Embodied Intelligent Agent Collaborative Perception System and Collaborative Perception Method

By constructing a decentralized distributed self-organizing collaborative network and adopting a local collaborative agent module and an adaptive data abstraction mechanism, the single point of failure and scalability issues of centralized architecture are solved, and efficient, robust and flexible collaborative perception of embodied intelligent agent clusters is achieved.

CN123123941APending Publication Date: 2026-11-17LINKER
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611074091.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Centralized collaborative control frameworks suffer from single-point failure risk, high communication energy consumption, limited scalability, and weak environmental adaptability, making it difficult to achieve large-scale, highly intelligent embodied intelligent agent cluster collaboration.

Method used

A decentralized distributed self-organizing collaborative network is constructed, and a local collaborative agent module is used for resource status monitoring, adaptive data abstraction, multi-objective optimization decision-making, and contract bidding mechanism to achieve distributed self-organizing collaborative perception.

Benefits of technology

It achieves improvements in communication energy consumption and bandwidth efficiency, enhances system robustness and scalability, enables individuals to have autonomous perception and decision-making capabilities, and can continuously maintain the optimal collaborative strategy in dynamic environments to adapt to complex tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure REF-OBJ-1784515340618-000001
    Figure REF-OBJ-1784515340618-000001
Patent Text Reader

Abstract

This invention discloses an embodied intelligent agent collaborative perception system, comprising several local collaborative agent modules, each corresponding to a specific entity within the embodied intelligent agent. These modules form a distributed, self-organizing collaborative network without a central control node. Each local collaborative agent module includes a resource status monitor, a task requirement parser, an adaptive data abstraction engine, a collaborative decision-maker, and a multi-source heterogeneous data fusion unit. This invention employs a decentralized, distributed, self-organizing architecture, eliminating the risk of single-point failure and improving system scalability. Through adaptive data abstraction and multi-objective optimization decision-making, it dynamically balances perception accuracy, latency, and energy consumption, improving system energy efficiency and environmental adaptability. It supports contract bidding and elastic reconfiguration, ensuring continuous and stable execution of perception tasks, and is suitable for various embodied intelligent agent cluster collaborative perception scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of embodied intelligent clusters and distributed collaborative sensing technology, and more specifically to an embodied intelligent agent collaborative sensing system and collaborative sensing method. Background Technology

[0002] With the development of embodied intelligence technology, collaborative systems of embodied intelligent agents, such as drone swarms, have been applied in scenarios such as public performances and emergency inspections. The current mainstream centralized collaborative control framework is mature in applications such as large-scale drone shows, but its inherent architectural defects have become the core obstacle to technological upgrades. This framework relies heavily on the central node to complete global planning and command distribution, which poses a risk of single point of failure. As the swarm size expands, communication and computing power bottlenecks easily occur, resulting in insufficient system scalability. Intelligent agents need to maintain high-frequency data interaction with the central node, resulting in high communication energy consumption and difficulty in dynamically optimizing energy efficiency. Individuals only have basic command execution capabilities and lack autonomous perception, decision-making, and neighbor-machine collaboration capabilities, resulting in weak environmental adaptability. At the same time, the system's functional logic is pre-fixed, making it difficult to dynamically adjust and expand according to scenario requirements, resulting in insufficient flexibility and reusability, which seriously limits the development of embodied intelligent agent swarms to larger scales and higher levels of intelligence. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an embodied intelligent agent collaborative perception system and method. By constructing a decentralized distributed self-organizing collaborative network and combining adaptive data abstraction and multi-objective optimization decision-making mechanisms, it solves the problems of high single-point failure risk, limited scalability, low energy efficiency, and weak environmental adaptability of centralized architectures.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an embodied intelligent agent collaborative perception system, comprising a plurality of local collaborative agent modules, wherein the plurality of local collaborative agent modules are respectively disposed within the embodied intelligent agent body, and the plurality of local collaborative agent modules constitute a distributed self-organizing collaborative network without a central control node, wherein the local collaborative agent modules include: The resource status monitor is used to monitor the local multi-dimensional resource status vector R_i(t)=[E_rem(t),C_load(t),B_avail(t),L_qual(t)] in real time, where E_rem is the remaining energy, C_load is the computing load, B_avail is the available communication bandwidth, and L_qual is the link quality with neighbors. The task requirement parser is used to receive the collaborative task descriptor issued by the upper layer application and parse it into a quantifiable collaborative requirement vector N_i={Data_Type,Accuracy_Req,Latency_Max,Update_Rate}. An adaptive data abstraction engine is used to process raw, perceived data in real time into multiple optional abstraction levels L∈{1,2,3,4}; The collaborative decision-maker is used to solve a multi-objective constrained optimization problem online based on local resources R_i(t), task requirements N_i, and neighbor contract information received from the network, in order to generate the optimal collaborative strategy A_i*. A multi-source heterogeneous data fusion unit is used to receive and fuse sensing data from different neighbors, potentially different abstraction levels L, and asynchronous data. A robust fusion algorithm is used to output the fused environmental state estimation results.

[0005] As a further improvement of the present invention, the multiple optional abstraction levels output by the adaptive data abstraction engine include: L1 represents raw / lightly impaired data: D_1 = Compress(RawData), high fidelity, high bandwidth; L2 is feature-level data: D_2={Feature_Points,Bounding_Boxes,Optical_Flow}, with a medium data volume; L3 is semantic-level data: D_3={Object_ID,Position,Velocity,Confidence}, with low data volume; L4 is decision-level data: D_4={Probability_Map,Decision_Suggestion}, with a very low data volume; Among them, the higher the abstraction level L, the smaller the data volume S(D_L), but the information loss L(D_L) may increase.

[0006] As a further improvement to the present invention, the optimization model used by the collaborative decision-maker is as follows: Minimize J_i(A_i)=ω_e*E_com(A_i)+ω_c*E_comp(A_i)+ω_q*Q_loss(A_i) Constraints: Perf(A_i)≥Perf_th(N_i); E_com(A_i)+E_comp(A_i)≤E_budget(R_i(t)); Latency(A_i)≤Latency_max(N_i) Where A_i is the action strategy, including [selecting the data abstraction level L for sending / receiving] and [establishing / maintaining cooperative links with which neighbors]; E_com and E_comp are the estimated communication and computing energy consumption; Q_loss is the information quality loss caused by data abstraction; ω_e, ω_c, and ω_q are dynamically adjusted weight coefficients. The cooperative decision-maker outputs the optimal strategy A_i* and the cooperative contract C_i published externally.

[0007] As a further improvement of the present invention, the collaborative contract C_i output by the collaborative decision-maker is a structured message: C_i = {ID_i,Offer(L_offer,Quality,S(D_L_offer)),Demand(L_demand,Req),Cost(cost calculated based on R_i(t)} Contracts are broadcast in the network periodically or triggered by events. After receiving the broadcast, neighboring nodes will perform distributed matching and bidding: if their own Demand matches the other party's Offer, they will calculate a "bid" response based on their own resources. The initiator will then select the neighbor with the best "cost-effectiveness" to establish a point-to-point collaborative link based on the optimization goal J_i.

[0008] Another aspect of the present invention provides an embodied intelligent agent collaborative perception method, which is mounted on the aforementioned embodied intelligent agent collaborative perception system and includes the following steps: Step 1, Requirement Publication: After agent A detects the target, the task requirement parser receives the collaborative task descriptor issued by the upper layer application and parses it into a quantifiable collaborative requirement vector N_A (containing the target pose data type, accuracy requirements, maximum latency, and 5Hz update rate indicators); simultaneously, the resource status monitor collects and generates a local multidimensional resource status vector R_A(t) in real time. Both are input into the collaborative decision-maker to start the collaborative contract generation process. Step 2, Contract Bidding: A's collaborative decision-maker combines local resource status and task requirements, matches the fourth-level data abstraction capability of the adaptive data abstraction engine, generates a structured collaborative contract C_A, and broadcasts it to the distributed self-organizing network; after receiving the contract, neighboring nodes B and C verify the requirement matching degree through their own task requirement parsers, evaluate their local resource carrying capacity through their resource status monitors, and finally generate response contracts through their respective collaborative decision-makers: node B outputs a contract C_B that provides L=3 semantic pose data and has a lower collaborative cost, while node C outputs a contract C_C that provides L=2 feature-level data and has a higher collaborative cost; Step 3, Decision Making and Chain Establishment: The collaborative decision-maker of A takes into account local resource constraints and task performance constraints, solves the multi-objective constraint optimization problem online, and compares the overall cost-effectiveness of the neighbor response contract; it determines that the L=3 semantic level data meets the task accuracy and latency requirements, and that the collaborative cost of node B is better, outputs the optimal collaborative strategy A_i*, selects to establish a point-to-point collaborative link with node B, agrees to transmit the sensing data at the L=3 level, after the link is established, the adaptive data abstraction engine of B processes the original sensing data into L3 semantic level data and sends it, the multi-source heterogeneous data fusion engine of A receives the data and performs robust fusion, and outputs the fused environmental state estimation result; Step 4, Adaptive Adjustment: During collaborative execution, the resource status monitors of A and B continuously monitor multi-dimensional resource parameters such as link quality L_qual, remaining energy, and computational load. When a deterioration in the AB link quality is detected, B's collaborative decision-maker re-solves the optimization model based on the current real-time resource status and generates optional adjustment strategies: a) Schedule the adaptive data abstraction engine to raise the data abstraction level to L=2, improving transmission robustness through feature-level redundancy information; b) Reduce transmission power and adapt to fault tolerance mechanisms to control energy consumption within an acceptable packet loss rate range; c) Propose to A the introduction of a relay node C, with the collaborative decision-makers of both parties completing the matching and negotiation of the relay link. Step 5, Elastic Reorganization: When B's resource status monitor detects that its remaining energy E_rem is lower than the safety threshold, B's collaborative decision-maker sets the Cost parameter in the externally broadcast collaborative contract to the maximum value; after A's collaborative decision-maker captures this contract change, it automatically triggers the optimization model to be resolved, switches and selects node C as the new collaborative partner and establishes a link; during the switching process, A's multi-source heterogeneous data fusion device smoothly transitions the data source, ensuring that the perception task is executed continuously without interruption.

[0009] Beneficial effects of the present invention 1. Order-of-magnitude improvement in energy consumption and bandwidth efficiency: By replacing "constant flow" with "on-demand abstraction," communication energy consumption E_com can be reduced by 1-2 orders of magnitude. The optimization model J_i directly aims to minimize energy consumption.

[0010] 2. Achieve true system robustness and scalability: No single point of failure, using a peer-to-peer network. The system reliability model changes from serial to parallel / networked: R_dist=1-∏_(m∈M)(1-R_m), where M is the minimum set of nodes required to complete the task, and M changes dynamically, resulting in a leap in reliability.

[0011] 3. Emergent high-order swarm intelligence: Individuals can emerge complex, adaptive, and flexible swarm behaviors (such as dynamic role assignment, self-healing networks, and collaborative reasoning) through simple local optimization (minJ_i) and contract interaction, enabling them to perform complex tasks in open environments.

[0012] 4. Providing a basic framework for general collaborative intelligence: This invention is not limited to specific carriers or tasks, but provides a unified collaborative perception and decision-making infrastructure for various embodied intelligent groups, and has broad prospects for industrial applications. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to the given embodiments.

[0014] The embodied intelligent agent cooperative perception system in this embodiment includes several local cooperative agent modules, which are installed one-to-one within the embodied intelligent agent. These local cooperative agent modules constitute a distributed self-organizing cooperative network without a central control node. Each local cooperative agent module includes: a resource status monitor, used to monitor the local multi-dimensional resource status vector R_i(t)=[E_rem(t),C_load(t),B_avail(t),L_qual(t)] in real time, where E_rem is the remaining energy, C_load is the computational load, B_avail is the available communication bandwidth, and L_qual is the link quality with neighbors; and a task requirement parser, used to receive data from the upper layer. The application issues a collaborative task descriptor, which is then parsed into a quantifiable collaborative requirement vector N_i={Data_Type,Accuracy_Req,Latency_Max,Update_Rate}. An adaptive data abstraction engine processes raw sensing data in real-time into multiple selectable abstraction levels L∈{1,2,3,4}. A collaborative decision-maker solves a multi-objective constrained optimization problem online based on local resources R_i(t), task requirements N_i, and neighbor contract information received from the network, to generate the optimal collaborative strategy A_i*. A multi-source heterogeneous data fusion unit receives and fuses sensing data from different neighbors, potentially different abstraction levels L, and asynchronous data. The system employs a robust fusion algorithm to output the fused environmental state estimation results. Each embodied intelligent agent independently completes resource perception, task parsing, and collaborative decision-making through a local collaborative agent module, without needing to upload status to the central node or wait for instructions. This fundamentally solves the problems of single point of failure and insufficient scalability in centralized architectures. The distributed self-organizing network supports dynamic joining and leaving of nodes, and the cluster size is not limited by the performance of the central node. Local multi-source heterogeneous data fusion can directly output environmental perception results, reducing dependence on the central node, improving the system's response speed and environmental adaptability, and reducing the energy consumption caused by long-distance high-frequency communication, thus achieving energy efficiency optimization.

[0015] Furthermore, the adaptive data abstraction engine outputs multiple optional abstraction levels L∈{1,2,3,4}, including: L1 for raw / slightly damaged data: D_1=Compress(RawData), high fidelity, high bandwidth; L2 for feature-level data: D_2={Feature_Points,Bounding_Boxes,Optical_Flow}, medium data volume; L3 for semantic-level data: D_3={Object_ID,Position,Velocity,Confidence}, low data volume; L4 for decision-level data: D_4={Probability_Map,Decision_Suggestion}, extremely low data volume. Among these, the higher the abstraction level L, the smaller the data volume S(D_L), but the information loss L(D_L) may increase. The adaptive data abstraction engine can output corresponding level of perception data according to the instructions of the collaborative decision-maker, providing an adapted data format for collaborative scenarios with different accuracy requirements and different bandwidth conditions, and helping the system balance perception quality and communication overhead in resource-constrained scenarios. An additional technical effect is that the four-level abstraction covers the complete granularity from raw data to decision results, which is compatible with different types of collaborative tasks and improves the system's task adaptability.

[0016] Furthermore, the optimization model used by the collaborative decision-maker is as follows: Minimize J_i(A_i)=ω_eE_com(A_i)+ω_cE_comp(A_i)+ω_q*Q_loss(A_i) with the following constraints: Perf(A_i)≥Perf_th(N_i); E_com(A_i)+E_comp(A_i)≤E_budget(R_i(t)); Latency(A_i)≤Latency_max(N_i). Where A_i is the action strategy, including selecting the data abstraction level L for sending / receiving and establishing / maintaining collaborative links with which neighbors; E_com and E_comp are the estimated communication and computing energy consumption; Q_loss is the information quality loss caused by data abstraction; ω_e, ω_c, and ω_q are dynamically adjusted weight coefficients. The collaborative decision-maker outputs the optimal strategy A_i* and the externally published collaborative contract C_i. The collaborative decision-maker uses local real-time resources and task requirements as constraints, and solves the optimal collaborative strategy through multi-objective optimization. It can dynamically adapt to the current environment and resource status at each decision moment, helping to solve the problems of fixed strategies in centralized architecture and the inability to dynamically optimize energy efficiency. An additional technical effect is that the weight coefficients can be dynamically adjusted according to the scenario priority, and can adapt to various working modes such as energy saving priority, accuracy priority, and low latency priority, thereby improving the system's scenario flexibility.

[0017] Furthermore, the collaborative contract C_i output by the collaborative decision-maker is a structured message: C_i={ID_i,Offer(L_offer,Quality,S(D_L_offer)),Demand(L_demand,Req),Cost(cost calculated based on R_i(t))}. This contract is broadcast periodically or triggered by events in the network. Upon receiving it, neighboring nodes' decision-makers perform distributed matching and bidding: if their own Demand matches the other party's Offer, they calculate a "bid" response based on their own resources. The initiator selects the neighbor with the best "cost-effectiveness" based on the optimization objective J_i to establish a point-to-point collaborative link. All collaborative nodes complete the distributed interaction of capabilities and needs through the structured contract, eliminating the need for a central node to perform global matching and scheduling, thus facilitating completely decentralized self-organizing collaboration. An additional technical benefit is that the bidding mechanism enables optimized resource allocation, prioritizing the selection of nodes with the lowest cost and matching capabilities to establish collaboration, further improving the overall system efficiency. Simultaneously, the event-triggered broadcast mechanism reduces ineffective communication overhead.

[0018] Furthermore, the embodied intelligent agent collaborative perception method is integrated into the aforementioned embodied intelligent agent collaborative perception system, including the following steps: Step 1, Request Publication: After intelligent agent A detects a target, the task request parser receives the collaborative task descriptor issued by the upper-layer application and parses it into a quantifiable collaborative request vector N_A (containing target pose data type, accuracy requirements, maximum latency, and 5Hz update rate indicators); simultaneously, the resource status monitor collects and generates a local multi-dimensional resource status vector R_A(t) in real time, and both are input into the collaborative decision-maker to initiate the collaborative contract generation process; Step 2, Contract Bidding: A's collaborative decision-maker combines local resource status and task requirements, matches the four-level data abstraction capability of the adaptive data abstraction engine, and generates a structured contract. The same contract C_A is broadcast to the distributed self-organizing network. After receiving the contract, neighboring nodes B and C verify the matching degree of the requirements through their own task requirement parsers and evaluate the local resource carrying capacity through their resource status monitors. Finally, their respective collaborative decision-makers generate reply contracts: node B outputs contract C_B, which can provide L=3 semantic pose data and has a lower collaborative cost, and node C outputs contract C_C, which can provide L=2 feature-level data and has a higher collaborative cost. Step 3, decision-making and chain establishment: A's collaborative decision-maker substitutes local resource constraints and task performance constraints to solve the multi-objective constraint optimization problem online and compares the overall cost-effectiveness of the neighbor reply contracts. It determines that L=3 semantic-level data has met the task accuracy and latency requirements, and that node B has a better collaborative cost, and outputs the optimal solution. The collaborative strategy A_i* selects to establish a point-to-point collaborative link with node B, agreeing to transmit sensing data at level L=3. After the link is established, the adaptive data abstraction engine on B processes the original sensing data into L3 semantic-level data and sends it. The multi-source heterogeneous data fusion unit on A receives the data and performs robust fusion, outputting the fused environmental state estimation result. Step four, adaptive adjustment: During the collaborative execution, the resource status monitors of A and B continuously monitor multi-dimensional resource parameters such as link quality L_qual, remaining energy, and computational load. When the A / B link quality deterioration is detected, B's collaborative decision-maker re-solves the optimization model based on the current real-time resource status and generates optional adjustment strategies: a) Schedule the adaptive data abstraction engine to upgrade the data abstraction level. b) Improve transmission robustness by using feature-level redundancy information to L=2; c) Reduce transmission power and adapt to fault tolerance mechanism to control energy consumption within an acceptable packet loss rate range; d) Propose to A the introduction of relay node C, and have the two parties' collaborative decision-makers complete the matching and negotiation of the relay link; Step 5, elastic reconfiguration: When B's resource status monitor detects that the remaining energy E_rem of the local machine is lower than the safety threshold, B's collaborative decision-maker sets the Cost parameter in the externally broadcast collaborative contract to the maximum value; after A's collaborative decision-maker captures the contract change, it automatically triggers the optimization model to be resolved, switches and selects node C as the new collaborative partner and establishes a link; during the switching process, A's multi-source heterogeneous data fusion device smoothly transitions the data source to ensure that the perception task is executed continuously without interruption.This method covers the entire process of collaborative tasks, from initiation and chain establishment to dynamic adjustment and fault reorganization. Each step is completed by the local autonomous decision-making of the node without the need for intervention from the central node, which helps to achieve stable collaborative perception under the distributed architecture. The additional technical effect is that the adaptive adjustment and elastic reorganization mechanism can cope with dynamic changes such as link fluctuations and insufficient resources, ensuring that the perception task is continuous and uninterrupted, and greatly improving the robustness and environmental adaptability of the system.

[0019] This embodiment provides the following two examples: 1: Wide-area disaster search and rescue Scenario: In a post-earthquake area, a cluster of 50 drones is conducting coordinated search and rescue operations. There is no central command and communication bandwidth is limited.

[0020] Implementation: 1. The cluster takes off in a randomized manner, with each machine broadcasting an initial contract (Offer: local sensor status; Demand: task area information).

[0021] 2. Drone U1, located at the edge, detects a large uncovered area, and its mission requirement N_1 is updated to "cover area A". Its decision-maker encodes this requirement into the contract broadcast.

[0022] 3. After receiving the drones U2 and U3, which were located in the central area and had sufficient power, they evaluated their own resources and then bid against each other. U2 won the bid because it was closer and had better link quality.

[0023] 4. U2 establishes collaboration with U1. U2 sends information to U1 and neighboring nodes with L=3 (semantic: "suspected vital signs heat source on the east side of area A @ coordinates (x,y)") instead of L=1 (infrared video stream), which greatly saves bandwidth.

[0024] 5. Upon receiving the information, the three drones automatically formed a temporary subgroup, using an L=1 data stream to perform detailed confirmation and multi-view fusion of the target area, and broadcast the confirmation results in an L=4 (Decision: "Survivors confirmed, supplies recommended") format. The entire process was completely self-organizing and adaptive, with no central node involved in scheduling.

[0025] 2: Collaborative Sensing at Smart Urban Intersections Scenario: Complex intersections, where multi-brand intelligent connected vehicles and roadside units (RSUs) work together to eliminate blind spots.

[0026] Implementation: 1. The vehicle V_car about to turn left has a left front blind spot. The task requirement N_car is to "obtain information on pedestrians / non-motorized vehicles in the blind spot with a delay of <100ms".

[0027] 2. V_car broadcasts the contract C_car via C-V2X. The oncoming vehicle V_truck (tall, with good visibility) and roadside unit RSU_1 (equipped with a panoramic camera) both receive the broadcast.

[0028] 3. V_truck decision-maker computation: Providing L=2 (accurate pedestrian bounding box and motion vector) can meet the requirements, but its own computational load is already high, and the estimated cost Cost_truck is relatively high. RSU_1 computation: Providing L=3 (pedestrian position and velocity estimation) is sufficient, and its computing power is sufficient and power supply is unlimited, with extremely low Cost_rsu.

[0029] 4. V_car's decision-maker selects to establish collaboration with RSU_1 based on the optimization objectives (focusing on low latency and low communication power consumption), receives its L=3 semantic data, and fuses it with its own sensor data to provide early warning of risks.

[0030] 5. When V_car leaves the intersection, the collaborative contract is automatically terminated, and related resources are released. This framework achieves optimized collaborative awareness of resources among heterogeneous devices.

[0031] In summary, this invention adopts a decentralized distributed self-organizing collaborative architecture, deploying a local collaborative agent module within each embodied intelligent agent. Through resource status monitoring, adaptive data abstraction, multi-objective optimization decision-making, and contract bidding matching mechanisms, it achieves fully distributed collaborative perception scheduling. This solves the problems of high single-point failure risk, limited scalability, low energy efficiency, and weak environmental adaptability in traditional centralized architectures. It can continuously maintain the optimal collaborative strategy under dynamic resource and environmental conditions, ensuring the stability and continuity of perception tasks, and is suitable for various large-scale embodied intelligent agent cluster collaborative scenarios.

[0032] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An embodied intelligent agent collaborative perception system, characterized in that: It includes several local collaborative agent modules, each corresponding to a specific entity within the embodied intelligent body. These local collaborative agent modules constitute a distributed, self-organizing collaborative network without a central control node. Each local collaborative agent module includes: The resource status monitor is used to monitor the local multi-dimensional resource status vector R_i(t)=[E_rem(t),C_load(t),B_avail(t),L_qual(t)] in real time, where E_rem is the remaining energy, C_load is the computing load, B_avail is the available communication bandwidth, and L_qual is the link quality with neighbors. The task requirement parser is used to receive the collaborative task descriptor issued by the upper layer application and parse it into a quantifiable collaborative requirement vector N_i={Data_Type,Accuracy_Req,Latency_Max,Update_Rate}. An adaptive data abstraction engine is used to process raw, perceived data into multiple optional abstraction levels L∈{1,2,3,4} in real time. The collaborative decision-maker is used to solve a multi-objective constrained optimization problem online based on local resources R_i(t), task requirements N_i, and neighbor contract information received from the network, in order to generate the optimal collaborative strategy A_i*. A multi-source heterogeneous data fusion unit is used to receive and fuse sensing data from different neighbors, potentially different abstraction levels L, and asynchronous data. A robust fusion algorithm is used to output the fused environmental state estimation results.

2. The embodied intelligent agent collaborative perception system according to claim 1, characterized in that: The adaptive data abstraction engine processes data into multiple optional abstraction levels L∈{1,2,3,4}, including: L1 represents raw / lightly impaired data: D_1 = Compress(RawData), high fidelity, high bandwidth; L2 is feature-level data: D_2={Feature_Points,Bounding_Boxes,Optical_Flow}, with a medium data volume; L3 is semantic-level data: D_3={Object_ID,Position,Velocity,Confidence}, with low data volume; L4 is decision-level data: D_4={Probability_Map,Decision_Suggestion}, with a very low data volume; Among them, the higher the abstraction level L, the smaller the data volume S(D_L), but the information loss L(D_L) may increase.

3. The embodied intelligent agent collaborative perception system according to claim 2, characterized in that: The optimization model used by the collaborative decision-maker is as follows: Minimize J_i(A_i)=ω_e*E_com(A_i)+ω_c*E_comp(A_i)+ω_q*Q_loss(A_i) Constraints: Perf(A_i)≥Perf_th(N_i); E_com(A_i)+E_comp(A_i)≤E_budget(R_i(t)); Latency(A_i)≤Latency_max(N_i) Where A_i is the action strategy, including [selecting the data abstraction level L for sending / receiving] and [establishing / maintaining cooperative links with which neighbors]; E_com and E_comp are the estimated communication and computing energy consumption; Q_loss is the information quality loss caused by data abstraction; ω_e, ω_c, and ω_q are dynamically adjusted weight coefficients. The cooperative decision-maker outputs the optimal strategy A_i* and the cooperative contract C_i published externally.

4. The embodied intelligent agent collaborative perception system according to claim 3, characterized in that: The collaborative contract C_i output by the collaborative decision-maker is a structured message: C_i = {ID_i,Offer(L_offer,Quality,S(D_L_offer)),Demand(L_demand,Req),Cost(cost calculated based on R_i(t)} Contracts are broadcast in the network periodically or triggered by events. After receiving the broadcast, neighboring nodes will perform distributed matching and bidding: if their own Demand matches the other party's Offer, they will calculate a "bid" response based on their own resources. The initiator will then select the neighbor with the best "cost-effectiveness" to establish a point-to-point collaborative link based on the optimization goal J_i.

5. An embodied intelligent agent collaborative perception method, mounted on the embodied intelligent agent collaborative perception system as described in any one of claims 1 to 4, characterized in that: Includes the following steps: Step 1, Requirement Publication: After detecting the target, Agent A receives the collaborative task descriptor from the upper-layer application through the task requirement parser, and parses it into a quantifiable collaborative requirement vector N_A (containing the target pose data type, accuracy requirements, maximum latency, and 5Hz update rate). Simultaneously, the resource status monitor collects and generates a local multidimensional resource status vector R_A(t) in real time. Both are input into the collaborative decision-maker to start the collaborative contract generation process. Step 2, Contract Bidding: A's collaborative decision-maker combines local resource status and task requirements, matches the fourth-level data abstraction capability of the adaptive data abstraction engine, generates a structured collaborative contract C_A, and broadcasts it to the distributed self-organizing network; after receiving the contract, neighboring nodes B and C verify the requirement matching degree through their own task requirement parsers, evaluate their local resource carrying capacity through their resource status monitors, and finally generate response contracts through their respective collaborative decision-makers: node B outputs a contract C_B that provides L=3 semantic pose data and has a lower collaborative cost, and node C outputs a contract C_C that provides L=2 feature-level data and has a higher collaborative cost; Step 3, Decision Making and Chain Establishment: A's collaborative decision-maker incorporates local resource constraints and task performance constraints to solve the multi-objective constraint optimization problem online, comparing the overall cost-effectiveness of neighbor response contracts; it determines that L=3 semantic-level data meets the task accuracy and latency requirements, and that node B's collaborative cost is better, outputting the optimal collaborative strategy A_i*, selecting to establish a point-to-point collaborative link with node B, agreeing to transmit sensing data at L=3 level. After the link is established, B's adaptive data abstraction engine processes the original sensing data into L3 semantic-level data and sends it, while A's multi-source heterogeneous data fusion engine receives the data and performs robust fusion, outputting the fused environmental state estimation result; Step 4, Adaptive Adjustment: During collaborative execution, the resource status monitors of A and B continuously monitor multi-dimensional resource parameters such as link quality L_qual, remaining energy, and computational load; when a deterioration in the AB link quality is detected, B's collaborative decision-maker re-solves the optimization model based on the current real-time resource status and generates optional adjustment strategies: a) The adaptive data abstraction engine is scheduled to raise the data abstraction level to L=2, and improve transmission robustness through feature-level redundancy information; b) Reduce transmission power and adapt to fault tolerance mechanism to control energy consumption within an acceptable packet loss rate range; c) Propose to A the introduction of relay node C, with both parties' collaborative decision-makers completing the matching and negotiation of relay links; Step 5, Elastic Reorganization: When B's resource status monitor detects that its remaining energy E_rem is lower than the safety threshold, B's collaborative decision-maker sets the Cost parameter in the externally broadcast collaborative contract to the maximum value; after A's collaborative decision-maker captures this contract change, it automatically triggers the optimization model to be resolved, switches and selects node C as the new collaborative partner and establishes a link; during the switching process, A's multi-source heterogeneous data fusion device smoothly transitions the data source, ensuring that the perception task is executed continuously without interruption.