is: wherein, is the bandwidth inhibition intensity applied to the data synchronization channel, with a value range of The larger the value is, the stronger the restriction is. is the maximum inhibition intensity, with a preset upper limit of the system; is the safety decay factor, used to control the degree of nonlinearity and steepness of the inhibition intensity as the risk rises; is the safety margin between the current performance of the control channel and the tolerance upper limit, defined as , is the current measured end-to-end time delay of the control channel, is the tolerance upper limit of the kth level task; is the preset safety margin threshold. When , the inhibition mechanism is triggered.

[0084] The back pressure trigger condition introduced by the service level deviation is: wherein, is the service level objective (SLO) of the kth task, i.e. the ideal expected latency; is a preset safety margin ratio, used to trigger back pressure in advance to avoid latency approaching the limit.

[0085] In this embodiment, by actively suppressing the bandwidth occupation of non-critical data flow when the control channel is at risk of latency deterioration, the deterministic execution of high-critical tasks is prioritized; the suppression intensity is dynamically adjusted according to the risk level to avoid data flow interruption caused by one-size-fits-all; the back pressure mechanism ensures that the suppression instruction can quickly reach the robot terminal source, reducing the load at the generation end; and the recovery condition introduces a stable time length judgment to prevent frequent oscillation caused by transient fluctuations.

[0086] In some embodiments, the ultra-close edge node is configured to: perform online differential encoding on the semantic feature vector, generate change summary data based on the time sequence difference between the semantic feature vector and the historical semantic feature vector; in the case where the time sequence difference exceeds a difference threshold, trigger asynchronous uploading of the change summary data to the regional edge node, the change summary data being compressed in a differential encoding format, including the semantic dimension that has changed and the incremental value of the semantic dimension, for internal state log recording and anomaly detection; the regional edge node is configured to receive the change summary data to update the global perception data.

[0087] It can be understood that online differential encoding is a compression method that calculates the difference between the current feature and the historical feature in real time while the data is being generated, and only retains the changed part; the historical semantic feature vector is a semantic feature snapshot cached by the UPEN in the previous period or the last valid state; the time sequence difference is a quantitative index for measuring the degree of change between the current and historical semantic feature vectors, which can be calculated by Manhattan distance or weighted Hamming distance, for example, giving higher weight to key dimensions such as risk level; the change summary data is a compact data structure that only contains the semantic dimension that has changed and the incremental value of the semantic dimension; the semantic dimension is a structured attribute component in the semantic feature vector that has an independent physical or logical meaning, used to represent the high-level state or environmental understanding of the robot system at a certain time.

[0088] the difference threshold is a preset sensitivity threshold for judging whether the change has semantic meaning; asynchronous uploading is a non-blocking way of sending data to the regional edge node (REN) immediately after an event is triggered, without relying on a fixed period; The differential encoding format is a self-defined binary protocol that only transmits field identity document (ID), old value, new value, or increment, greatly reducing data volume; Internal state logging is writing a change summary by UPEN into a local ring buffer for debugging or auditing; Anomaly detection is triggering local alarm or security policy through mutation pattern.

[0089] In actual execution, at the end of each control cycle, UPEN uses the AI model optimized by TensorRT inference framework to infer the fused micro-sensing and scene data, outputting the current semantic feature vector; the kernel differential engine is called to compare the semantic feature vector with the historical semantic feature vector cached in the double data rate (DDR) memory dimension by dimension, calculate the weighted L1 difference, for example, the weight of the emergency stop signal dimension is set to 10, and the weight of the normal position dimension is set to 1; if the difference exceeds the preset threshold, traverse all dimensions, record only the fields whose changes exceed the minimum resolution, and generate a compact change summary data; the change summary data is serialized into a self-defined binary format and sent to REN asynchronously through the TSN low-priority queue; at the same time, UPEN writes the change summary data into the local ring log buffer and sends it to the lightweight anomaly detection module based on 1D-CNN to determine whether to trigger local security action; the REN runs a listening service that, after receiving the change summary data, parses the change dimension and increment value, directly updates the global perception state table in its memory, replacing the full refresh without the need to re-aggregate all original data.

[0090] For example, the semantic perception weighted time series difference is: wherein, is the weighted time series difference between the current and historical semantic feature vectors, used to determine whether to trigger change summary upload; is the semantic feature vector extracted at the current time t; is the historical semantic feature vector at the last valid time t 1; is the element-wise (Hadamard) product; is the state-dependent weight matrix, each element is dynamically adjusted according to the historical state.

[0091] The weight of the i-th semantic dimension is : wherein, a risk level dimension in the history vector; .

[0092] In this embodiment, by uploading data only when there is a meaningful change in the environment or task state, the waste of bandwidth caused by traditional periodic full reporting is avoided; by differential encoding, the amount of data uploaded at a time is reduced; asynchronous transmission does not affect the local control loop; the global state of REN is continuously and accurately updated, and UPEN can also use change summaries to achieve fast local exception response, ensuring the timeliness of perception while reducing uplink load and prolonging the endurance of mobile robots, and enabling sparse and efficient communication driven by semantics.

[0093] In some embodiments, the deployment of the master node and the shared backup node in the micro-edge cluster is jointly optimized and determined based on radio frequency propagation characteristics, network hop count and switching delay; The radio frequency propagation characteristics include path loss, multipath delay spread and channel coherence time, which are used to evaluate the lower limit of the delay jitter of the wireless link; The network hop count is used to represent the number of intermediate forwarding devices from the ultra-close edge node to the regional edge node; The switching delay is used to represent the transmission delay introduced by queuing and processing of network intermediate nodes in the edge network path from the ultra-close edge node to the regional edge node; The regional edge node is used to construct a deployment cost function based on the radio frequency propagation characteristics, the network hop count and the switching delay, to determine the optimal deployment coordinate set of the micro-edge cluster.

[0094] It can be understood that the master node is a UPEN currently undertaking main computing and control tasks, responsible for processing high-criticality tasks, and the end-to-end communication delay between the master node and the robot terminal under its jurisdiction meets the upper limit of the tolerance of high-criticality tasks; the shared backup node is a UPEN in a low-power standby state, which can be multiplexed by multiple micro-edge clusters, and takes over the tasks of the master node when the master node fails.

[0095] The radio frequency propagation characteristics are used to describe the physical behavior of wireless signal transmission in space. In the radio frequency propagation characteristics, the path loss is used to represent the degree of signal strength attenuation with distance, the multipath delay spread is used to represent the time delay spread caused by the multi-path arrival of signals due to reflection and refraction, and the channel coherence time is used to represent the length of time that the channel remains stable, reflecting the speed of channel time variation. These parameters together determine the lower limit of the delay jitter that the wireless link can achieve, i.e. the theoretically optimal delay stability.

[0096] Network hops is the number of intermediate forwarding devices, such as switches, routers or MEC access points, that data travels through from UPEN to the regional edge node (REN), and each additional hop introduces additional delay.

[0097] Switching delay refers to the cumulative delay of data packets in each intermediate network device due to queuing, buffer processing, scheduling decisions, etc., which can be collected through in-band network telemetry (INT).

[0098] The deployment cost function is used to weight and integrate the radio propagation characteristics, network hops and switching delay into a single optimization objective, and the optimization model takes the weighted sum of the network hops and the switching delay as the objective function, and the radio propagation characteristics being no less than the preset communication quality threshold as the constraint condition. The node deployment position combination that minimizes the deployment cost function value is found through an optimization algorithm, and the final determined specific installation coordinates of the master node and the shared backup node in the physical space are taken as the optimal deployment coordinate set.

[0099] In actual execution, the REN obtains candidate deployment points through factory digital twin models or field surveys, calculates the path loss, multipath delay spread and channel coherence time between each candidate point using the WinProp radio simulation tool or through Wi-Fi / 5G channel detection measurement data, determines the path and corresponding hops from any UPEN to the REN based on the network topology graph, and actively probes and measures the end-to-end switching delay of each path by sending timestamp probe packets. The REN inputs the radio propagation characteristics, network hops and switching delay into the preset deployment cost function, iterates through all feasible master / backup node combinations by running an integer linear programming optimization algorithm, calculates the total cost of each combination, and finally selects the combination with the minimum cost as the deployment scheme of the micro-edge cluster, and issues configuration instructions to related UPENs to complete role assignment and network binding.

[0100] For example, the deployment task-aware deployment cost function is: wherein, is the comprehensive deployment cost of the micro-edge cluster in the deployment position set , and the smaller the value, the better the deployment scheme, including the coordinates of each UPEN; used to represent the lower limit of wireless jitter; used to represent the end-to-end deterministic delay; is a task classification set; is the traffic weight of the kth level task in this area, satisfying obtained by REN through statistical history task distribution; is a normalized weighted coefficient, which can reflect the relative importance of wireless link quality and network transmission delay in the total cost, and satisfies ; is the set of deployment locations , the wireless path loss required to support the kth level task, obtained by ray tracing or measurement, the larger the value, the more serious the signal attenuation; is the reference path loss, for example, the free space path loss, used to normalize , eliminating the dimension effect; is the set of deployment locations , the number of network hops, i.e. the number of intermediate forwarding devices, that the data stream corresponding to the kth level task passes through from UPEN to REN; is the single-hop average switching delay, which is the average delay introduced by queuing and processing in a single switch / router, obtained by network probe measurement; is the set of deployment locations , the standard deviation of end-to-end delay jitter corresponding to the kth level task, used to reflect the delay uncertainty caused by wireless multipath and network congestion, which is a key indicator to measure control stability.

[0101] In this embodiment, by considering the computing load, the wireless channel physical characteristics and the wired network transmission performance are included in the unified evaluation framework, ensuring that the uplink between the master node and REN has low enough delay jitter and high reliability; In the hybrid wired-wireless deployment scenario, the control instruction delay caused by wireless multipath or switch congestion is avoided, the fault switching success rate is improved, the deployment service quality guarantee capability of the edge cluster in complex industrial environment is enhanced, and the transformation of micro-edge cluster deployment from experience site selection to performance driven is realized.

[0102] In some embodiments, the regional edge node is configured to analyze assistance intention information of each robot terminal in the robot queue, and in a case where it is detected that multiple robot terminals have a cooperative operation demand, group a super near-end edge node involved in the robot queue and a corresponding shared backup node thereof as a virtual joint node, the assistance intention information including task dependency relationship, spatial cooperation demand and peripheral device sharing request. All super near-end edge nodes in the virtual joint node share resource views and scheduling strategies, and are subjected to cooperative task allocation and load balancing by the regional edge node.

[0103] It can be understood that the robot queue is a list of tasks to be executed in the system in priority order, including requests of multiple robot terminals.

[0104] The assistance intention information is the cooperative demand data actively reported by the robot terminal, including task dependency relationship, space cooperation demand and peripheral sharing request.

[0105] The virtual joint node is a UPEN set logically grouped by REN, used to support collaborative tasks.

[0106] The resource view includes real-time resource status of UPEN, such as CPU usage, memory occupation, network bandwidth, etc.

[0107] The scheduling strategy is the task allocation rule, including priority, load threshold and other parameters.

[0108] The collaborative task allocation is to split and allocate multiple robot tasks to different UPENs; the load balancing is to dynamically adjust the workload of each UPEN to avoid overload.

[0109] In actual execution, REN periodically collects the assistance intention information of each robot terminal from the robot queue, receives JSON format data through gRPC protocol, and parses the task dependency relationship, space cooperation demand and peripheral sharing request therein; REN runs the collaboration detection module based on Drools rule engine, and when it detects that multiple robots have common needs, for example, two robots request space cooperation and peripheral sharing conflict at the same time, REN configures the master nodes and their corresponding shared backup nodes involved through logical group configuration in memory, and groups them into a virtual joint node by creating a virtual node ID; REN generates a global view after aggregating all UPENs in the virtual joint node, distributes a unified resource view and scheduling strategy, for example, a weighted allocation algorithm based on task urgency; REN uses a dynamic weighted round-robin load balancing algorithm based on current load to split the collaborative task into subtasks, allocate them to UPENs in the virtual joint node, and monitor load changes in real time. When the load of a UPEN exceeds the threshold, the allocation strategy is automatically adjusted to migrate part of the task to the backup node.

[0110] In addition, when task dependency conflict or space constraint overlap is detected, the conflict resolution module can be started to perform DAG topological sorting and calculate a safe cooperation path, and the super near edge nodes involved and their shared backup nodes are grouped into a virtual joint node through a hash table structure, and a real-time resource view is shared, including CPU utilization, memory occupation and network status. The regional edge node dynamically allocates collaborative tasks based on the real-time resource view, and triggers task migration mechanism when resource is overloaded, ensuring task execution continuity.

[0111] For example, the collaborative perception load indicator: wherein, a cooperative perception integrated load indicator of the jth UPEN, used for task scheduling within the virtual joint node; a CPU usage rate of the jth UPEN; a memory occupancy rate of the jth UPEN; a cooperative distance penalty, a cooperative robot terminal set of the jth UPEN based on assistance intention information; an Euclidean distance between the jth UPEN and the ith robot terminal, used to reflect physical cooperation closeness; weighting coefficients of resources, memory, and cooperative distance respectively, satisfying .

[0112] In this embodiment, the originally dispersed UPEN resources are logically integrated through the virtual joint node, reducing the cooperative task switching delay, improving the resource utilization and execution efficiency, and avoiding control interruption caused by single-point overload through dynamic load balancing; real-time analysis of peripheral sharing requests can support complex cooperation scenarios, realizing efficient and adaptive scheduling of multi-robot cooperative tasks.

[0113] In some embodiments, the task classification information includes high criticality tasks, medium criticality tasks, and low criticality tasks, The operating system kernel layer of the hyper-proximal edge node is configured to construct mutually isolated hardware resource slices according to the task classification information; the reflection control task in the high criticality task is allocated an exclusive CPU core, a continuous physical memory interval, and a last-level cache partition; the semantic feature extraction or data aggregation task in the medium criticality task and the low criticality task is allocated a limited computing resource slice.

[0114] It can be understood that the operating system kernel layer is a bottom module of a real-time kernel such as Linux PREEMPT_RT or Xenomai in the hyper-proximal edge node.

[0115] The hardware resource slice is a computing resource partition achieved through physical isolation. The resource slice is statically reserved at startup, and memory access and interrupt preemption across slices are prohibited during runtime to eliminate control channel latency jitter caused by resource contention.

[0116] The exclusive CPU core is a dedicated processor core allocated to the high criticality task, which is avoided through CPU affinity binding; The continuous physical memory interval refers to a memory block allocated continuously in DRAM, which is implemented through mmap in cooperation with HugeTLB page table.

[0117] The last-level cache partition is a dedicated area of the CPU L3 cache or similar cache, configured by Intel Cache Allocation Technology (CAT) or ARM Memory Partitioning and Monitoring (MPAM).

[0118] The reflection control task is a real-time control operation in a high-criticality task, for example, closed-loop feedback; The limited computing resource slice is a resource pool shared by medium-criticality tasks or low-criticality tasks, and the CPU usage rate and memory bandwidth are limited by Control Groups (CGroup).

[0119] In actual execution, when the UPEN is started, the task hierarchy information is received from the regional edge node, and the customized resource slice configuration module is loaded at the operating system kernel layer; for high-criticality tasks, the kernel calls the CPU affinity API, for example, sched_setaffinity, to bind the specified core to the task, and at the same time, the physical address continuous memory block is allocated by the mmap call, the HugeTLB page table is used to avoid page table rollover, and the cache partition is set by the Allocation Technology of Intel CAT; for medium-criticality tasks such as semantic feature extraction, the kernel creates a resource control group by CGroup v2 to limit the CPU usage rate and memory bandwidth; for low-criticality tasks such as data aggregation, a shared computing resource pool is allocated, and the bandwidth is controlled by the token bucket mechanism; during task execution, the kernel monitors the resource usage in real time, and if cache contention is detected for a high-criticality task, the cache partition is automatically adjusted; the entire configuration is loaded at startup by the kernel module such as kmod-cpu-alloc, without the need for application layer intervention.

[0120] In this embodiment, by exclusively using the CPU core, continuous memory, and cache partition for high-criticality tasks, the jitter caused by shared resource contention is eliminated, hardware-level resource isolation is achieved, the control instruction end-to-end latency jitter is reduced, and the security level is improved; medium-criticality tasks and low-criticality tasks run in the limited slice, avoiding interference with high-criticality tasks; the control reliability and resource utilization in the industrial robot scene are improved, providing physical protection for high-security and high-real-time control.

[0121] In some embodiments, the ultra-close edge node is further configured to extract a semantic feature vector by progressive semantic feature encoding; The progressive semantic feature encoding generates a hierarchical data structure including at least a first encoding layer and a second encoding layer. The first encoding layer is a key state code including a limited state identifier, used for transmission within a first time window, representing a security level of current operation and an outline state of the system. The second encoding layer is a complete semantic feature description, used for transmission within a second time window, including complete dimension and precision information of the semantic feature vector. The regional edge node sends a feature request instruction to the super near edge node according to the current network state and the emergency degree corresponding to the task classification information, to receive the first encoding layer or the first encoding layer and the second encoding layer.

[0122] It can be understood that the progressive semantic feature encoding is used for hierarchical data compression, and the semantic feature vector is split into two layers of key state and complete description; the first encoding layer is a key state code including a security level identifier and a system outline state, used for transmission within a very short time window, for example, in the security level, 0 = safe, 1 = warning, 2 = emergency, and the system outline state can be running / paused; the second encoding layer is a complete semantic feature description including all dimensions and precision information of the semantic feature vector.

[0123] The feature request instruction is a control command sent by the REN, which specifies the layer of encoding transmitted by the UPEN.

[0124] In the first encoding layer, a fixed 2-byte key state code is used to encode the security level and the system state by bit mask; the second encoding layer is a 128-dimensional floating point vector.

[0125] In actual execution, the UPEN runs the TensorRT-optimized lightweight MobileNetV3 in each control cycle to perform real-time inference on micro-sensing data and scene state data, and outputs a 128-dimensional semantic feature vector; the progressive semantic feature encoding module splits the vector into the first encoding layer and the second encoding layer; the UPEN transmits the first encoding layer to the shared memory control block in the exclusive time slot of the control channel through the time-gated scheduling mechanism; in the subsequent periodic time slot, the second encoding layer is written into the data block; the regional edge node monitors the network state and task classification information such as bandwidth utilization and end-to-end delay in real time, and since high-criticality tasks require priority response, when network congestion or task urgency is detected, a feature request instruction is sent to the UPEN, requiring only the first encoding layer to be transmitted; when the network is idle, the instruction requires the first encoding layer and the second encoding layer to be transmitted simultaneously; after receiving the instruction, the UPEN immediately performs the corresponding transmission operation in the specified time slot without additional data copying.

[0126] For example, the feature layer request probability is: wherein, a probability of requesting the regional edge node to receive the second encoding layer; a Sigmoid activation function, for mapping the input to the interval (0, 1); an urgency of a current task, obtained by task grading mapping; a current network normalized load; an urgency sensitivity coefficient, for amplifying the influence of ; a network load suppression coefficient, reflecting the restriction of bandwidth on the upload decision.

[0127] In this embodiment, by preferentially transmitting the key status code, the decision response time of the high criticality task is shortened, the second encoding layer is uploaded only when the bandwidth is sufficient, the occupation of redundant data is avoided, and the uplink traffic is reduced; the regional edge node adjusts the request strategy according to the network dynamics, ensures that the key information is delivered preferentially, and maintains the overall bandwidth efficiency of the system; the delay of task decision in the industrial robot control scene is reduced, and adaptive transmission of semantic information is realized.

[0128] The application also provides a real-time interaction method of a robot and an external device based on edge computing, applied to the real-time interaction system of the robot and the external device based on edge computing as described in any of the above embodiments.

[0129] As shown in Figure 2 , the real-time interaction method of the robot and the external device based on edge computing includes: Step 210, determining task grading information according to a semantic feature vector of global perception data through a regional edge node, making a global decision, and scheduling the shared backup node in a regional redundant resource pool, wherein the global perception data includes scene state data and micro-sensing data; Step 220, collecting the scene state data through a super-near edge node, and extracting the semantic feature vector from the micro-sensing data and the scene state data; Step 230, collecting the micro-sensing data through a robot terminal, and performing a task based on the task grading information.

[0130] According to the method for real-time interaction between the robot and the peripheral device based on edge computing provided by the embodiment of the application, through deep coupling of the super-near edge node and the robot terminal through shared memory zero-copy and time-gating scheduling, uncertain delay caused by a traditional network protocol stack is avoided, and deterministic execution of a high-critical task is ensured; the regional edge node performs task grading and global scheduling based on a semantic feature vector, and uses a shared backup node pool across clusters to realize elastic disaster recovery, thereby avoiding resource redundancy waste and ensuring system high availability; the system can adapt to environmental changes, reduce interaction delay in a complex wireless environment, ensure deterministic delay of a high-critical task, improve utilization efficiency of overall system resources, and improve real-time perception and control capability of the robot on the peripheral device.

[0131] The embodiments of the application are described above with reference to the drawings; however, the application is not limited to the specific implementation described above, which is merely illustrative rather than restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the application without departing from the purpose of the application and the scope protected by the claims, and all the forms belong to the protection of the application.

[0132] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an illustrative embodiment", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0133] Although the embodiments of the application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and purposes of the application, and the scope of the application is defined by the claims and their equivalents.

Claims

1. An edge computing based real-time interaction system of robot and peripheral, characterized in that, The system comprises at least one regional edge node, a plurality of super-close edge nodes and a plurality of robot terminals, each of the regional edge nodes is connected with a micro edge cluster, the micro edge cluster comprises a master node and at least one shared standby node, the master node and the shared standby node are divided based on task hierarchical information for the super-close edge nodes; The super-close edge nodes are connected based on shared memory and time-gated scheduling coupling; The regional edge nodes are deployed based on regional division of the robot terminals, used to determine the task hierarchical information according to a semantic feature vector of global perception data, make global decisions, and schedule the shared standby nodes in a regional redundant resource pool, the regional division is determined based on task throughput, data aggregation demand and energy consumption constraints, and the global perception data comprises scene state data and micro-sensing data; The super-close edge nodes are deployed based on task delay limit and communication reliability demand, used to collect the scene state data, and extract the semantic feature vector from the micro-sensing data and the scene state data; The robot terminals are used to collect the micro-sensing data, and perform tasks based on the task hierarchical information.

2. The edge computing based real-time interaction system of robots and peripherals according to claim 1, characterized in that, Control channels and data synchronization channels are configured between the robot terminals and the super-close edge nodes; A gated shared memory manager is configured on the super-close edge node, used to open access rights of a mapped logical block in an activated time slot according to a time slot trigger signal generated by a time-gated scheduling mechanism, so that an authorized robot terminal or super-close edge node performs zero-copy read and write on a corresponding logical block in a shared memory region; The time-gated scheduling mechanism is used to divide a communication period into exclusive time slots, periodic time slots and dynamic time slots in time sequence, allocates exclusive time slots to the control channels, and allocates corresponding periodic time slots or dynamic time slots to the data synchronization channels, the priority of the exclusive time slots is higher than that of the periodic time slots and the dynamic time slots; The shared memory region is divided into a plurality of logical blocks, the logical blocks comprise control blocks and at least one data block; The control blocks are used to map the control channels, so that the super-close edge nodes write the task hierarchical information and control instructions through kernel-mode memory mapping, and the robot terminals read; The data blocks are used to map the data synchronization channels, so that the robot terminals write the micro-sensing data, and the super-close edge nodes read.

3. The edge computing based real-time interaction system of robots and peripherals according to claim 2, wherein, The super-close edge nodes are used to adjust the period and duty cycle of the time gates of the control channels and the data synchronization channels according to the task hierarchical information issued by the regional edge nodes, issue control instructions containing task hierarchical information to the robot terminals through the control blocks in the exclusive time slots of the control channels, allocate data synchronization channels with fixed positions and redundant periods to data streams with high criticality of the task hierarchical information, and the mapped data blocks are continuous on the physical memory. The robot terminal is configured to receive the control instruction and respond by synchronizing the micro-sensing data in the mapped data block in an authorized and activated time slot according to the task hierarchy information through zero-copy; The exclusive time slot of the control channel is located at the start of each communication cycle, which is synchronized with the control decision cycle of the robot terminal.

4. The edge computing based real-time interaction system of robots and peripherals according to claim 2, wherein, The ultra-near edge node is integrated with a joint arbitrator; The joint arbitrator is configured to adjust the suppression intensity of the data transmission bandwidth of the data synchronization channel according to a deviation value between the performance index of the control channel and a service level target corresponding to the task hierarchy information, when the deviation between the end-to-end delay of the control channel and the upper limit of tolerance corresponding to the task hierarchy information is less than a preset safety margin, wherein the deviation value is negatively correlated with the suppression intensity; The joint arbitrator is configured with a back pressure feedback unit, which is configured to send a flow control signal to the robot terminal to suspend or reduce the rate of writing micro-sensing data to the data block by the robot terminal; The ultra-near edge node is further configured to release the bandwidth suppression of the data synchronization channel and restore the transmission rate of the data synchronization channel when the end-to-end delay of the control channel continues to be lower than the safety recovery threshold for a preset stable duration.

5. The edge computing based real-time interaction system of robots and peripherals according to claim 1, wherein, The ultra-near edge node is configured to: perform online differential encoding on the semantic feature vector, generate change summary data based on the time sequence difference between the semantic feature vector and the historical semantic feature vector; trigger asynchronous uploading of the change summary data to the regional edge node when the time sequence difference exceeds a difference threshold, wherein the change summary data is compressed in a differential encoding format, including the semantic dimension that has changed and the incremental value of the semantic dimension, for internal state log recording and anomaly detection; The regional edge node is configured to receive the change summary data to update the global perception data.

6. The edge computing based real-time interaction system of robots and peripherals according to claim 1, wherein, The deployment of the master node and the shared standby node in the micro-edge cluster is determined based on joint optimization of radio frequency propagation characteristics, network hop count and switching delay; The radio frequency propagation characteristics include path loss, multipath delay spread and channel coherence time, which are used to evaluate the lower limit of the delay jitter of the wireless link; The network hop count is used to represent the number of intermediate forwarding devices from the ultra-near edge node to the regional edge node; The switching delay is used to represent the transmission delay introduced by the queuing and processing of network intermediate nodes in the edge network path from the ultra-near edge node to the regional edge node; The regional edge node is configured to construct a deployment cost function based on the radio frequency propagation characteristics, the network hop count and the switching delay to determine the optimal deployment coordinate set of the micro-edge cluster.

7. The edge computing based robot and peripheral real-time interaction system according to claim 1, wherein, The region edge node is used to parse the assistance intention information of each robot terminal in the robot queue. When multiple robot terminals are detected to have collaborative operation needs, the ultra-near-end edge node involving the robot queue and its corresponding shared backup node are grouped into a virtual joint node. The assistance intention information includes task dependency relationship, spatial collaboration needs and peripheral sharing requests. All ultra-near-end edge nodes within the virtual federated node share resource views and scheduling policies, and the regional edge nodes coordinate task allocation and load balancing.

8. The edge computing based real-time interaction system of robots and peripherals according to claim 1, wherein, The task classification information includes high-criticality tasks, medium-criticality tasks, and low-criticality tasks. The operating system kernel layer of the ultra-near-end edge node is configured to construct mutually isolated hardware resource slices based on task classification information; The reflection control task in a highly critical task is allocated a dedicated CPU core, a contiguous physical memory region, and a last-level cache partition; Slices of limited computing resources are allocated to semantic feature extraction or data aggregation tasks in medium-critical and low-critical tasks.

9. The edge computing based robot and peripheral real-time interaction system according to claim 1, wherein, The ultra-near-end edge node is further used to extract semantic feature vectors through progressive semantic feature encoding; The progressive semantic feature encoding generates a hierarchical data structure that includes at least a first encoding layer and a second encoding layer; The first encoding layer is a critical status code, including a limited number of status identifiers, which are used for transmission within a first time window to characterize the security level of the current operation and the approximate state of the system. The second encoding layer is a complete semantic feature description, used for transmission within the second time window, including the complete dimension and precision information of the semantic feature vector; The regional edge node sends a feature request instruction to the ultra-near-end edge node according to the current network status and the urgency level corresponding to the task classification information, so as to receive the first coding layer, or the first coding layer and the second coding layer.

10. A real-time interaction method between a robot and peripheral devices based on edge computing, characterized in that, The method, applied to a real-time interaction system between a robot and a peripheral device based on edge computing as described in any one of claims 1-9, comprises: By using regional edge nodes, task classification information is determined based on the semantic feature vector of the global perception data, global decisions are made, and the shared backup nodes in the regional redundant resource pool are scheduled. The global perception data includes scene status data and micro-sensor data. The scene state data is collected through ultra-near-end edge nodes, and the semantic feature vector is extracted from the micro-sensing data and the scene state data; The robot terminal collects the microsensor data and performs task execution based on the task classification information.

Images