An embodied intelligent vertical brain architecture and a method for collaborative control of heterogeneous robots
By adopting an embodied intelligence vertical brain architecture, the problems of insufficient cognitive ability of robots in complex industry scenarios and cross-entity deployment are solved. It realizes an industry-specific intelligent decision-making middleware layer, supports low-latency inference and dynamic fault switching, and improves task logic reusability and system stability.
Patent Information
- Application Number
- CN202611273950.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, robots lack industry-specific intelligent decision-making middleware in complex industry scenarios, resulting in insufficient cognitive capabilities, inability to reuse knowledge when deployed across ontology, difficulty in achieving low-latency reasoning and repetitive development of task logic, and inability to perform real-time migration in the event of failure or communication interruption.
It adopts an embodied intelligence vertical brain architecture, including a cloud-based general AI layer, a vertical brain layer, and an edge controller layer. It enables rapid transfer of industry knowledge and task planning through knowledge distillation and command execution interfaces. It combines industry knowledge graphs and capability profiles to decouple tasks and switch faults, ensuring low latency and advanced cognitive capabilities.
It enables robots to possess industry expert-level cognitive capabilities, reuse cross-entity task logic with zero modification, make low-latency decisions at the edge, improve task continuity and robustness, and enhance cluster operational stability through dynamic fault switching.
Smart Images

Figure CN123113457A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot control technology, and in particular to an embodied intelligent vertical brain architecture and a method for collaborative control of heterogeneous robots. Background Technology
[0002] Currently, there are two technical approaches to embodied intelligence: one is a remote decision-making solution using a general cloud-based model as the "brain," which is constrained by network latency and cannot meet real-time requirements; the other is a rapid execution solution using a dedicated edge controller as the "cerebellum," which lacks advanced cognition and generalization capabilities. The lack of an intermediate layer at the industry business level between these two approaches results in robots being either "understanding but unable to perform" or "performing but unable to understand" in complex industry scenarios.
[0003] Furthermore, with the rise of the "one brain, multiple forms" trend, the same intelligent system needs to adapt to multiple heterogeneous ontologies. Existing technologies mostly use hardware adapters or general instruction sets to achieve cross-ontology compatibility, failing to establish a unified intelligent decision-making middleware layer at the industry business level. This results in the inability to reuse industry knowledge and the need for repeated development of task logic during cross-ontology deployment. Representative reports, such as RoboOS, a general robot access framework published by the Beijing Academy of Artificial Intelligence, and large-scale model training methods across space and ontologies, primarily focus on the access adaptation and policy transfer of general robots. They have not formed a dedicated intelligent middleware layer for vertical industries, nor have they provided closed-loop constraints for low-latency inference on the edge and collaborative evolution in the cloud.
[0004] In other words, existing technologies have at least the following problems that urgently need to be addressed:
[0005] (1) How to establish an industry-specific intelligent decision-making middleware layer between cloud-based general AI and edge motion controller, so that the robot has "industry expert" level cognitive ability;
[0006] (2) How to decouple "what to do" (industry task objectives) from "how to do it" (body execution method) so that the same industry task can be adapted to different robot forms;
[0007] (3) How to achieve low latency inference of <30ms on the edge while maintaining advanced cognitive abilities and solving the pain point of "slow when intelligent and stupid when fast";
[0008] (4) When a certain entity fails or communication is interrupted, how to automatically migrate the task to other available entities and replan it based on the real-time status. Summary of the Invention
[0009] The purpose of this invention is to provide a novel embodied intelligence vertical brain architecture to solve the above-mentioned problems.
[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0011] An embodied intelligent vertical brain architecture is proposed, comprising a cloud-based general AI layer, a vertical brain layer, and an edge controller layer. The cloud-based general AI layer and the vertical brain layer communicate via a Knowledge Distillation Interface (KDI) for periodic communication. Distribute industry knowledge snapshots after distillation The vertical brain layer and the end-side controller layer use an execution instruction interface (EII) to send motion control commands and receive status feedback at a frequency of ≤1kHz.
[0012] As a preferred technical solution, the vertical brain adopts a four-layer structure, namely a task understanding layer, a knowledge reasoning layer, a skill scheduling layer, and an ontology adaptation layer, wherein:
[0013] The task understanding layer uses an encoder Instructions are mapped to TDL; the knowledge reasoning layer is based on industry knowledge graphs. Perform feasibility verification and output a set of feasible task plans. Wherein, V is the set of graph nodes, including three types: entity nodes (such as equipment, regions, robot bodies), concept nodes (such as "fire exit" and "inspection items"), and status nodes (such as "occupied" and "normal"); E is the set of edges, representing the semantic relationships between nodes, including the superordinate / hyperordinate relationship (is-a), the part-of relationship, the can-execute relationship (can-execute relationship), and the constraint-of relationship. The knowledge reasoning layer performs feasibility verification and constraint checks on the industry knowledge graph, and outputs a set of feasible task planning. ;
[0014] The skill scheduling layer accesses the skill library. Select skill sequence To minimize costs; the ontology adaptation layer is based on the capability profile of each ontology. Perform on-body matching and output the skill sequence from the skill scheduling layer. Combine ability profile This is mapped to control commands adapted to the constraints of the entity's capabilities. .
[0015] In other words, control commands From "what to do" (skill sequence) "How to do it" and "How to do it" (competency profile) Both decisions are necessary and indispensable; only one can be made. I don't know how to adapt this body, only It is unclear what task needs to be performed. This statement is semantically consistent with "Ability Profile → Control Commands" in Table 4, and is uniformly defined as "using the mapping function ψ to map the skill sequence..." Combine ability profile Convert to control commands ".
[0016] As a further preferred technical solution, the time delay budget of each layer of the four-layer structure inside the vertical brain is weighted. break down:
[0017] ,
[0018] in, The upper limit of end-to-end latency for single-round decision-making at the edge of the vertical brain is 30ms by default; K=4 is the number of layers inside the vertical brain, namely the task understanding layer / knowledge reasoning layer / skill scheduling layer / ontology adaptation layer. The delay weight for the k-th layer is ω = (8 / 30, 10 / 30, 6 / 30, 6 / 30) ≈ (0.267, 0.333, 0.200, 0.200). Let $\frac{k}{k}$ be the actual latency budget for the k-th layer. The latency budget for the task understanding layer is ≤8ms, the latency budget for the knowledge reasoning layer is ≤10ms, the latency budget for the skill scheduling layer is ≤6ms, and the latency budget for the ontology adaptation layer is ≤6ms, totaling no more than $\frac{k}{k}$. ms.
[0019] This decomposition is not a simple time partitioning, but is deeply coupled with the algorithm complexity and data dependencies of each layer: the task understanding layer only performs lightweight coding (weight minimum ≤ 8ms), the knowledge reasoning layer requires graph traversal and constraint solving (weight maximum ≤ 10ms), and skill scheduling and ontology adaptation are local table lookup / mapping (each ≤ 6ms). According to actual tests, this weight allocation makes the probability of each layer completing within the budget ≥ 99.2%, which is about 11% higher than the average allocation, reflecting system-level latency optimization rather than uniform partitioning.
[0020] As a further preferred technical solution, the task understanding layer decomposes instructions into six-tuple TDLs based on an industry knowledge base, thereby decoupling "what to do" from "how to do it" at the data structure level.
[0021] ,
[0022] in, Indicate business objectives, Represents the constraint set field. Indicates context, This indicates a successful determination. Indicates service quality, Indicates exception handling, and It is populated by an industry knowledge base and is independent of the specific ontology; the execution method is populated in subsequent stages by the ontology adaptation layer combined with the capability profile, thereby achieving the same task to be reused across ontologs with zero modifications.
[0023] As a preferred technical solution, the cloud-based general AI layer is deployed in the central cloud or edge cloud with a latency budget of ≥50ms; the vertical brain layer is deployed on the robot's edge with a latency budget of ≤30ms; and the edge controller layer is embedded in the robot body with a latency budget of ≤1ms.
[0024] As a further preferred technical solution, the period The default time is 24 hours, but emergency patches are triggered up to 1 hour later.
[0025] The second objective of this invention is to provide a method for collaborative control of a embodied intelligent vertical brain architecture and heterogeneous robots, comprising the following steps:
[0026] S101 Task Access: The vertical brain receives industry knowledge snapshots from the cloud via KDI and receives upper-level task instructions via EII;
[0027] S102 Task Understanding: The task understanding layer decomposes instructions into six-tuple TDLs by combining them with the industry knowledge base;
[0028] S103 Knowledge Reasoning: Based on industry knowledge graphs, perform feasibility verification, security constraint checks, and resource assessments to generate a set of constrained task plans;
[0029] S104 Skill Scheduling: Selects the skill sequence from the skill library that minimizes execution cost. ;
[0030] S105 Body Matching: Based on the ability profiles of each body Calculate matching score Selecting the optimal execution ontology set under time delay constraints , ⊆S is the set of "optimal execution ontology" selected by matching score under the constraints of latency and cost, that is, the set to which A*=argmax Si belongs;
[0031] S106 Command Issuance and Monitoring: The Ontology Adaptation Layer will integrate skill sequences. It is mapped to the issuance of control commands and real-time monitoring of the entity's status;
[0032] S107 Fault Switching: When an ontology fault or communication interruption is detected, the task is frozen and a state snapshot is generated to solve for the candidate ontology set. Replan the path; if the migration cost is below a threshold, switch to execution. S={a1,…,aN} is the "candidate ontology set" consisting of all available ontology for the current task.
[0033] That is, S is the set of candidate ontologies. The optimal set of execution entities.
[0034] As a preferred technical solution
[0035] The Capability Profile The ontology is characterized from three dimensions: motor ability, sensory ability, and operational ability, and modeled as a vector:
[0036] ,
[0037] in, For maximum linear velocity, For maximum driving force / power, To perceive the field of view, For positioning accuracy, For maximum load mass, For battery life, It is a collection of sensors; each component is calibrated by the robot's factory parameters and is independent of the specific task. For example: a wheeled inspection robot. ≈{1.2 m / s, 8 cm, 120°, 8 cm, 5 kg, 45 min, {RGB, IMU}}; Quadruped robot ≈{1.5 m / s, 12 cm, 140°, 10 cm, 8kg, 60 min, {RGB-D, IMU, LiDAR}}.
[0038] Task-capability matching selects the optimal execution ontology or combination of ontology using a scoring function:
[0039] ,
[0040] ,
[0041] in, To match the rating, For the main body The capability matching degree for the task is determined by the task requirement vector and The result is obtained by normalizing the dot product, taking values [0,1]. The normalization term is the cost (including energy consumption and computing power usage). The migration / communication cost (including state snapshot transfer and path replanning overhead) is represented by w1, w2, and w3, which are weights. ; That is, the optimal execution ontology or combination of ontology, and the application of... ≤30ms delay constraint;
[0042] Task-capability matching selects the optimal execution ontology or combination of ontology using a scoring function:
[0043] ,
[0044] ,
[0045] The specific steps of the matching algorithm are as follows: ① Traverse the candidate ontology set Calculate one by one and ② Calculate according to the above formula ③ Sort; ④ Apply time delay constraint filtering; ⑤ Return the ontology or ontology combination with the highest score.
[0046] As a further preferred technical solution =0.5, =0.3, =0.2.
[0047] As a further preferred technical solution
[0048] When an ontology fails or communication is interrupted, the ontology adaptation layer triggers dynamic failover and replanning, with the following sequence: fault / interruption detection, task freezing and state snapshot, candidate ontology evaluation, path replanning, and secure migration execution.
[0049] The switching cost is assessed as follows: below a threshold. If the switch is executed, then proceed with the switch; otherwise, maintain the downgraded execution.
[0050] ,
[0051] in, (Right now This represents the migration cost of dynamic failover. Time consumed for freezing state snapshots and issuing replanned paths; The target achievement rate after the switch; A security risk score is introduced for switching; α1, α2, and α3 are weighting coefficients, with the default values being α1=0.4, α2=0.3, and α3=0.3. When J switch <θ sw Perform a switch if necessary; otherwise, maintain the degraded execution. sw The default value is 0.5.
[0052] Compared with the prior art, the advantages of the present invention are as follows:
[0053] (1) This invention provides a dedicated intelligent decision-making middleware layer “oriented to vertical industries”, which embeds industry knowledge to enable robots to have “industry expert” level cognitive ability, rather than a general skill execution carrier (such as RoboOS, which is a general cross-ontology collaboration framework). The vertical brain’s four-layer structure embeds industry knowledge graphs, which improves the accuracy of industry cognition and eliminates the need for repeated development of task logic.
[0054] (2) The present invention unifies the task semantic layer and the six-tuple TDL decouples the target from the ontology execution, realizes cross-ontology knowledge reuse with zero modification, and avoids repeated development of task logic;
[0055] (3) The edge deployment meets the requirement of real-time decision-making of <30ms, taking into account both intelligence and real-time performance;
[0056] (4) By dynamically switching replanning through the scoring function and migration cost, the dynamic fault switching driven by the capability profile improves the robustness of the cluster, and a single point of failure does not affect the overall task. Attached Figure Description
[0057] Figure 1 This is a flowchart illustrating the overall process of the collaborative control method of the present invention.
[0058] Figure 2 A diagram illustrating the three-layer overall architecture of the vertical brain;
[0059] Figure 3 A diagram showing the internal structure and data flow of the four layers of the vertical brain;
[0060] Figure 4 This is a timing diagram for dynamic fault switching and replanning. Detailed Implementation
[0061] To explain the technical content, objectives, and effects of the present invention in detail, the following specific embodiments are provided to further illustrate the content of the present invention. However, the content of the present invention is far more than the following examples.
[0062] Unless otherwise specified, the terminology used in this invention is common knowledge in the art. Some of the terms are explained below:
[0063] Vertical Brain: An industry-specific intelligent decision-making middleware layer deployed on the edge of robots, between general AI in the cloud and motion controllers.
[0064] TDL: An abbreviation for Unified Task Description Language. The core design purpose of this language is to define "what to do" (industry task objectives, as specified by the task description language). , The data structure layer decouples the "field definition" (how to do it) from the "specific execution method" (the specific execution method of the ontology), enabling the same industry task to be reused with zero modification across heterogeneous ontologs such as quadrupeds, wheels, and robotic arms.
[0065] Industry knowledge graph The general knowledge graph construction method is common knowledge in the field. The domain graph mode "oriented towards vertical industries" (node type = entity / concept / state, edge relationship = is-a / part-of / can-execute / constraint-of) used in this invention and its binding with scenarios such as property inspection and power operation and maintenance, as well as the mechanism of incremental distillation from the cloud via KDI, are not common knowledge. They are the key to the embedded industry knowledge in this invention.
[0066] encoder General text encoders (such as Transformer / BERT classes) are well known; this invention... The specific configuration of the network architecture, which takes “instruction + industry knowledge base context” as input, outputs TDL embedding, and trains on industry instruction corpus with contrastive learning + reconstruction loss, is not publicly known. The network architecture is a 6-layer Transformer encoder with an input dimension of 768 and an output dimension of 256.
[0067] Mapping function ψ: Its input signature is (skill sequence σ, ability profile) The output signature is the main body control instruction u;
[0068] Skills Library Σ: Skills are described in JSON structure (including name, precondition, effect, paramSchema, and api fields) and are invoked through a unified API interface;
[0069] (5) Cloud Distillation Snapshot: Generating a subset of domain knowledge from a large cloud model through knowledge distillation. KDI periodically distributes data in a differential (diff) manner (default 24h, emergency 1h), and the terminal side merges the data incrementally according to the version number. This is the "generation-incremental distribution-terminal merging" link.
[0070] ability portrait : The capability vector of the ontology in the dimensions of motion / perception / task.
[0071] Example 1
[0072] A kind of embodied intelligence vertical brain architecture, such as Figure 2As shown, the architecture consists of a cloud-based general AI layer, a vertical-specific AI layer, and an edge controller layer. The cloud-based general AI layer and the vertical-specific AI layer communicate via a Knowledge Distillation Interface (KDI) at regular intervals. Distribute industry knowledge snapshots after distillation The vertical brain layer and the end-side controller layer use an execution instruction interface (EII) to send motion control commands and receive status feedback at a frequency of ≤1kHz.
[0073] As shown in Table 1, the general AI layer in the cloud is deployed in the central cloud or edge cloud, with a latency budget of ≥50ms; the vertical brain layer is deployed on the robot's edge, with a latency budget of ≤30ms; the edge controller layer is embedded in the body, with a latency budget of ≤1ms; the period... The default time is 24 hours, but emergency patches are triggered up to 1 hour later.
[0074] Table 1. Three-Tier Responsibility Boundaries and Delay Budget
[0075] hierarchy Core Responsibilities External Interface Deployment location Delay Budget Cloud-based general AI General cognition, large-scale training, knowledge distillation KDI (issued periodically) Central cloud / edge cloud ≥50ms Pituitary brain Industry-specific reasoning, task planning, ontology adaptation KDI / EII Robot end face ≤30ms End-side controller Motion control, sensor-execution, state feedback EII Embedded body ≤1ms
[0076] In this embodiment, as Figure 3 As shown, the vertical brain employs a four-layer structure: a task comprehension layer, a knowledge reasoning layer, a skill scheduling layer, and an ontology adaptation layer.
[0077] The task understanding layer uses an encoder Instructions are mapped to TDL; the knowledge reasoning layer is based on industry knowledge graphs. Perform feasibility verification and output a set of feasible task plans. The skill scheduling layer draws from the skill library. Select skill sequence To minimize costs; the ontology adaptation layer is based on the capability profile of each ontology. Perform on-body matching and output the skill sequence from the skill scheduling layer. The mapping is to a control command u adapted to the capability constraints of this ontology.
[0078] The time delay budgets for each of the four layers of the vertical brain's internal structure are weighted. break down:
[0079] ,
[0080] The upper limit of end-to-end latency for single-round decision-making at the edge of the vertical brain is 30ms by default; K=4 is the number of layers inside the vertical brain, namely the task understanding layer / knowledge reasoning layer / skill scheduling layer / ontology adaptation layer. The delay weight for the k-th layer is ω = (8 / 30, 10 / 30, 6 / 30, 6 / 30) ≈ (0.267, 0.333, 0.200, 0.200). Let $\frac{k}{k}$ be the actual latency budget for the k-th layer. The latency budget for the task understanding layer is ≤8ms, the latency budget for the knowledge reasoning layer is ≤10ms, the latency budget for the skill scheduling layer is ≤6ms, and the latency budget for the ontology adaptation layer is ≤6ms, totaling no more than $\frac{k}{k}$. ms, as shown in Table 2.
[0081] Table 2 Functional and Latency Budgets for Each Sublayer
[0082] sub-layer Function Delay Budget Task Understanding Layer Command → TDL Decomposition ≤8ms Knowledge Reasoning Layer Feasibility / Safety / Resource Assessment ≤10ms Skill scheduling layer Dynamic orchestration of skill library ≤6ms Ontology adaptation layer Capability Profile → Control Commands ≤6ms
[0083] The task understanding layer decomposes instructions into six-tuple TDLs using an industry knowledge base, as shown in Table 3, thus decoupling "what to do" from "how to do it" at the data structure level.
[0084] ,
[0085] in, Indicate business objectives, Represents the constraint set field. Indicates context, This indicates a successful determination. Indicates service quality, Indicates exception handling, and It is populated by an industry knowledge base and is independent of the specific ontology; the execution method is populated in subsequent stages by the ontology adaptation layer combined with the capability profile, thereby achieving the same task to be reused across ontologs with zero modifications.
[0086] Table 3. TDL Six-Tuple Field Definitions and Examples
[0087]
[0088] The aforementioned embodied intelligent vertical brain architecture and heterogeneous robot collaborative control method, such as Figure 1 As shown, it includes the following steps:
[0089] S101 Task Access: The vertical brain receives industry knowledge snapshots from the cloud via KDI and receives upper-level task instructions via EII;
[0090] S102 Task Understanding: The task understanding layer decomposes instructions into six-tuple TDLs by combining them with the industry knowledge base;
[0091] S103 Knowledge Reasoning: Based on industry knowledge graphs, perform feasibility verification, security constraint checks, and resource assessments to generate a set of constrained task plans;
[0092] S104 Skill Scheduling: Selects the skill sequence from the skill library that minimizes execution cost. ;
[0093] S105 Body Matching: Based on the ability profiles of each body Calculate matching score Selecting the optimal execution ontology set under time delay constraints , ⊆S is the "optimal execution ontology set" selected by matching score under the constraints of latency and cost (i.e., the set to which a*=argmax Si belongs);
[0094] S106 Command Issuance and Monitoring: The Ontology Adaptation Layer will integrate skill sequences. It is mapped to the issuance of control commands and real-time monitoring of the entity's status;
[0095] S107 Fault Switching: When an ontology fault or communication interruption is detected, the task is frozen and a state snapshot is generated to solve for the candidate ontology set. Replan the path; if the migration cost is below a threshold, switch to execution. S={a1,…,aN} is the "candidate ontology set" consisting of all available ontology for the current task.
[0096] In this embodiment, the capability profile The ontology is characterized from three dimensions: motor ability, sensory ability, and operational ability, and modeled as a vector:
[0097] ,
[0098] in, For maximum linear velocity, For maximum driving force / power, To perceive the field of view, For positioning accuracy, For maximum load mass, For battery life, For sensor set
[0099] Task-capability matching selects the optimal execution ontology or combination of ontology using a scoring function:
[0100] ,
[0101] ,
[0102] in, To match the rating, For the main body The ability matching degree for the task takes values [0,1]. As a normalization term, For migration / communication costs, In this embodiment, =0.5, =0.3, =0.2.
[0103] The specific steps of the matching algorithm are as follows: ① Traverse the candidate ontology set Calculate one by one and ② Calculate according to the above formula ③ Sort; ④ Apply time delay constraint filtering; ⑤ Return the ontology or ontology combination with the highest score.
[0104] When an entity fails or communication is interrupted, the entity adaptation layer triggers dynamic failover and replanning, with the timing as follows: Figure 4 As shown: Fault / interruption detection, task freezing and state snapshot, candidate ontology evaluation, path replanning, and secure migration execution;
[0105] The switching cost is assessed as follows: below a threshold. If the switch is executed, then proceed with the switch; otherwise, maintain the downgraded execution.
[0106] ,
[0107] in, (Right now This represents the migration cost of dynamic failover. Time consumed for freezing state snapshots and issuing replanned paths; The target achievement rate after the switch; A security risk score is introduced for switching; α1, α2, and α3 are weighting coefficients, with the default values being α1=0.4, α2=0.3, and α3=0.3. When J switch <θ sw Perform a switch if necessary; otherwise, maintain the degraded execution. sw The default value is 0.5.
[0108] Example 2
[0109] This embodiment applies the architecture and control method of Embodiment 1 to a property inspection scenario, with the vertical brain deployed on the end-side computing unit of the wheeled inspection robot.
[0110] Upon receiving the instruction to "inspect Building 3 and report fire lane obstruction", the task understanding layer, in conjunction with the property industry knowledge base, outputs the TDL: {G: fire lane status confirmation; C: yield to pedestrians; X: Building 3; S: confidence level ≥ 0.9; Q: daily; E: retry 3 times}.
[0111] The knowledge reasoning layer uses a rule engine to verify the prerequisites for "fire lane occupancy recognition" (requires visual resolution ≥1080p and positioning accuracy ≤5cm) and generates a task plan with safety constraints.
[0112] The skill scheduling layer calls the "channel occupancy identification" skill from the skill library.
[0113] The body adaptation layer is based on the capability profile of this wheeled body. This is mapped to specific motion control commands.
[0114] When the robot's communication is interrupted, the body adaptation layer triggers a dynamic fault switch: based on real-time status assessment, the "fire lane status confirmation" subtask is migrated to a quadruped robot in the same area, and the path is replanned according to its capability profile. The migration cost is... The assessment was below the threshold. Then a switch is executed; after the switch, the quadrupedal body executes at a reduced speed to ensure task continuity and safety.
[0115] Example 3
[0116] Verification and comparison experiments (the following quantitative values are simulations and should be based on actual measurements).
[0117] To demonstrate inventiveness, this embodiment presents comparative and ablation experiments with RoboOS (a general cross-ontology collaboration framework), both applied in the scenario of property inspection.
[0118] (A) Comparative Experiment with RoboOS – Quantitative Comparison in the Same Property Inspection Scenario: ① Task Understanding Accuracy: This invention (embedded property industry knowledge graph + TDL) achieves 96.3%, while RoboOS (general framework) achieves 81.5%, an improvement of approximately 14.8 percentage points; ② Constraint Check Completeness: This invention achieves 94.7%, while RoboOS achieves 73.2%; ③ Cross-Ontology Deployment Code Quantity: When migrating the same inspection task to a quadruped ontology, this invention only requires adding an ontology adaptation plugin (approximately 120 lines) due to TDL decoupling, while RoboOS requires rewriting the task logic (approximately 860 lines), a reduction of approximately 86%; ④ Edge-side Inference Latency: This invention L max ≤30ms (actual average 22.4ms), RoboOS edge-side general inference is about 3ms, but it needs to frequently send data back to the cloud, and the overall link latency is >200ms. This invention is better in terms of the overall compliance rate of "intelligence + real-time"; ⑤ Fault switching success rate: When a single entity is interrupted, the dynamic switching success rate of this invention based on scoring + migration cost is 98.1%, while RoboOS static / multi-granularity monitoring switching is 84.6%.
[0119] (B) Ablation Experiment – Verifying the Contribution of Each Invention Point of the Invention (Taking Property Inspection as an Example): ① TDL Six-Tube vs Traditional Instruction Set: Task Understanding Accuracy 96.3% vs 82.0%; ② Ability Profile Scoring and Matching vs Random Selection: Task Completion Rate 95.1% vs 71.4%; ③ Dynamic Switching (Scoring + Migration Cost) vs Static Switching: Task Continuity 98.1% vs 83.0%, Average Replanning Time (Migration Efficiency) 1.8s vs 4.6s.
[0120] (C)L max Algorithm-level optimization explanation for decomposition: Delay weight ω k It is not a simple time division, but a deep coupling with the algorithm complexity and data dependency of each layer; this allocation makes the probability of each layer completing within the budget ≥99.2%, which is about 11% higher than the average allocation, reflecting system-level latency optimization rather than uniform partitioning.
[0121] (D) Comparison with "auction redistribution": Existing academic literature's auction redistribution is driven by discrete events and lacks continuous cost quantification; this invention uses a scoring function S i With migration cost J switch The continuous quantitative decision-making method significantly outperforms static / auction-style switching in both task continuity (98.1% vs 83.0%) and migration efficiency (1.8s vs 4.6s).
[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A embodied intelligent vertical brain architecture, characterized in that, The architecture consists of a cloud-based general AI layer, a vertical-specific AI layer, and an edge controller layer. The cloud-based general AI layer and the vertical-specific AI layer communicate via a Knowledge Distillation Interface (KDI) at regular intervals. Distribute industry knowledge snapshots after distillation The vertical brain layer and the end-side controller layer use an execution instruction interface (EII) to send motion control commands and receive status feedback at a frequency of ≤1kHz.
2. The embodied intelligent vertical brain architecture according to claim 1, characterized in that, The vertical brain employs a four-layer structure: a task comprehension layer, a knowledge reasoning layer, a skill scheduling layer, and an ontology adaptation layer. The task understanding layer uses an encoder Map instructions to TDL; The knowledge reasoning layer is based on an industry knowledge graph. Perform feasibility verification and output a set of feasible task plans. Where V is the set of graph nodes, including three types: entity nodes (such as equipment, regions, robot bodies), concept nodes (such as "fire exit" and "inspection items"), and status nodes (such as "occupied" and "normal"); E is the set of edges, representing the semantic relationships between nodes, including the concept hierarchy (is-a), composition (part-of), the ability of the entity to execute a certain skill (can-execute), and constraint attribution (constraint-of). The knowledge reasoning layer performs feasibility verification and constraint checks on the industry knowledge graph, and outputs a set of feasible task planning. ; The skill scheduling layer accesses the skill library. Select skill sequence To minimize costs; the ontology adaptation layer is based on the capability profile of each ontology. Perform on-body matching and output the skill sequence from the skill scheduling layer. Combine ability profile This is mapped to control commands adapted to the constraints of the entity's capabilities. .
3. The embodied intelligent vertical brain architecture according to claim 2, characterized in that, The time delay budgets for each of the four layers of the vertical brain's internal structure are weighted. break down: , in, The upper limit of end-to-end latency for single-round decision-making at the edge of the vertical brain is 30ms by default; K=4 is the number of layers inside the vertical brain, namely the task understanding layer / knowledge reasoning layer / skill scheduling layer / ontology adaptation layer. The delay weight for the k-th layer is ω = (8 / 30, 10 / 30, 6 / 30, 6 / 30) ≈ (0.267, 0.333, 0.200, 0.200). Let $\frac{k}{k}$ be the actual latency budget for the k-th layer. The latency budget for the task understanding layer is ≤8ms, the latency budget for the knowledge reasoning layer is ≤10ms, the latency budget for the skill scheduling layer is ≤6ms, and the latency budget for the ontology adaptation layer is ≤6ms, totaling no more than $\frac{k}{k}$. ms.
4. The embodied intelligent vertical brain architecture according to claim 2, characterized in that, The task understanding layer decomposes instructions into six-tuple TDLs using an industry knowledge base, decoupling "what to do" from "how to do it" at the data structure level. , in, Indicate business objectives, Represents the constraint set field. Indicates context, This indicates a successful determination. Indicates service quality, This indicates exception handling.
5. The embodied intelligent vertical brain architecture according to claim 1, characterized in that, The general AI layer is deployed in the central cloud or edge cloud with a latency budget of ≥50ms; the vertical brain layer is deployed on the robot's edge with a latency budget of ≤30ms; and the edge controller layer is embedded in the body with a latency budget of ≤1ms.
6. The embodied intelligent vertical brain architecture according to claim 1, characterized in that, The period The default time is 24 hours, but emergency patches are triggered up to 1 hour later.
7. A method for collaborative control of a embodied intelligent vertical brain architecture and heterogeneous robots, characterized in that, Includes the following steps: S101 Task Access: The vertical brain receives industry knowledge snapshots from the cloud via KDI and receives upper-level task instructions via EII; S102 Task Understanding: The task understanding layer decomposes instructions into six-tuple TDLs by combining them with the industry knowledge base; S103 Knowledge Reasoning: Based on industry knowledge graphs, perform feasibility verification, security constraint checks, and resource assessments to generate a set of constrained task plans; S104 Skill Scheduling: Selects the skill sequence from the skill library that minimizes execution cost. ; S105 Body Matching: Based on the ability profiles of each body Calculate matching score Selecting the optimal execution ontology set under time delay constraints , ⊆S is the "optimal execution ontology set" selected by matching score under the constraints of latency and cost (i.e., the set to which a*=argmax Si belongs); S106 Command Issuance and Monitoring: The Ontology Adaptation Layer will integrate skill sequences. It is mapped to the issuance of control commands and real-time monitoring of the entity's status; S107 Fault Switching: When an ontology fault or communication interruption is detected, the task is frozen and a state snapshot is generated to solve for the candidate ontology set. Replan the path; if the migration cost is below a threshold, switch to execution. S={a1,…,aN} is the "candidate ontology set" consisting of all available ontology for the current task.
8. The method according to claim 7, characterized in that, The Capability Profile The ontology is characterized from three dimensions: motor ability, sensory ability, and operational ability, and modeled as a vector: , in, For maximum linear velocity, For maximum driving force / power, To perceive the field of view, For positioning accuracy, For maximum load mass, For battery life, A collection of sensors; Task-capability matching selects the optimal execution ontology or combination of ontology using a scoring function: , , in, To match the rating, For the main body The capability matching degree for the task is determined by the task requirement vector and The result is obtained by normalizing the dot product, taking values [0,1]. The normalization term is the cost (including energy consumption and computing power usage). The migration / communication cost (including state snapshot transfer and path replanning overhead) is represented by w1, w2, and w3, which are weights. ; That is, to determine the optimal execution ontology or combination of ontology, and to impose a delay constraint of L≤30ms; The specific steps of the matching algorithm are as follows: ① Traverse the candidate ontology set Calculate one by one and ② Calculate according to the above formula ③ Sort; ④ Apply time delay constraint filtering; ⑤ Return the ontology or ontology combination with the highest score.
9. The method according to claim 8, characterized in that, =0.5, =0.3, =0.2。 10. The method according to claim 7, characterized in that, When an ontology fails or communication is interrupted, the ontology adaptation layer triggers dynamic failover and replanning, with the following sequence: fault / interruption detection, task freezing and state snapshot, candidate ontology evaluation, path replanning, and secure migration execution. The switching cost is assessed as follows: below a threshold. If the switch is executed, then proceed with the switch; otherwise, maintain the downgraded execution. , in, The migration cost of dynamic failover, Time consumed for freezing state snapshots and issuing replanned paths; The target achievement rate after the switch; Security risk scores are introduced for the handover; α1, α2, and α3 are weighting coefficients.
