Body-composite robot shared perception edge collaborative control method and system

Abstract

The application discloses a body-combined robot shared perception edge collaborative control method and system, and relates to the technical fields of robot collaborative control and edge computing. The method collects ontology state, visual data and task instructions based on a unified time baseline, unloads visual subtasks according to time delay budget and residual computing power of an edge node; the edge node generates a perception result package and writes the package into a shared perception space; a controller performs total time delay calculation, time delay compensation, space-time correlation and consistency verification on the result package, determines shared perception credibility, and only uses a result meeting the time delay budget and having credibility reaching an access threshold for rolling optimization solution; when the credibility is lower than a degradation threshold or a link time delay continuously exceeds a limit, a safe degradation control mode is switched to.

Description

Technical Field

[0001] This invention relates to the field of robot collaborative control and edge computing technology, and more specifically, to a collaborative control method and system for a hybrid robot with shared sensing edge. Background Technology

[0002] Embossed robots typically consist of a mobile chassis, a robotic arm, an end effector, vision sensors, and a vision actuator for adjusting the field of view or posture. They can perform composite tasks that involve moving and working simultaneously in scenarios such as inspection, grasping, handling, assembly, and operation assistance. During actual operation, these robots not only need to maintain the motion control precision of the mobile chassis and robotic arm, but also need to continuously acquire information about the target object's state, obstacle distribution, environmental semantics, and operational context to form a stable closed loop of perception, decision-making, and execution.

[0003] In existing technologies, some solutions deploy visual tasks such as target detection, target tracking, pose estimation, and depth estimation to be performed on the robot itself. While this type of solution has a relatively direct processing chain, it places high demands on the robot's computing resources, power supply, and heat dissipation capabilities. In continuous operation, frequent occlusion, or high-frequency control scenarios, it can easily lead to increased processing latency, control cycle jitter, and higher overall machine costs.

[0004] Some solutions attempt to offload some vision tasks to the edge to reduce the burden on the robot itself. However, existing edge offloading solutions mostly use edge nodes as independent inference resources. The visual results returned from the edge are usually used as auxiliary display information or reference information for higher-level decisions, lacking a unified time reference that matches the robot's current control cycle, making it difficult to stably enter the motion control closed loop. When there is jitter in network transmission, the acquisition time, inference completion time, and control solution time are disconnected, which can easily lead to outdated or mismatched shared results, or inability to stably participate in control state updates.

[0005] In multi-robot collaborative operation scenarios, neighboring robots can provide target observation results and environmental constraint information from different perspectives, which can theoretically alleviate problems such as occlusion, blind spots, and lack of local perception. However, in current technologies, information sharing among multiple robots often remains at the level of result forwarding, simple stitching, or static map sharing, lacking spatiotemporal correlation mechanisms around target identification, spatial location, and time windows, as well as consistency verification and credibility assessment mechanisms for observation results from multiple robots. As a result, different observation results of the same target object may be duplicated, conflicting, or erroneously reused, making it difficult to directly use as closed-loop control input.

[0006] Furthermore, existing technologies typically handle edge resource scheduling and control problem-solving separately. Whether the resource is unloaded at the edge, to which node it is unloaded, and how much resource it consumes are often disconnected from the robot's current control objectives, latency budget, and safety constraints. There is also a lack of clear admission criteria and degradation switching mechanisms for whether shared results can be directly incorporated into control. Therefore, existing technologies struggle to simultaneously meet the following requirements: first, ensuring time consistency between shared perception results and the control cycle under edge computing conditions; second, improving the availability and reliability of perception results in multi-robot shared scenarios; and third, maintaining real-time motion control and system security during network state and edge load fluctuations.

[0007] Therefore, it is necessary to provide a collaborative control method and system for shared perception edges of embodied composite robots to solve the problems in the existing technology, such as the difficulty in stably entering the control closed loop with shared perception results, insufficient coupling between edge unloading and control solution, and insufficient safety control capability under network fluctuation conditions. Summary of the Invention

[0008] The purpose of this invention is to provide a collaborative control method and system for shared sensing edge of an embodied composite robot, in order to solve the problems in the prior art such as inconsistent edge sensing results and control cycle time reference, difficulty in directly reusing shared sensing results of multiple robots in control closed loop, decoupling of edge computing power scheduling and control, and lack of stable degradation mechanism when shared sensing is abnormal.

[0009] To achieve the above objectives, this invention provides a collaborative control method for embodied composite robots with shared sensing edges, applied to a collaborative control system including the current embodied composite robot, at least one neighboring embodied composite robot, at least one edge node, a shared sensing communication layer, and a shared sensing space, comprising the following steps: Within the current control cycle, the body state, visual data, and task instructions of the current embodied composite robot are collected based on a unified time baseline. The body state includes at least the pose and velocity of the mobile chassis, the position and velocity of the robotic arm joints, the state of the end effector, and the posture of the visual actuator.

[0010] Based on the task instructions, a visual subtask is determined for updating target state information and / or environmental constraint information. Then, based on the current control cycle delay budget, visual subtask priority, current network transmission delay, and remaining computing power of the edge node, the visual subtask is offloaded to at least one edge node for execution.

[0011] The edge node performs the visual subtask on the visual data from the current embodied composite robot and / or nearby embodied composite robots, generates a perception result package, and writes it into the shared perception space via the shared perception communication layer. The perception result package includes at least the target identifier, acquisition timestamp, result generation timestamp, target state information, environmental constraint information, and result confidence level.

[0012] The system reads the perception result package corresponding to the current control task from the shared perception space, calculates the total perception link delay of each perception result package based on a unified time baseline, and performs delay compensation for the target state information by combining the motion state of the mobile chassis, the motion state of the robotic arm, and the attitude change of the vision actuator within the current control cycle.

[0013] For multiple perception result packages belonging to the same target object, spatiotemporal correlation and consistency verification are performed, and the shared perception credibility is determined based on the consistency verification results, result confidence, and total perception link latency.

[0014] Only perception results that meet the following conditions are determined as valid shared perception results: the total latency of the perception link is not greater than the current control cycle latency budget, and the confidence level of the shared perception is not lower than the shared perception control admission threshold T1.

[0015] Based on the effective sharing of perception results, ontology state, and motion execution target, a motion vision shared state is constructed, and the control quantities of the mobile chassis, robotic arm, and vision actuator are uniformly solved in a rolling optimization problem that includes edge resource occupancy constraints.

[0016] When the shared perception confidence level is lower than the shared perception control admission threshold T1 but not lower than the shared perception degradation threshold T2, the weight of the corresponding perception result in state fusion is reduced. When the shared perception confidence level is lower than the shared perception degradation threshold T2, or when the total delay of the perception link exceeds the current control cycle delay budget for multiple consecutive preset control cycles, the system switches to the safety degradation control mode and outputs a limited safety control command based on the most recent effective motion vision shared state and the remaining perception information of the body. The shared perception control admission threshold T1 is greater than the shared perception degradation threshold T2.

[0017] Furthermore, the unified time baseline is preferably established by the robot body control unit, vision sensors, and edge nodes through at least one of the following methods: local master clock synchronization, network time synchronization, or time-sensitive network synchronization, in order to ensure the consistency of the acquisition timestamp, the result generation timestamp, and the control cycle time reference.

[0018] Furthermore, the perception result packages in the shared perception space are preferably indexed and stored according to the target identifier, the spatial index of the task coordinate system, and the time window, and historical results that exceed the time delay budget are invalidated and eliminated to reduce the probability of stale results entering the control closed loop.

[0019] Furthermore, the spatiotemporal correlation and consistency verification preferably screens candidate observation results based on target identification, time window and spatial proximity, and performs weight reduction processing and confidence calculation on multi-source observation results by calculating at least one consistency index among target position deviation, target attitude deviation and result confidence difference.

[0020] Furthermore, the visual subtask unloading is preferably performed based on the scheduling cost to select edge nodes and allocate computing power slices. The scheduling cost takes into account the visual subtask priority, expected computational load, latency exceedance, edge node load, and remaining computing power ratio, and uses the slice allocation result as the basis for edge resource occupancy constraints in subsequent rolling optimization.

[0021] Furthermore, the rolling optimization problem includes at least a motion tracking objective, a visual observability objective, an obstacle safety constraint, a control stability objective, and an edge resource occupancy constraint. The obstacle safety constraint is determined at least in part based on environmental constraint information, and the edge resource occupancy constraint is determined at least in part based on the edge computing power slice allocation results corresponding to the visual subtask.

[0022] Furthermore, the safety degradation control mode preferably executes at least one of the following strategies: reducing the upper limit of the mobile chassis speed, reducing the upper limit of the robotic arm joint speed or joint torque, freezing the high-speed rotation command of the vision actuator, switching to a short-term holding strategy based on the most recent effective target pose, switching to a local obstacle avoidance strategy based on the body's near-field sensor, and pausing the addition of unloading requests for low-priority vision subtasks.

[0023] This invention establishes a unified time baseline and performs total latency calculation and latency compensation on the shared sensing results, enabling the sensing results returned by the edge nodes to participate in the state update of the current control cycle according to the unified time reference, thereby improving the time consistency and real-time availability of the shared sensing results entering the control closed loop.

[0024] This invention standardizes and organizes the perception results of multiple robots through a shared perception space, and avoids duplicate, conflicting or low-reliability results from different sources from directly entering the control closed loop through spatiotemporal correlation, consistency verification and shared perception reliability assessment, thereby improving the reuse effectiveness and control security of shared perception results.

[0025] This invention uses shared sensing reliability and total sensing link delay as control admission conditions, allowing only sensing results that meet the delay budget and whose reliability reaches the shared sensing control admission threshold to participate in rolling optimization, thereby reducing the impact of outdated or low-reliability sensing results on control accuracy and system stability.

[0026] This invention introduces the edge computing power slicing allocation result into the edge resource occupancy constraint term in the rolling optimization problem, realizing the synergistic coupling of edge resource scheduling and robot motion control, which is beneficial to maintaining the real-time performance and control solution stability of key vision tasks under the conditions of edge load fluctuation and network state changes.

[0027] This invention improves the system's fault tolerance and operational security under conditions of shared sensing quality fluctuations, network congestion, or edge node anomalies by setting a shared sensing control admission threshold T1 and a shared sensing degradation threshold T2, performing weighted fusion within the threshold range, and switching to a safe degradation control mode when the threshold is below the degradation threshold or the latency exceeds the limit continuously.

[0028] This invention is applicable to scenarios such as simultaneous movement and grasping, multi-robot collaborative inspection, dynamic obstacle collaborative obstacle avoidance, and edge load fluctuation. It can reduce the computational burden on the robot itself while improving the utilization efficiency of nearby robots sharing perception results into the control closed loop.

[0029] Furthermore, this invention does not simply concatenate edge offloading, shared sensing, and rolling optimization. Instead, it maps the acquisition time, edge result generation time, shared sensing space read / write time, and control solution time to the same time reference through a unified time baseline. Then, it uses the total latency of the sensing link and the confidence level of shared sensing to jointly limit the result admission conditions. This ensures that whether a shared sensing result can enter the control closed loop no longer depends solely on a single identification confidence level, but rather on the joint satisfaction of time validity and multi-source consistency. This joint constraint mechanism can directly suppress the interference of stale, mismatched, and conflicting results on closed-loop control. Its technical effect cannot be naturally obtained by implementing traditional edge offloading or traditional result fusion individually.

[0030] Furthermore, this invention feeds back the edge offloading results of the visual subtasks to the rolling optimization solution process in the form of edge resource occupancy constraints, enabling the perception scheduling decision and the control output decision to share the same solution objective and the same budget boundary. Compared with the method of independently completing edge task allocation outside the controller and then passively sending the results to the controller, this invention can actively adjust the balance between the complexity of the perception link and the control solution pressure when edge load fluctuations and network jitter exist, thereby maintaining a high shared perception availability rate while ensuring timely output of control commands, and thus has stronger closed-loop real-time stability. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a diagram illustrating the overall system architecture of an embodiment of the present invention.

[0033] Figure 2 This is a flowchart of the edge computing power slice scheduling process according to an embodiment of the present invention.

[0034] Figure 3 This is a schematic diagram of the shared sensing space data flow according to an embodiment of the present invention.

[0035] Figure 4 This is a schematic diagram of motion vision sharing state fusion according to an embodiment of the present invention.

[0036] Figure 5 This is a flowchart of motion vision collaborative scrolling optimization according to an embodiment of the present invention.

[0037] Figure 6 This is a flowchart illustrating the safety degradation control process according to an embodiment of the present invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the embodiments shown in the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various changes and substitutions can be made without departing from the concept of the present invention, and all such changes and substitutions should fall within the scope of protection of the present invention.

[0039] I. System Composition and Unified Time Baseline In one specific embodiment, the embodied composite robot shared perception edge collaborative control system includes a current embodied composite robot, at least one edge node, a shared perception communication layer, a shared perception space, and a control processor; when the system is in a multi-robot collaborative mode, it also includes at least one nearby embodied composite robot. The current embodied composite robot includes a mobile chassis, a robotic arm, an end effector, a vision sensor, and a vision execution mechanism. The control processor can be deployed in the robot body control unit, the edge node, or a collaborative computing unit of both.

[0040] The unified time baseline is used to unify the time reference across the robot body, vision sensors, edge nodes, and shared sensing space. Specifically, the unified time baseline can be established using at least one of the following methods: local master clock synchronization, network time synchronization, or time-sensitive network synchronization. In one specific embodiment, a time-sensitive network synchronization method can be used to establish the unified time baseline: the robot body control unit, vision sensors, and edge nodes are clock-aligned using a time synchronization mechanism supporting the IEEE 802.1AS protocol, ensuring that the deviation between the local clock of each node and the unified time reference is controlled within a preset error range, such as the microsecond range. After synchronization is completed, each node generates a corresponding timestamp based on the unified time reference when collecting, generating, writing, and reading data related to shared sensing. After the unified time baseline is established, the acquisition timestamp, result generation timestamp, result writing time, and result reading time within the current control cycle are all mapped to the same time reference to eliminate time offsets caused by multi-source asynchronous acquisition.

[0041] In a specific embodiment, let the starting point of the current control cycle be t_k, the acquisition timestamp in the sensing result packet i be t_cap,i, the result generation timestamp be t_gen,i, and the result reading time be t_read,i. Then, the total latency τ_i of the sensing link can be understood as the total time elapsed from the acquisition and generation of the sensing result to its entry into the control closed loop, which can be schematically represented as follows: τ_i = (t_gen,i - t_cap,i) + (t_read,i - t_gen,i) The first term corresponds to the edge-side processing time, and the second term corresponds to the transmission and queuing time introduced by the shared sensing communication layer and the shared sensing space. If there is a residual synchronization error ε_sync in the unified time baseline, it can be included as a correction factor in the total delay assessment to improve the consistency of delay judgment.

[0042] II. Perception Result Package and Delay Compensation In one specific embodiment, the edge node performs a visual subtask on visual data from the current embodied composite robot and / or neighboring embodied composite robots to obtain a perception result package. The perception result package includes at least a target identifier, acquisition timestamp, result generation timestamp, target state information, environmental constraint information, and result confidence level.

[0043] The target state information may include at least one of the following: target position, target attitude, target velocity, target category, or target trajectory segment; the environmental constraint information may include at least one of the following: obstacle location, passable area, hazardous area, local geometric boundary, or restricted area for operation; the result confidence level is used to characterize the degree of availability of the current perception result package for subsequent control loops.

[0044] In one specific embodiment, the time delay compensation for the target state information preferably adopts a short-time extrapolation method. If the target state vector is denoted as x_i(t), it can be predicted and updated according to the compensation time interval Δτ_i, which can be schematically represented as follows: x_i(t_k) = f(x_i(t_cap,i), Δτ_i) Where f(·) represents the prediction function that matches the target motion characteristics. For translational targets, uniform velocity approximation or other short-time prediction methods can be used for compensation; for the posture-related information of the embodied composite robot itself, correction can be made by combining the speed of the moving chassis, the speed of the robotic arm joints, and the posture changes of the vision actuator to reduce the timing deviation within the control cycle.

[0045] III. Spatiotemporal correlation, consistency verification, and shared perception credibility In one specific embodiment, when there are multiple perception result packets for the same target object in the shared perception space, the control processor first filters candidate observation results based on target identifier, time window and spatial proximity relationship, and then performs spatiotemporal correlation and consistency verification.

[0046] Preferably, the target positions from different sources can be first mapped uniformly to the task coordinate system, and then a consistency index S can be constructed based on the target position deviation, target attitude deviation, and result confidence differences to comprehensively characterize the degree of coordination of multi-source observation results. For example: S = exp(-α|Δp| - β|Δθ| - γ|Δc|) Here, α, β, and γ are weighting coefficients used to balance the effects of position bias, attitude bias, and confidence level differences. The larger the S value, the more consistent the observation results from multiple sources.

[0047] In one specific embodiment, the shared perception confidence level C can be jointly determined by the result confidence level, consistency index, and total latency of the perception link, and is used to simultaneously reflect the result quality, consistency, and timeliness, for example: C = w1× c + w2 × S + w3×exp(-λ×τ) Where c is the average result confidence of the candidate sensing result package, S is the average consistency index, τ is the average total latency of the sensing link, and λ is the latency attenuation coefficient, used to characterize the attenuation strength of the total latency of the sensing link on the shared sensing confidence. The larger the value of λ, the more sensitive the system is to the increase in latency, and the stronger the inhibitory effect of the latency term on confidence. w1, w2, and w3 are weighting coefficients, and satisfy w1 + w2 + w3 = 1. When C is high, it indicates that the shared sensing result has high spatiotemporal consistency, confidence, and real-time performance, and can be used for control closed loop; when C decreases, it preferentially enters the weighted fusion or security degradation control mode.

[0048] 4. Threshold Setting and Access Rules In a specific embodiment, the shared perception control access threshold T1 and the shared perception degradation threshold T2 satisfy T1>T2. Wherein, T1 is used to determine whether the shared perception result can directly enter the rolling optimization closed-loop, and T2 is used to determine whether the shared perception result can only be used for weight reduction fusion or should be directly switched to the safety degradation control mode.

[0049] Preferably, T2 can be preset according to the system safety requirements, task risk level and historical control stability test results; T1 can be obtained by adding an access margin on the basis of T2, for example: T1 can be determined by adding an access margin on the basis of T2.

[0050] The access margin can be adjusted according to the real-time requirements of the control task, the edge load level and the network status. For high-risk tasks, the access margin can be appropriately increased to improve the quality of shared perception results entering the control closed-loop; for low-risk tasks, the access margin can be appropriately reduced to balance both real-time performance and computing power utilization.

[0051] In a specific embodiment, if the shared perception reliability C satisfies C ≥ T1 and the total time delay of the perception link τ_i is not greater than the current control cycle time delay budget B_k, the perception result is determined as a valid shared perception result; if T2 ≤ C<T1, the perception result is retained but its weight in state fusion is reduced; if C<T2, the perception result no longer participates in closed-loop control, and the safety degradation control mode is triggered when necessary.

[0052] 5. State Fusion and Rolling Optimization In a specific embodiment, the motion vision shared state z_k is constructed based on the valid shared perception result, the ontology state and the motion execution target, and the moving chassis control quantity, the manipulator control quantity and the vision actuator control quantity are uniformly solved in the rolling optimization problem including the edge resource occupancy constraint term. The rolling optimization problem can be understood as comprehensively balancing the control effect under the premise of satisfying the safety constraint and the resource constraint, and its objective function can be schematically expressed as: min J = J_track + J_obs + J_smooth + J_edge Wherein, J_track is the motion tracking objective term, which is used to reduce the error of the robot relative to the task target; J_obs is the visual observability objective term, which is used to maintain target visibility and perception effectiveness; J_smooth is the control smoothness objective term, which is used to reduce control jitter; J_edge is the edge resource occupancy constraint term, which is used to limit edge node overload.

[0053] In a specific embodiment, the obstacle safety constraint item is determined at least partially based on the environmental constraint information, for example, limiting the minimum distance between the mobile chassis and an obstacle, limiting the contact probability between the end of the robotic arm and a dangerous area, and limiting the duration of the blind area of the visual actuator during high-speed rotation. The edge resource occupation constraint item is determined at least partially based on the edge computing slice allocation result corresponding to the visual subtask, and is used to ensure that the edge node does not break through the resource upper limit while completing the visual subtask.

[0054] VI. Weight Reduction Fusion and Safety Degradation Control In a specific embodiment, when the shared perception credibility C is lower than T1 but not lower than T2, the control processor reduces the weight of the corresponding perception result in state fusion, so that the high-credibility result occupies a relatively higher position in state estimation. Preferably, the fusion weight ω_i can be positively correlated with the credibility, for example: ω_i = C_i / ΣC_i Alternatively, a weight update method based on threshold segmentation is adopted, so that high-credibility results occupy a higher weight in state estimation, and low-credibility results are only used as auxiliary references.

[0055] In a specific embodiment, when C<T2, or the total time delay of the perception link exceeds the limit for consecutive multiple preset control cycles, the system switches to the safety degradation control mode. This mode may include at least one or more of the following strategies: reducing the upper speed limit of the mobile chassis, reducing the upper joint speed limit or upper joint torque limit of the robotic arm, freezing high-speed rotation commands of the visual actuator, switching to a short-term maintenance strategy, switching to a local obstacle avoidance strategy, and suspending new offloading requests for low-priority visual subtasks.

[0056] In the safety degradation control mode, the system preferably outputs the amplitude-limited safety control instruction based on the last effective motion vision shared state and the remaining ontology perception information. In this way, even if the shared perception quality decreases in a short period, the system can still maintain basic motion control and obstacle avoidance capabilities under low risk.

[0057] VII. Parameter Configuration and Admission Threshold Setting In one specific implementation, before the task begins, the controller jointly determines the control cycle delay budget B_k, the shared perception control admission threshold T1, and the shared perception degradation threshold T2 for the current task based on the task risk level, environmental dynamism, network jitter statistics, and historical load of edge nodes. For tasks with dense dynamic obstacles, narrow grasping windows, or small end-effector attitude error tolerance, T1 can be increased and the allowable delay margin reduced to prioritize ensuring higher reliability of shared perception results entering the control closed loop. For tasks primarily involving inspection, long-range pre-aiming, or relaxed path tracking, T1 can be appropriately reduced and the delay margin relaxed to improve edge resource utilization and multi-robot observation reuse rate. By binding the threshold settings to task risk, the admission rules can be aligned with the control objectives, rather than using fixed empirical thresholds.

[0058] Furthermore, the threshold tuning can be linked with the granularity of edge node computing power slices and task priorities. If the current task requires simultaneous target detection, pose estimation, and depth estimation, the controller can allocate smaller but more stable computing power slices to high-priority subtasks based on the computational load and deadline of different visual subtasks, while reserving reclaimable resources for low-priority subtasks. When the network condition deteriorates or the load on edge nodes suddenly increases, the controller prioritizes maintaining the target state update link required for dominant control, while postponing the execution of subtasks such as environmental semantic supplementation and long-distance non-critical target recognition. This allows for an interpretable engineering trade-off between shared perception quality and resource consumption.

[0059] VIII. Implementation Methods for Shared Perception Conflict Resolution and Role Switching In one specific implementation, when multiple neighboring robots simultaneously provide observation results of the same target object, the controller, in addition to constructing a consistency index based on position deviation, posture deviation, and result confidence, can also introduce an observation source stability factor, a target tracking continuity factor, and a viewpoint effectiveness factor to perform a secondary ranking of results from different sources. Specifically, for sources that stably track the same target for multiple consecutive control cycles and have a low occlusion rate, a higher source stability weight can be assigned; for sources that only appear for a short time in a single cycle and have large timestamp jitter, their shared perception confidence contribution value can be reduced. This strategy helps avoid target misassociation or abrupt changes in control input caused by short-term cross-observations from multiple robots.

[0060] Furthermore, the system can also be configured with a master-slave observation role switching mechanism. When visual observability of the current embodied composite robot decreases due to field-of-view occlusion, strong glare, or posture constraints between the robot and the target object, a nearby robot with a better lateral view can be temporarily promoted to the master observation source. Once the robot regains observability and outputs stable results for several consecutive control cycles, the shared perception mode, dominated by the robot and supplemented by nearby robots, is gradually restored. In one embodiment, the role switching process is constrained by lease duration, heartbeat confirmation, and result sequence number to prevent multiple robots from repeatedly competing for the master observation role under network jitter, thereby ensuring a smooth transition of the control closed-loop input source.

[0061] IX. Deployment Verification, Log Review and Recovery Mechanism In one specific implementation, a unified interface verification checklist can be established for different robot platforms during the system deployment phase. This checklist should include at least: a unified time baseline synchronization error upper limit, average write latency from edge nodes to the shared sensing space, average read latency from the shared sensing space to the controller, visual actuator posture feedback refresh rate, near-field obstacle avoidance sensor refresh rate, and the maximum allowable duration for rolling optimization solutions. Before formal operation, the controller sequentially performs static verification, idle link verification, and low-speed closed-loop integration testing. If any key indicator exceeds the limit, the shared sensing results are temporarily not allowed to enter the main control closed loop, and the system will only operate in ontology verification mode. By establishing quantitative acceptance conditions before deployment, the risk of link bottlenecks being exposed only after deployment can be reduced.

[0062] Furthermore, during operation, the controller preferably writes the latency budget usage, shared perception reliability, source of adopted results, edge node slice allocation, degradation triggering reasons, and recovery confirmation results for each control cycle into the operation log. During the offline analysis phase, the frequency of timeout events, reliability fluctuations, role switching times, and degradation duration can be statistically analyzed by task number, target type, environment area number, and edge node number. Based on this, the threshold template, resource slicing strategy, and abnormal recovery hysteresis parameters for the next round of tasks can be corrected. By combining online control with offline review in a closed loop, the parameter adaptation efficiency and on-site deployment stability under similar tasks can be gradually improved.

[0063] In one specific implementation, the exit from the safety degradation control mode does not employ a strategy of immediate switchback after a single threshold recovery. Instead, a recovery confirmation window is introduced. That is, when the shared perception confidence level rises above the degradation threshold T2 again, the total latency of the perception link returns to within the budget, and the body's near-field sensors do not detect any new high-risk obstacles, the controller still needs to verify, over several consecutive control cycles, that the continuity of target state updates, the stability of observation sources, and the rolling optimization solution time have all returned to normal. Only then does it gradually restore high-speed vision actuator commands, edge high-priority unloading requests, and the normal speed limit. This delayed exit mechanism effectively avoids the system frequently switching between normal and degradation modes under abnormal edge conditions.

[0064] Furthermore, to facilitate subsequent review and engineering implementation, the unified time baseline, shared perception space, admission threshold, degradation threshold, slice allocation results, and role switching mechanism involved in the embodiments of the specification can all be pre-mapped as parameter templates or configuration entries, allowing for limited online updates for different scenarios without changing the core control flow. This demonstrates both the cross-platform deployment capability of the invention and its engineering feasibility under conditions of network fluctuations, multi-robot sharing, and varying edge loads.

[0065] 10. Parameter Templates, Comparison Verification, and Supplementary Implementation Methods for Extended Deployment In one specific implementation, a correspondence between task types and parameter templates can be pre-established during system deployment. The parameter templates include at least: initial values ​​for control cycle delay budget, shared perception control admission threshold T1, shared perception degradation threshold T2, result consistency index weight, observation source stability weight, maximum slice occupancy rate of edge nodes, rolling optimization solution time domain length, chassis speed upper limit, robotic arm joint speed upper limit, and degradation exit confirmation window length. For tasks involving simultaneous movement and grasping, the parameter templates prioritize target state update frequency and end-effector pose stability; for multi-robot collaborative inspection tasks, the parameter templates prioritize shared observation coverage and edge resource reuse efficiency; for dynamic obstacle collaborative obstacle avoidance tasks, the parameter templates prioritize improving near-field constraint response speed and degradation switching sensitivity. Template-based configuration reduces on-site debugging complexity and improves the ability to quickly switch between different tasks.

[0066] Furthermore, the parameter template can be divided into two layers: a static template and an online correction template. The static template is calibrated during the offline testing phase based on the robot platform's capabilities, network device performance, and typical operating scenarios, and is used to define the basic range of key thresholds and constraints. The online correction template performs minor adjustments to some parameters in the static template based on the network jitter distribution, edge node busyness, observation occlusion rate, and local environmental risk level in the current task. Preferably, online correction is only allowed within the reviewed and defined safety boundaries to avoid threshold drift or safety policy failure due to over-adaptation of the controller in complex scenarios. This maintains the flexibility of parameter adjustment while ensuring that the control closed loop has interpretable behavioral boundaries.

[0067] In one specific embodiment, to verify the actual effect of the present invention, three sets of comparative experiments can be performed according to the same task script, the same target object, and the same ground conditions: one with shared perception access control disabled, one with edge offloading enabled but without access verification, and one with the complete control closed loop of the present invention enabled. Evaluation indicators include at least: target state update latency, effective shared perception result adoption rate, rolling optimization solution success rate, task completion time, peak end-point pose deviation, minimum safe distance, number of degradation triggers, and recovery time after degradation. If, under conditions of edge load fluctuation and network jitter, the present invention's solution outperforms the control solution in multiple indicators, it demonstrates that the combination of a unified time baseline, shared perception access, and degradation switching mechanism has clear engineering value. The comparative experiments are used to illustrate the verification method of the present invention and do not constitute a limitation on the scope of protection.

[0068] Furthermore, in scenarios involving large-scale multi-robot deployments, the shared sensing space can be deployed using a hierarchical organization approach. Specifically, shared sensing results can be partitioned and stored according to work area, target category, or task priority, with independent result reading channels or stricter failure elimination strategies configured for high-priority tasks. For cross-regional movement tasks, the controller can pre-fetch environmental constraint information and candidate target results from the next partition before the robot approaches the region boundary, reducing the risk of sensing gaps at the moment of zone cutting. For systems with a large number of edge nodes, a node selection order can be established based on the edge node's geographical location, historical response latency, and reliability level to further improve the stability of shared sensing under large-scale deployment conditions.

[0069] In one specific embodiment, the present invention is also applicable to application conditions where ontological computing resources are limited but edge resources can be dynamically and elastically expanded. When a new edge node is detected to be connected or an existing edge node releases idle computing power, the controller can increase the offloading ratio of computationally intensive visual subtasks such as high-precision pose estimation, semantic segmentation, or local environment reconstruction. When edge resources shrink again, the task chain that contributes the most to the main control closed loop is prioritized, and the remaining visual subtasks are reverted to low-frequency update or event-triggered update modes. By maintaining stable admission logic and control mainline under resource expansion and contraction conditions, it can be further demonstrated that the present invention does not rely on fixed edge configuration, but has the ability to expand and adapt to actual industrial scenarios.

[0070] Furthermore, in special scenarios involving human-machine interaction, narrow aisle traffic, or highly reflective workstations, the controller can also set a scene correction factor based on the shared sensing results. This scene correction factor can be determined at least based on ground reflectivity, glass obstruction probability, personnel activity density, wireless channel congestion, and the availability of the onboard near-field sensors, and is used to synchronously correct the shared sensing reliability, degradation threshold, and local obstacle avoidance constraints. For scenarios with dense crowds and significant network jitter, the degradation switching sensitivity can be increased and the allowable output range of high-speed control commands can be compressed; for scenarios with high obstruction probability but where nearby robots can provide stable side views, the weight of results from auxiliary observation robots can be appropriately increased. By introducing a scene correction factor, this invention can maintain more stable control behavior under complex field conditions.

[0071] In one specific implementation, the controller can also perform short-window reuse analysis on historical results in the shared sensing space to identify the motion trend, occlusion duration, and observation source switching patterns of a target object over several consecutive control cycles. If the target object exhibits a stable motion trend within the short window, a higher proportion of the predicted compensation results is allowed to participate in the solution of the current control cycle; if the target object experiences drastic directional changes, frequent occlusion, or continuous conflicts between results from different sources within the short window, the proportion of predicted compensation is preferentially reduced, and the weight of ontological near-field observations and safety constraints in rolling optimization is increased. This historical short-window reuse analysis not only improves the utilization efficiency of shared sensing results in continuous control cycles but also provides an executable basis for predicting abnormal trends and early degradation.

[0072] In one specific implementation, the rolling optimization solver can adaptively select between a full optimization mode and a simplified optimization mode based on the number of available shared sensing results, edge resource occupancy, and remaining margin of the main actuator within the current control cycle. When there are sufficient available shared sensing results and a sufficient solution time budget, the controller adopts a full optimization model that includes motion tracking, visual observability, obstacle safety, and edge resource constraints. When the number of shared sensing results decreases, edge resource fluctuations increase, or the solution time approaches its upper limit, it switches to a simplified optimization model that retains key safety constraints and the dominant tracking target to prioritize ensuring that control commands can be output on time within the current control cycle. By switching the solution complexity according to time limits and information quality, it can be further illustrated that the present invention not only focuses on whether the sensing results are admitted, but also on the real-time executability of the control solution after admission.

[0073] Furthermore, the switching boundary between the aforementioned complete optimization mode and simplified optimization mode is preferably recorded in the parameter template, and can be preset according to the robot platform's computing power, controller frequency, and on-site network conditions. This avoids unexplained solution mode jitter for the same task on different deployment platforms, thereby improving the consistency and reproducibility of the control strategy.

[0074] XI. Supplementary Implementation Methods for Real-time Scheduling, Conflict Arbitration, and Resumption of Execution In one specific implementation, before the start of each control cycle, the control processor preferably decomposes the current cycle latency budget B_k into acquisition and refresh budget, link transmission budget, edge inference budget, correlation verification budget, and rolling optimization budget. The controller performs predictive updates on each sub-budget based on task type, target risk level, network jitter statistics from the previous time window, and the busy / idle status of edge nodes. When the total predicted overhead approaches the upper limit of B_k, it prioritizes compressing the refresh frequency, result publication frequency, or historical result retention time of low-priority visual subtasks to ensure that the target state update and security constraint update corresponding to the main control link can still be completed on time within the current cycle.

[0075] Furthermore, to avoid insufficient link stability due to simply selecting edge nodes based on remaining computing power, the controller can establish a comprehensive score value H_j for candidate edge nodes. This score value can at least consider the expected queuing delay, round-trip communication delay, current load rate, available computing power ratio, historical successful completion rate, and delay jitter amplitude within the most recent control cycles. For nodes with lower scores but closer geographical locations to the robot, only visual subtasks with wider time limits or lower contribution to the closed loop can be assigned; for nodes with higher scores and lower jitter, priority is given to tasks that directly affect the output of the current control cycle, such as target state updates, critical obstacle perception, or end-effector pose correction.

[0076] In one specific implementation, when multiple visual subtasks such as target detection, target tracking, pose estimation, depth estimation, and semantic segmentation exist simultaneously within the same control cycle, the controller preferably assigns a deadline, a minimum result quality threshold, and a maximum slice occupancy ratio to each type of subtask. If edge resources cannot simultaneously meet the deadlines of all tasks, the task set is processed hierarchically according to the contribution of the main control closed loop. Subtasks directly related to target position, obstacle distance, and operational posture retain a synchronous update mechanism, while subtasks such as environmental semantic supplementation, long-distance candidate target screening, or low-risk area background mapping can be switched to low-frequency update, event-triggered update, or short-window historical results.

[0077] In multi-robot sharing scenarios, if multiple neighboring robots provide different observation results for the same target object within a similar time window, the controller can calculate a consistency index based on position deviation, posture deviation, and result confidence, and can also introduce arbitration factors such as source stability score, continuous tracking length, observation view coverage, and occlusion recovery speed. Sources that maintain stable output over multiple consecutive control cycles, have minimal timestamp jitter, and do not conflict with local observations of the target robot can be assigned a higher arbitration priority. Sources that only appear briefly in a single cycle or that significantly contradict the results of the target robot's near-field sensors are restricted to candidate auxiliary information and not directly included in the current closed-loop solution.

[0078] Furthermore, to reduce control input fluctuations caused by frequent admission and exit of shared sensing results under network jitter conditions, the controller preferably establishes sliding time window statistics for the shared sensing reliability C and the total latency tau of the sensing link, and uses entry hysteresis and exit hysteresis mechanisms for judgment. That is, a shared sensing result not only needs to satisfy C not less than T1 and tau not more than B_k in the current period, but also preferably needs to satisfy that the average reliability and latency fluctuation amplitude in the recent few periods are within the allowable range before it is promoted to the dominant control input; and if its reliability drops briefly but has not yet triggered the degradation condition, it is preferentially fused by weight reduction rather than immediately removed, so as to maintain the continuity of control state updates.

[0079] In one specific implementation, the limiting control command in the safety degradation control mode is not a fixed constant limit, but rather a hierarchical constraint executed based on the remaining sensing capability of the actuator, the task risk level, and the current actuator state. For the mobile chassis, the upper limit of speed and angular velocity can be contracted based on the near-field obstacle distance and the width of the passable area on the ground; for the robotic arm, the upper limit of joint speed, joint acceleration, or joint torque can be contracted based on the relative distance between the end effector and the target object, obstacle, and human activity area; for the vision actuator, high-speed rotation commands can be frozen and low-speed compensation range can be retained based on the field of view stabilization requirements. By setting degradation limiting ranges separately for different actuators, safety boundaries can be prioritized to be maintained under abnormal operating conditions rather than simply shutting down the entire system.

[0080] Furthermore, the recovery of the safety degradation control mode preferably adopts a phased confirmation mechanism. Specifically, when the shared perception confidence level rises above the degradation threshold T2 again, the total latency of the perception link returns to within the budget, and the body's near-field sensor does not detect any new high-risk obstacles, the controller does not immediately restore all control capabilities. Instead, it first restores the eligibility for auxiliary fusion of the shared perception results, then restores key visual subtasks that directly contribute to target tracking or pose correction, and finally restores the commands of the high-speed visual actuator and the upper limit of normal motion speed. A minimum continuous confirmation period can be set between each stage to avoid the system frequently switching between normal mode and degradation mode when the network status fluctuates repeatedly.

[0081] In one specific implementation, if the current embodied composite robot needs to simultaneously perform composite tasks such as grasping, obstacle avoidance, path tracking, and collaborative observation of the work area, the controller can establish independent data index areas, reading priorities, and failure elimination rules for different tasks within the shared perception space. For the result area directly serving the main control closed loop, a shorter failure window and a more stringent time consistency check are preferred; for the result area used for supplementary environmental understanding or downstream planning reference, a longer historical window and a higher reuse ratio can be retained. By implementing task-level partitioning management of the shared perception space, the risk of result competition and erroneous calls between different tasks can be reduced, and the interpretability of result scheduling behavior can be improved.

[0082] Furthermore, to facilitate on-site deployment and review support, the controller preferably establishes a fixed mapping relationship between parameter templates and calibration procedures. The calibration procedure includes at least unified time baseline synchronization error testing, idle link round-trip delay testing, edge node full-load and half-load response testing, shared perception reliability threshold tuning testing, and degradation recovery hysteresis testing. The calibrated parameter templates can be categorized and stored according to task type, network conditions, and robot platform capabilities, with defined limits on the allowable online correction magnitude. This demonstrates that the invention has a clear engineering implementation path in different scenarios and provides a traceable technical basis for subsequent amendments to claims, explanation of the scope of protection, and responses to examination opinions.

[0083] In one specific implementation, the operation log system preferably records the following information according to the control cycle: the unloading decision result of the visual subtask, the change of edge node score, the calculation result of shared perception credibility, the source of the adopted result, the cause of degradation triggering, the output range of the amplitude limiting control command, and the status of the recovery confirmation stage. During offline review, the frequency of timeout events, the distribution of conflict arbitration results, and the distribution of recovery time can be statistically analyzed by task number, node number, target type, and anomaly category, and the budget decomposition coefficient, arbitration weight, and recovery window length of the next round of tasks can be adjusted accordingly. By describing the online scheduling, anomaly handling, and offline review in a closed loop, the engineering feasibility of the present invention and the sufficiency of the specification can be further enhanced.

[0084] In one specific implementation, when the shared sensing communication layer experiences a momentary disconnection, continuous packet loss, or a sudden drop in bandwidth, the controller preferably activates a result caching and body-priority rollback mechanism. That is, the controller first determines whether the most recent valid shared sensing result is still within the allowed reuse window. If it is, short-term prediction compensation is performed by combining the results from the body encoder, inertial measurement unit, and near-field sensors. If it exceeds the allowed reuse window, the controller immediately stops using cached results as the primary control input and transfers unfinished low-priority tasks from the edge side to a waiting queue or cancels them directly. By distinguishing between reusable and non-reusable anomalies, short-term communication fluctuations can be avoided from being misjudged as long-term failures, and outdated results can be prevented from remaining in the closed-loop path for extended periods.

[0085] Furthermore, a unified data interface verification rule is preferably established between the control processor and the shared sensing space. This rule may include at least field integrity verification, timestamp monotonicity verification, coordinate system identifier consistency verification, sequence number jump verification, and confidence level range verification. When any sensing result package has missing fields, timestamp reversal, coordinate system mismatch, or abnormal sequence number jumps, the controller can directly mark the result package as an unadmitted result and write the anomaly category to the operation log. By adding an interface consistency check before result admission, it can be further demonstrated that this invention does not solely rely on the accuracy of the upper-layer algorithm, but also considers the verifiability and accountability of shared data objects in engineering deployment.

[0086] In high-risk scenarios involving human-machine interaction, narrow-channel encounters, or delicate end-point operations, the controller preferentially increases the shared sensing admission threshold T1 and reduces the allowable latency margin, while amplifying near-field obstacle constraints, end-point safety boundary constraints, and degradation switching sensitivity. For such scenarios, even if edge nodes can continue to provide sensing results, if the shared sensing reliability fluctuation exceeds a preset range, the controller can prematurely enter a reduced-weight fusion or local degradation state to ensure that safety boundaries are prioritized over pursuing higher operational efficiency in high-risk scenarios. By providing a scenario-based threshold tightening mechanism, the correspondence between "technical features, technical problems, and technical effects" in the specification can be further strengthened.

[0087] In one specific embodiment, to verify that the control logic of the present invention can remain consistent under different deployment conditions, the controller can also perform batch comparative analysis on the unified time baseline error, edge node queuing delay, result admission ratio, degradation duration, and recovery confirmation cycle. If the same parameter template can still maintain stable result admission logic, no loss of synchronization in the main control closed loop, and predictable safety degradation behavior under different robot platforms, different wireless link conditions, and different combinations of edge nodes, it can be shown that the present invention has cross-platform reuse capability. This batch verification method can serve as an acceptance basis during engineering deployment and as supplementary support for demonstrating the feasibility and reasonableness of the scope of protection of the present invention.

[0088] XII. Examples of Implementation Methods In a comparative verification example, a composite task of simultaneous movement and grasping with dynamic obstacle avoidance was selected as the test object. Three sets of comparative schemes were set up under the same robot platform, the same target object, the same working area, and the same edge node hardware conditions. The first scheme only performed edge vision task unloading, without setting a unified time baseline and shared perception admission rules; the second scheme performed a unified time baseline and shared perception result reading, but did not include edge resource occupancy constraints in the rolling optimization solution; the third scheme executed the complete control flow of this invention, including unified time baseline, calculation of total latency of perception link, shared perception credibility admission, edge resource occupancy constraint entry, and safety degradation switching mechanism.

[0089] In a specific set of tests, the control period was set to 20 ms, the current control period delay budget B_k was set to 18 ms, the average load rate of edge nodes during peak task periods was 75%, and network links experienced random jitter ranging from 2 ms to 9 ms. Test results show that, compared to the first scheme, the third scheme improved the effective shared sensing result adoption rate from 68.4% to 86.7%, reduced the average stale time of results entering the control loop from 11.2 ms to 4.6 ms, and reduced the rolling optimization solution timeout rate from 14.5% to 3.1%. These results demonstrate that the combined effect of a unified time baseline and shared sensing admission rules can significantly improve the time effectiveness and availability of shared results when entering the closed loop.

[0090] Furthermore, compared to the second set of schemes, the third set of schemes reduced the number of control cycle out-of-step occurrences under edge node load fluctuation conditions from 27 times per 1000 cycles to 8 times, the peak position deviation of the robotic arm end effector during task completion decreased from 34 mm to 18 mm, and the minimum safe distance under dynamic obstacle scenarios increased from 0.21 m to 0.29 m, while the task completion time only increased by 3.8%. These results demonstrate that introducing edge resource occupancy constraints into rolling optimization does not simply sacrifice control efficiency, but rather achieves a more stable closed-loop solution and a higher safety margin at a relatively low cost.

[0091] In another set of abnormal link tests, continuous packet loss and bandwidth drop disturbances were injected into the shared sensing communication layer. The results showed that, compared to the control scheme without degradation admission and recovery hysteresis mechanisms, the number of recovery oscillations after degradation triggering in the present invention was reduced from 7 to 2, and the average duration of the degradation mode was shortened from 4.8 s to 3.1 s. Furthermore, no sudden chassis stops or robotic arm attitude changes occurred due to erroneous admission of outdated sharing results. These results demonstrate that the confidence threshold differentiation, degradation switching, and phased recovery mechanisms in this invention can maintain more predictable control behavior under abnormal network conditions.

[0092] The above comparison results show that the uniform time baseline, shared perception credibility admission, edge resource occupation constraint closed-loop admission, and safety degradation control of the present invention are not stacks of independent conventional designs, but act together under mutual restriction and mutual support to improve the closed-loop timeliness, closed-loop stability and safety under abnormal working conditions of shared perception results. For those skilled in the art, even if they are respectively aware of edge unloading, multi-robot shared perception or rolling optimization control, they cannot directly obtain the comprehensive technical effect achieved by the present invention without targeted system reconstruction.

[0093] In an example scenario, an embodied composite robot performs a task of moving and grasping simultaneously, and an adjacent robot provides side view information. The current control cycle is set to 20 ms, the current control cycle time delay budget B_k is set to 18 ms, the shared perception degradation threshold T2 is set to 0.55, and the shared perception control admission threshold T1 is set to 0.75.

[0094] When the result confidence of a perception result packet is 0.82, the consistency index is 0.88, and the total time delay of the perception link is 15 ms, the result satisfies C ≥ T1 and τ_i ≤ B_k, so it can enter the rolling optimization closed-loop as an effective shared perception result.

[0095] When the result confidence of another perception result packet is 0.62, the consistency index is 0.60, and the total time delay of the perception link is 16 ms, its shared perception credibility is between T2 and T1, so it only participates in state fusion in a manner of reducing weight, and is not used as a dominant control basis.

[0096] When the system detects C<T2 or τ_i>B_k for multiple consecutive control cycles, the controller immediately switches to the safety degradation control mode, limits the moving chassis speed and the movement amplitude of the mechanical arm within a safe range, and preferentially uses the on-body near-field sensor to perform local obstacle avoidance, so as to avoid control instability caused by abnormal shared perception.

[0097] The above examples are only used to illustrate the working mode of the present invention, and do not constitute a limitation on the protection scope of the present invention.

Claims

1. A collaborative control method for a holographic composite robot with shared sensing edges, characterized in that, This invention is applied to a collaborative control system comprising a current embodied composite robot, at least one edge node, a shared sensing and communication layer, and a shared sensing space. The system includes: acquiring the current embodied composite robot's body state, visual data, and task instructions based on a unified time baseline; determining visual sub-tasks based on the task instructions, and offloading the visual sub-tasks to at least one edge node for execution based on the current control cycle delay budget, network transmission delay, and the remaining computing power of the edge nodes; generating a perception result package containing target state information from the edge node and writing it into the shared sensing space; reading the perception result package corresponding to the current control task from the shared sensing space, calculating the total delay of the sensing link based on the unified time baseline, and performing delay compensation on the target state information in conjunction with the body motion state within the current control cycle; and performing spatiotemporal correlation and one-to-one communication for multiple perception result packages belonging to the same target object. Consistency verification is performed, and the shared perception credibility is determined based on the consistency verification result, the result confidence level, and the total latency of the perception link. Only perception results whose total latency of the perception link is not greater than the current control cycle latency budget and whose shared perception credibility is not lower than the shared perception control admission threshold T1 are determined as valid shared perception results. Based on the valid shared perception results, the ontology state, and the motion execution target, a motion vision shared state is constructed, and the control command is uniformly solved in a rolling optimization problem that includes edge resource occupancy constraints. When the shared perception credibility is lower than the shared perception degradation threshold T2, or the total latency of the perception link exceeds the limit for multiple consecutive control cycles, the system switches to a safe degradation control mode and outputs a limited safe control command based on the most recent valid motion vision shared state and the remaining ontology perception information, wherein the shared perception control admission threshold T1 is greater than the shared perception degradation threshold T2.

2. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, The unified time baseline is established by the robot body control unit, vision sensors, and edge nodes through at least one of the following methods: local master clock synchronization, network time synchronization, or time-sensitive network synchronization, to ensure the consistency of the acquisition timestamp, result generation timestamp, and control cycle time reference.

3. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, The perception result package includes at least the target identifier, acquisition timestamp, result generation timestamp, target status information, environmental constraint information, and result confidence level; the perception result packages in the shared perception space are indexed and stored according to the target identifier, task coordinate system spatial index, and time window, and historical results that exceed the time delay budget are invalidated and eliminated.

4. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, The spatiotemporal correlation and consistency verification filters candidate observation results based on target identification, time window and spatial proximity, and performs weight reduction processing and confidence calculation on multi-source observation results by calculating at least one consistency index among target position deviation, target attitude deviation and result confidence difference.

5. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, The visual subtask unloading is performed based on the scheduling cost to select edge nodes and allocate computing power slices. The scheduling cost takes into account the visual subtask priority, expected computation volume, latency exceedance, edge node load and remaining computing power ratio, and uses the slice allocation result as the basis for edge resource occupation constraints in subsequent rolling optimization.

6. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, The rolling optimization problem includes at least a motion tracking objective, a visual observability objective, an obstacle safety constraint, a control stability objective, and an edge resource occupancy constraint. The obstacle safety constraint is determined at least in part based on environmental constraint information, and the edge resource occupancy constraint is determined at least in part based on the edge computing power slice allocation result corresponding to the visual subtask.

7. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, When the shared perception confidence level is lower than the shared perception control admission threshold T1 but not lower than the shared perception degradation threshold T2, the weight of the corresponding perception result in state fusion is reduced.

8. The embodied composite robot shared sensing edge collaborative control method according to claim 1, characterized in that, The safety degradation control mode executes at least one of the following strategies: reducing the upper limit of the mobile chassis speed, reducing the upper limit of the robotic arm joint speed or joint torque, freezing the high-speed rotation command of the vision actuator, switching to a short-term holding strategy based on the most recent effective target pose, switching to a local obstacle avoidance strategy based on the body's near-field sensor, and pausing the addition of unloading requests for low-priority vision subtasks.

9. A collaborative control system for a holographic composite robot with shared sensing edge, characterized in that, This includes a current embodied composite robot, at least one edge node, a shared perception communication layer, a shared perception space, and a control processor. The control processor is configured to: acquire the current embodied composite robot's body state, visual data, and task instructions based on a unified time baseline; perform offloading scheduling of visual subtasks according to the task instructions and in conjunction with the current control cycle delay budget, network transmission delay, and the remaining computing power of the edge node; perform total perception link delay calculation, delay compensation, spatiotemporal correlation, and consistency verification on the perception result packets written to the shared perception space by the edge node, and determine the shared perception reliability; only those packets that meet the delay budget and have a shared perception reliability not lower than the shared perception reliability... The perception result at the control admission threshold T1 is determined as a valid shared perception result; a motion vision shared state is constructed based on the valid shared perception result, the ontology state, and the motion execution target, and control commands are solved in a rolling optimization problem that includes edge resource occupancy constraints; when the shared perception confidence is lower than the shared perception degradation threshold T2 or the total delay of the perception link exceeds the limit for multiple consecutive control cycles, the system switches to a safe degradation control mode and outputs a limited safe control command; wherein, the shared perception space is configured to index and store the perception result package according to the target identifier, the task coordinate system spatial index, and the time window, and to perform failure elimination on historical results that exceed the delay budget.

10. The embodied composite robot shared perception edge collaborative control system according to claim 9, characterized in that, When executing the safety degradation control mode, the control processor outputs a limited safety control command based on the most recent effective motion vision sharing state and the remaining perception information of the body.

Images