A black light experiment cooperative control system and method based on multi-agent cooperative decision-making

Abstract

The present application relates to the technical field of artificial intelligence and black light experiment collaborative control, in particular to a black light experiment collaborative control system and method based on multi-agent collaborative decision. It includes multi-agent collaborative decision module, digital twin and protocol conversion module, black light experiment execution module, human-computer interaction quality control module and result analysis and skill sedimentation module. Through the cooperation of the five modules of multi-agent collaborative decision, digital twin and protocol conversion, multi-actuator arrangement, human-computer interaction quality control, result analysis and skill sedimentation, end-to-end full closed loop from natural language scientific research task input to 7*24 hours unattended experiment execution, full-dimensional data quality control to structured skill sedimentation reuse is realized. The pain points of high dependence on traditional scientific research manpower, poor experimental reproducibility and difficult knowledge inheritance are solved, and the scientific research efficiency and achievement quality are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and collaborative control technology for black-out experiments, specifically to a collaborative control system and method for black-out experiments based on multi-agent collaborative decision-making. Background Technology

[0002] With the deep integration of artificial intelligence, embodied intelligence, and automation technologies, experimental-intensive research fields are reaching a critical juncture in their transformation from "human-led" to "intelligent automation." The life sciences, materials science, and biomedicine fields have an increasingly urgent need for high-throughput, high-precision, and reproducible experiments, and traditional research models can no longer meet the efficiency and quality requirements of modern scientific research.

[0003] Current mainstream scientific research still adopts a "manual full-process participation" model, where researchers need to personally complete all aspects such as literature review, experimental design, sample preparation, instrument operation, data recording and analysis. More than 70% of the time is spent on repetitive mechanical labor, resulting in a serious waste of intellectual resources. At the same time, experimental operation relies on personal experience, and differences in operating techniques, environmental control, and parameter recording lead to a global success rate of less than 30% in reproducing scientific research results. A large amount of funding is invested in repeated verification rather than innovative breakthroughs.

[0004] Furthermore, existing laboratory automation systems are mostly single-function, independent devices that lack unified experimental protocol description standards and cross-device collaboration mechanisms. They cannot achieve full-process automation, nor can they transform successful experimental strategies and anomaly handling experience into reusable structured knowledge. This leads to different teams reinventing the wheel, resulting in extremely low efficiency in knowledge transfer. Summary of the Invention

[0005] This invention provides a collaborative control system and method for black-light experiments based on multi-agent collaborative decision-making. It constructs five core modules: multi-agent collaborative decision-making, digital twin and protocol conversion, black-light experiment execution, human-computer interaction quality control, result analysis and skill accumulation. It forms an end-to-end closed loop of "natural language task input → multi-agent collaborative scheme generation → digital twin verification → unattended experiment execution → data quality control → skill accumulation and reuse", thereby solving the problems mentioned in the background technology.

[0006] To achieve the above objectives, the collaborative control system for a blackout experiment based on multi-agent collaborative decision-making includes a multi-agent collaborative decision-making module, a digital twin and protocol conversion module, a blackout experiment execution module, a human-computer interaction quality control module, and a result analysis and skill accumulation module.

[0007] The multi-agent collaborative decision-making module includes, but is not limited to, management agents and dedicated agents such as literature and ideas, experimental design, security and resources, code generation, data analysis, equipment scheduling, visual quality control, and anomaly handling, and is equipped with shared memory and long-term context units.

[0008] The digital twin and protocol conversion module includes the necessary elements for building experimental tasks: a digital twin model, a feasibility assessment submodule, a ProtocolIR conversion submodule, and a multi-dimensional verification submodule.

[0009] The black light experiment execution module includes a multi-actuator orchestration submodule, an embodied intelligent execution unit cluster, and a full-dimensional data acquisition submodule.

[0010] The human-computer interaction quality control module includes a data reliability analysis submodule, an anomaly identification submodule, an experimental conclusion generation submodule, and a front-end interaction submodule;

[0011] Interactive terminals include, but are not limited to, web pages, mobile apps, mini-programs, desktop clients, laboratory control screens, robot interface interfaces, and hardware emergency stop and confirmation terminals.

[0012] The Results Analysis and Skills Accumulation module includes a key element extraction submodule, a skill template generation submodule, a hierarchical storage submodule, and a skill reuse recommendation submodule.

[0013] In the above technical solution, the multi-agent collaborative decision-making module replaces manual labor in literature research and scheme design, the digital twin and protocol conversion module builds a bridge between AI and physical laboratories, the black-light experiment execution module enables 24 / 7 unattended experiments, the human-computer interaction quality control module ensures the accuracy of results at key nodes, and the result analysis and skills accumulation module enables the structured reuse of scientific research experience. The five modules work together seamlessly to form a self-evolving scientific research closed loop.

[0014] Based on the above, a collaborative control method for a blackout experiment based on multi-agent collaborative decision-making includes:

[0015] S1. Multi-agent collaborative analysis of natural language research tasks: By managing agents, multiple specialized agents are coordinated to generate preliminary experimental plans, and a feasibility evaluation model is used for scoring and optimization.

[0016] S2. Verify the feasibility of the preferred solution using a laboratory digital twin model. If the verification is successful, convert the solution into a structured ProtocolIR and perform multi-dimensional verification. If the verification fails, automatically return to S1 for solution correction or trigger manual confirmation. If the verification is successful, proceed to S3.

[0017] S3. Based on the step type of ProtocolIR, each step is automatically assigned to the corresponding actuator, controlling the dark laboratory to complete the entire experimental operation under unattended conditions, and collecting experimental data in all dimensions in real time.

[0018] S4. Perform quality control analysis and outlier identification on all-dimensional experimental data, generate experimental conclusions and optimization suggestions, and push them to researchers for final confirmation or fine-tuning.

[0019] S5. Based on the experimental conclusions confirmed by humans, extract key success factors from successful experiments, transform them into structured skill templates, and store and reuse them in a hierarchical manner.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] Through the deep collaboration of five modules—multi-agent collaborative decision-making, digital twin verification, unattended execution, human-computer interaction quality control, and skill accumulation and reuse—an end-to-end intelligent scientific research closed loop has been constructed. This breakthrough overcomes the core bottlenecks of traditional scientific research models, freeing researchers from a large amount of repetitive mechanical labor, allowing them to focus on creative thinking and mechanistic research. Overall research efficiency has been significantly improved, and the R&D cycle has been greatly shortened. In several typical embodiments, efficiency can be increased several times and the R&D cycle can be reduced by more than half. Through standardized ProtocolIR commands and precise automated operation, the fluctuations in results caused by differences in human operation have been eliminated. The reproducibility of experimental results has been fundamentally improved, and the success rate of reproduction in several standardized experimental scenarios can reach a high level, fundamentally solving the reproducibility crisis that has long plagued the scientific community. At the same time, it transforms the implicit experimental experience that relies on the apprenticeship system into a searchable and callable structured skill template, realizing the standardized accumulation and cross-organizational reuse of scientific research knowledge, avoiding the loss of experience caused by personnel turnover, and breaking through the limitations of single-device automation, realizing 24 / 7 unattended experiments across devices and processes, significantly reducing the overall scientific research cost, and promoting the fundamental transformation of the scientific research model from "labor-intensive" to "intelligent and automated". Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall system architecture of the present invention, showing the hierarchical architecture and data interaction flow of the five core modules;

[0023] Figure 2 This is a schematic diagram of the internal composition and shared memory frame of the multi-agent collaborative decision-making module in this invention;

[0024] Figure 3 This is a schematic diagram of the internal processing flow of the digital twin and protocol conversion module in this invention, showing the complete process of digital twin verification, ProtocolIR generation and multi-dimensional verification;

[0025] Figure 4 This is a schematic diagram of the internal structure of the black-light experiment execution module in this invention;

[0026] Figure 5 This is a schematic diagram of the collaborative control method for a black-light experiment based on multi-agent collaborative decision-making according to the present invention, which includes a complete closed loop of skill accumulation and feedback iteration. Detailed Implementation

[0027] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] Due to fundamental problems such as high reliance on human resources, poor experimental reproducibility, difficulty in knowledge transfer, and fragmentation of automated systems, the existing scientific research model has long suffered from these issues. As a result, a large number of researchers spend more than 70% of their time on mechanical labor such as sample preparation, data recording, and repeated trial and error, resulting in a serious waste of valuable intellectual resources. At the same time, due to factors such as imprecise control of experimental conditions, differences in operating methods, and incomplete recording of environmental factors, the success rate of scientific research experiments is generally less than 30%. A large amount of funding and human resources are invested in verification rather than innovation. In addition, valuable experimental "feel" and experience in handling anomalies rely on oral and hands-on teaching through apprenticeship. When senior personnel leave, their experience is lost, making it difficult for the team's research and development capabilities to be continuously accumulated.

[0029] Therefore, in view of the problems mentioned above, refer to Figure 1-4 As shown, this invention illustrates a collaborative control system for a blackout experiment based on multi-agent collaborative decision-making.

[0030] correspond Figure 2 Multi-agent collaborative decision-making module

[0031] For details, please refer to the following: Figure 2 As shown, the system includes a multi-agent collaborative decision-making module, configured to receive scientific research experimental tasks in natural language form. Through a management agent, it coordinates multiple dedicated agents to collaboratively complete literature reviews, experimental design, safety assessments, and resource verification, generating a preliminary experimental plan. The multi-agent collaborative decision-making module adopts a distributed collaborative architecture, with the management agent acting as a central scheduling node, possessing capabilities for task decomposition, context synchronization, conflict coordination, and progress control.

[0032] The module has a built-in shared memory and long-term context unit to uniformly store project goals, laboratory resources, historical experimental data, researcher operation preferences and human feedback. All dedicated agents can read and update the data in this unit, ensuring information consistency in multi-agent collaboration.

[0033] When the management agent receives a natural language task input by the researcher (such as “design an experiment for T-cell CAR transfection and functional verification based on lentiviral vectors”), it first performs intent recognition, constructs a multimodal context including project goals, existing laboratory resources, historical experimental data, and researcher operation preferences, and stores it in the shared memory unit.

[0034] The task is broken down into four parallel sub-tasks: literature review, scheme design, security assessment, and resource verification. These sub-tasks are assigned to corresponding dedicated intelligent agents, and the output results of each agent are synchronized in real time to ensure logical coherence and parameter consistency.

[0035] Specifically, the dedicated agents in the multi-agent collaborative decision-making module include, but are not limited to:

[0036] The document and creative intelligence agent can connect to major global academic databases such as arXiv, PubMed, and CNKI, supports multilingual document retrieval and full-text analysis, and can automatically extract the core methods, key parameters and innovative points in the documents, and generate differentiated candidate research ideas by combining the laboratory's historical successful experimental data.

[0037] An experimental design agent is used to generate a scheme by combining historical experimental data with standard schemes. The generated scheme includes precise time points, reagent dosages, equipment parameters, operating steps and quality control standards, which can be directly used for subsequent verification and execution.

[0038] The safety and resource intelligence agent can query the inventory of laboratory consumables and the status of equipment reservations in real time, and automatically assess the biosafety level and compliance requirements according to the type of experiment to avoid resource shortages or safety violations.

[0039] Protocol conversion agents are used to convert experimental schemes verified by digital twins into a standardized ProtocolIR instruction set;

[0040] Code generation agents, data analysis agents, equipment scheduling agents, visual quality control agents, and exception handling agents can be extended and invoked as needed according to experimental scenarios.

[0041] correspond Figure 3 Digital Twin and Protocol Conversion Module

[0042] Then refer to Figure 3 As shown, the digital twin and protocol conversion module is configured to verify the executability of the preliminary experimental scheme by constructing a digital twin model of the necessary elements for the current experimental task, convert the verified scheme into a structured protocol intermediate representation ProtocolIR, and verify ProtocolIR.

[0043] The task-oriented digital twin model only covers the core elements involved in the current experiment, including the 3D model of the relevant automated equipment, operating parameters, motion trajectory and fault mode, as well as the specifications, storage location and usage status of the corresponding consumables. At the same time, it maps the laboratory space layout, robot motion range, collision boundaries and safety constraints, without the need to completely replicate the entire laboratory 1:1, thus taking into account both verification accuracy and implementation feasibility.

[0044] During the verification process, the digital twin model will simulate the entire process according to the steps of the experimental plan to detect whether there are problems such as equipment conflicts, path interference, parameter exceeding limits, and insufficient consumables, and predict the success rate and potential risks of the experiment.

[0045] For problems that can be automatically corrected during the simulation (such as adjusting the pipetting order to avoid equipment collisions), the digital twin module will directly correct the solution; for problems that cannot be automatically corrected (such as specific equipment failures or consumable shortages), a detailed problem report will be generated and returned to the corresponding dedicated intelligent agent for solution adjustment, or manual confirmation will be triggered.

[0046] The digital twin and protocol conversion module uses an experimental scheme feasibility evaluation model to optimize candidate schemes, and the calculation formula is as follows:

[0047] ①

[0048] In formula ①, For the comprehensive evaluation of the proposal, To predict the success rate of experiments based on historical data, For resource matching degree, For compliance and safety index, These are the weighting coefficients for the system's adaptive adjustment.

[0049] In a preferred embodiment, the initial recommended weight values ​​are w1=0.5, w2=0.3, and w3=0.2.

[0050] ProtocolIR is a standardized experimental protocol intermediate representation format defined in this invention. It adopts JSON structured data and includes fields such as device ID, robot type, end effector type, target object identifier, spatial coordinates and relative pose, grasping posture, force control parameters, speed limit, anomaly rollback strategy, manual confirmation node, permission level, and log tracking ID, in addition to basic experimental information, step sequence, parameter definition, resource requirements, quality control standards and anomaly handling strategy. It can be directly parsed and executed by all actuators, thus decoupling the experimental scheme from the underlying device.

[0051] The multi-dimensional verification submodule performs multi-dimensional verification on the generated ProtocolIR, including integrity, parameter range, pore layout, liquid volume, device compatibility, and biosafety, to ensure that the instructions are accurate.

[0052] correspond Figure 4 : Black Light Experiment Execution Module

[0053] Next, refer to Figure 4 As shown, the black-light experiment execution module is configured to include a multi-actuator orchestration submodule and an embodied intelligent execution unit cluster. It automatically assigns each step to the corresponding actuator according to the step type of ProtocolIR, completes the entire experimental operation under unattended conditions, and collects full-dimensional experimental data in real time.

[0054] The multi-executor orchestration submodule employs one or more scheduling algorithms combining rule-based scheduling, model predictive control, reinforcement learning, and heuristic optimization. Based on the type, priority, resource requirements, and equipment status of the steps, and with the goals of operational safety, execution efficiency, equipment conflict avoidance, path interference avoidance, and sample timing constraints, it can automatically allocate tasks to the optimal executor and achieve multi-task parallel scheduling and conflict avoidance.

[0055] For example, delicate operations such as liquid transfer and clone picking are assigned to robot actuators;

[0056] Control of equipment such as centrifuges, PCR instruments, and incubators is assigned to the equipment actuators; computational tasks such as data conversion and image analysis are assigned to the software actuators.

[0057] For dynamic adjustment steps that require real-time decision-making (such as adjusting the medium change time according to the cell growth status), the task is assigned to the intelligent agent actuator.

[0058] For special steps that require manual intervention (such as handling high-risk biological samples, operating flammable and explosive reagents, confirming suspected contamination, resetting equipment malfunctions, etc.), the generated manual task is pushed to the corresponding personnel's terminal. The system records the confirmer's identity, permission level, confirmation time, and processing opinion. After the manual completion and permission verification are passed, the subsequent process will continue automatically.

[0059] The actuator types and functions in the multi-actuator orchestration submodule are as follows:

[0060] Robotic actuators are used for precise and embodied operations such as liquid handling, sample transfer, clone picking, consumable identification and grasping, culture container handling, pipette / tip / well plate / culture dish operation, collaborative operation of robotic arm and dexterous hand, opening and closing of equipment door cover, sample loading and unloading, and microscopic imaging assistance.

[0061] Equipment actuators are used to control automated equipment such as centrifuges, PCR instruments, and incubators.

[0062] Manual actuators are used to handle special steps that require human intervention, covering scenarios such as handling flammable, explosive, toxic and volatile reagents, handling fire or leakage risks, handling scenarios where biosafety levels exceed limits, confirming suspected contamination, resetting failed robot grasping, handling equipment alarms, verifying abnormal sample status, and confirming that experimental results deviate significantly from expectations.

[0063] Software executors are used for data conversion, algorithm calls, and device communication;

[0064] The intelligent agent actuator is used for experimental steps that require real-time intelligent decision-making. It uses the vision-language-action (VLA) model to convert ProtocolIR steps into robot action sequences and corrects the grasping posture, motion path, force control parameters, motion speed and obstacle avoidance strategy in real time based on visual perception.

[0065] The full-dimensional data acquisition submodule can collect all data in real time during the experiment, including: environmental parameters (temperature, humidity, air pressure, CO2 concentration), equipment operation logs (rotation speed, temperature, running time, fault information), sample microscopic images (cell morphology, clonal morphology, staining results), and online quality control test results (absorbance, fluorescence intensity, cell viability). All data are accurately timestamped and identified by the equipment, stored in a distributed database, and are traceable, queryable, and can be used for subsequent data analysis and model training.

[0066] The human-computer interaction quality control module is configured to perform reliability analysis and anomaly identification on all-dimensional experimental data, generate experimental conclusions and suggestions for the next round of optimization, and push them to the front end for researchers to make final confirmation or fine-tuning. The final review and interpretation rights of the experiment are controlled by human researchers with the necessary permissions.

[0067] The data reliability analysis submodule uses a multi-dimensional data verification algorithm, combined with historical experimental data and quality control standards, to automatically identify abnormal data points, such as abnormal equipment operation, sample contamination, and test results deviating from the normal range.

[0068] For minor anomalies (such as fluctuations in individual data points), the system will automatically correct and label the data; for serious anomalies (such as equipment failure or experiment failure), the experiment will be immediately paused, an anomaly report will be generated and pushed to the researcher's terminal, and the current experimental status will be saved for subsequent investigation and recovery.

[0069] The experimental conclusion generation submodule performs statistical analysis and visualization processing on the valid data, generating a complete experimental report that includes experimental results, parameter influence analysis, conclusions and discussions. Based on the experimental results, it automatically generates optimization suggestions for the next round of experiments, such as adjusting reagent concentration, optimizing incubation time, and changing detection methods.

[0070] The comprehensive experimental data includes environmental parameters, equipment operation logs, cell or material microscopic images, and online quality control test results.

[0071] The results analysis and skills accumulation module uses the following calculation model during the analysis process:

[0072] Formula for calculating cell confluence:

[0073] ②

[0074] In formula ② The effective cell coverage pixel area identified by the visual model. For the total pixel area of ​​the field of view, when Automatically generate a succession instruction when the preset threshold is reached;

[0075] Cloning quality confidence score formula:

[0076] ③

[0077] In formula ③ For the overall score of clones, The morphological score is determined based on boundary clarity, nucleocytoplasmic ratio, cell packing density, and morphological uniformity. The positive signal intensity of pluripotent markers such as OCT4 / SOX2, This is a penalty for differentiating heterogeneous cells.

[0078] The results analysis and skills accumulation module is configured to extract key success factors from the entire process data of successful experiments based on manually verified experimental conclusions, and transform them into structured skill templates for reuse. The key factor extraction submodule uses natural language processing and machine learning algorithms to automatically extract key parameters, operational procedures, anomaly handling methods, and successful experiences affecting experimental results from experimental logs, operation records, and data reports. The skills template generation submodule organizes these elements into standardized structured templates, including basic experimental information, core parameters, detailed operating steps, quality control standards, common anomalies and handling methods, and reproducibility success rate. Templates must be approved and confirmed by authorized researchers before being added to the database.

[0079] The hierarchical storage submodule stores skill templates in the project skill library (visible only to project members), the organization skill library (shared by the entire organization), or publishes them to an external capability reuse platform after anonymization, based on their applicable scope and confidentiality level, for global researchers to retrieve, access, and trade. The skill reuse recommendation submodule intelligently recommends relevant skill templates based on the user's experimental tasks, historical operation records, and skill reuse value index, helping users quickly find the optimal experimental solution.

[0080] correspond Figure 5 Closed-loop control method flow

[0081] Next, refer to Figure 5 As shown, the collaborative control method for the black-light experiment based on multi-agent collaborative decision-making is as follows:

[0082] S1. The management agent recognizes the task intent, constructs multimodal data including project context, experimental context, and user context, and synchronizes it to the shared memory unit;

[0083] Literature and Creative Agent: Retrieve and analyze literature in the field, and propose at least 3 candidate research ideas; Experimental Design: Design an experimental scheme for the agent to generate a complete time sequence;

[0084] The security and resource intelligent agent completes security assessment and resource verification;

[0085] Multiple candidate solutions are scored using a feasibility assessment formula, and the solution with the highest overall score is selected.

[0086] S2. Verify the feasibility of the optimal solution by constructing a digital twin model of the necessary elements for the experimental task. If the verification is successful, convert the solution into a structured ProtocolIR and perform multi-dimensional verification.

[0087] If the verification fails, it will automatically return to S1 for scheme modification or trigger manual confirmation; if the verification passes, it will proceed to S3.

[0088] S3. Based on the step type of ProtocolIR, each step is automatically assigned to the corresponding actuator, controlling the dark laboratory to complete the entire experimental operation under unattended conditions, and collecting experimental data in all dimensions in real time.

[0089] When experiments involve cell culture, the VLA visual perception module in the lights-out lab calculates cell growth status using a cell confluence formula. When a preset threshold is reached, a passage instruction is automatically generated. The preset threshold can be configured according to the cell type, preferably 70%-80%. A new instruction set including actions such as PBS washing, trypsin digestion, and passage plating is automatically generated.

[0090] S4. Perform quality control analysis and outlier identification on all-dimensional experimental data, generate experimental conclusions and optimization suggestions, and push them to researchers for final confirmation or fine-tuning.

[0091] S5. Based on the experimental conclusions confirmed by humans, extract key success factors from successful experiments, transform them into structured skill templates, and store and reuse them in a hierarchical manner; at the same time, update the experimental conclusions, optimization experience and human feedback to the shared memory unit to support the iterative optimization of the next round of multi-agent scheme generation, forming a self-evolutionary closed loop.

[0092] The structured skill templates mentioned therein include at least: experimental parameters, operating procedures, experience in handling anomalies, and success rate of reproduction; the skill templates are stored hierarchically in the project skill library or the organization skill library according to their applicable scope, or are published to an external capability reuse platform for retrieval, access, or trading after being de-identified.

[0093] For existing skill templates, the system calculates their value using a skill reuse value index formula and dynamically sorts and recommends them based on this index.

[0094] ④

[0095] In formula ④, As a skill reuse value index, The frequency with which this skill is invoked by other teams or agents on the platform. The success rate of reproduction is calculated based on feedback data. The average recurrence period is denoted as .

[0096] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. The implementation of the present invention is not limited to software programs, applets, web pages, mobile apps, desktop clients, or other software carriers, nor is it limited to embedded hardware, dedicated chips, industrial control equipment, robot terminals, or other hardware forms. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A collaborative control system for a blackout experiment based on multi-agent collaborative decision-making, characterized in that, include: The multi-agent collaborative decision-making module is configured to receive scientific research experimental tasks in natural language form, manage agents, and rely on shared memory and long-term context units to synchronize project data, experimental status and historical feedback, collaboratively complete literature review, experimental design, safety assessment and resource verification, and generate preliminary experimental plans. The digital twin and protocol conversion module is configured to verify the feasibility of the preliminary experimental scheme by constructing a digital twin model of the necessary elements for the current experimental task, convert the verified scheme into a structured protocol intermediate representation ProtocolIR, and verify ProtocolIR. The black-light experiment execution module is configured to include a multi-actuator orchestration submodule and an embodied intelligent execution unit cluster. It automatically assigns each step to the corresponding actuator according to the step type of ProtocolIR, completes the entire experimental operation under unattended conditions, and collects full-dimensional experimental data in real time. The human-computer interaction quality control module is configured to perform reliability analysis and anomaly identification on all-dimensional experimental data, generate experimental conclusions and suggestions for the next round of optimization, and push them to the front end for researchers with the necessary permissions to make final confirmation or fine-tuning. The right to review the experiment and interpret the results is ultimately controlled by human researchers. The results analysis and skills accumulation module is configured to extract key success factors from the entire process data of successful experiments based on manually verified experimental conclusions, and transform them into structured skills templates for reuse.

2. The collaborative control system for a blackout experiment based on multi-agent collaborative decision-making as described in claim 1, characterized in that: The dedicated agents in the multi-agent collaborative decision-making module include, but are not limited to: The document and creative intelligence agent is used to retrieve and analyze literature in the field and propose at least three candidate research ideas based on the user context. An experimental design agent is used to generate complete time-series experimental plans by combining historical experimental data with standard schemes; Safety and resource intelligence agents are used to assess the biosafety and compliance of experiments, as well as the resource feasibility of consumables and equipment. Protocol conversion agents are used to convert experimental schemes verified by digital twins into a standardized ProtocolIR instruction set; Code generation agent, data analysis agent, equipment scheduling agent, visual quality control agent, and exception handling agent; The multi-agent collaborative decision-making module is also equipped with a shared memory and long-term context unit, which is used to store project context, experimental context, equipment operating status, historical decision records and human feedback data, providing unified data context support for all dedicated agents.

3. The collaborative control system for a blackout experiment based on multi-agent collaborative decision-making as described in claim 1, characterized in that: The digital twin and protocol conversion module uses an experimental scheme feasibility evaluation model to optimize candidate schemes, and the calculation formula is as follows: ① In formula ①, For the comprehensive evaluation of the proposal, To predict the success rate of experiments based on historical data, For resource matching degree, For compliance and safety index, These are the weighting coefficients for the system's adaptive adjustment. In a preferred embodiment, the initial recommended values ​​for the weight coefficients are w1=0.5, w2=0.3, and w3=0.2, and the system can adaptively and dynamically adjust them according to the experimental scenario and task type.

4. The collaborative control system for a blackout experiment based on multi-agent collaborative decision-making as described in claim 1, characterized in that: The multi-actuator orchestration submodule employs one or more scheduling algorithms combining rule-based scheduling, model predictive control, reinforcement learning, and heuristic optimization. With operational safety, execution efficiency, device conflict avoidance, path interference avoidance, and sample timing constraints as scheduling objectives, it automatically completes task allocation. The actuator types and functions are as follows: Robotic actuators are used for precise and embodied operations such as liquid handling, sample transfer, clone picking, consumable identification and grasping, culture container handling, pipette, pipette tip, well plate, culture dish operation, collaborative operation of robotic arm and dexterous hand, opening and closing of equipment door cover, sample loading and unloading, and microscopic imaging assistance. Equipment actuators are used to control automated equipment such as centrifuges, PCR instruments, and incubators. The manual actuator is used to handle special steps that require human intervention, including the handling of flammable, explosive, toxic and volatile reagents, handling of fire or leakage risks, handling of scenarios exceeding biosafety limits, confirmation of suspected contamination, reset of failed robot grasping, handling of equipment alarms, verification of abnormal sample status, and confirmation of significant deviations between experimental results and expectations. The system automatically records the identity of the confirmer, their access level, confirmation time and processing opinion. If the access requirements or the conditions for double verification are not met, the system will not allow the automatic continuation of the subsequent process. Software executors are used for data conversion, algorithm calls, and device communication; The intelligent agent actuator is used for experimental steps that require real-time intelligent decision-making. It uses the vision-language-action (VLA) model to convert ProtocolIR steps into robot action sequences and corrects the grasping posture, motion path, force control parameters, motion speed and obstacle avoidance strategy in real time based on visual perception.

5. The collaborative control system for a blackout experiment based on multi-agent collaborative decision-making as described in claim 1, characterized in that: The comprehensive experimental data includes environmental parameters, equipment operation logs, cell or material microscopic images, and online quality control test results; the result analysis and skill accumulation module uses the following calculation model during the analysis process: Formula for calculating cell confluence: ② In formula ② The effective cell coverage pixel area identified by the visual model. For the total pixel area of ​​the field of view, when Automatically generate a succession instruction when the preset threshold is reached; Cloning quality confidence score formula: ③ In formula ③ For the overall score of clones, The morphological score is determined based on boundary clarity, nucleocytoplasmic ratio, cell packing density, and morphological uniformity. The positive signal intensity of pluripotent markers such as OCT4 / SOX2, This is a penalty for differentiating heterogeneous cells.

6. A collaborative control method for a blackout experiment based on multi-agent collaborative decision-making. The collaborative control system and method for a blackout experiment based on multi-agent collaborative decision-making as described in claim 1 is characterized by comprising the following steps: S1. Multi-agent collaborative analysis of natural language research tasks: By managing agents, multiple specialized agents are coordinated to generate preliminary experimental plans, and a feasibility evaluation model is used for scoring and optimization. S2. Verify the feasibility of the optimal solution by constructing a digital twin model of the necessary elements for the current experimental task. If the verification is successful, convert the solution into a structured ProtocolIR and perform multi-dimensional verification. If the verification fails, it will automatically return to S1 for scheme modification or trigger manual confirmation; if the verification passes, it will proceed to S3. S3. Based on the step type of ProtocolIR, each step is automatically assigned to the corresponding actuator, controlling the dark laboratory to complete the entire experimental operation under unattended conditions, and collecting experimental data in all dimensions in real time. S4. Perform quality control analysis and outlier identification on all-dimensional experimental data, generate experimental conclusions and optimization suggestions, and push them to researchers for final confirmation or fine-tuning. S5. Based on the experimental conclusions confirmed by humans, extract key success factors from successful experiments, transform them into structured skill templates, and store and reuse them in a hierarchical manner. The accumulated skill templates and experimental feedback data are synchronously updated to the shared memory unit of the multi-agent collaborative decision-making module, supporting the optimization and iteration of the next round of experimental schemes, forming a complete closed loop.

7. The collaborative control method for a blackout experiment based on multi-agent collaborative decision-making as described in claim 6, characterized in that: Step S1 specifically includes: The management agent recognizes task intent, constructs multimodal data including project context, experimental context, and user context, and synchronizes it to the shared memory and long-term context unit to achieve full agent data sharing; The literature and creative intelligence agent retrieves and analyzes literature in the field, and proposes at least three candidate research ideas; Experimental design: An experimental scheme for generating complete time sequences of intelligent agents; The security and resource intelligent agent completes security assessment and resource verification; Multiple candidate solutions are scored using a feasibility assessment formula, and the solution with the highest overall score is selected.

8. The collaborative control method for a blackout experiment based on multi-agent collaborative decision-making as described in claim 6, characterized in that: In step S3, when the experiment involves cell culture, the visual-verbal-motor (VLA) visual perception module in the dark lab calculates the cell growth status using the cell confluence formula. When a preset threshold is reached, a passage instruction is automatically generated. The preset threshold can be configured according to the cell type, preferably 70%-80%. A new instruction set including actions such as PBS washing, trypsin digestion, passage and plate-planting is automatically generated.

9. The collaborative control method for a blackout experiment based on multi-agent collaborative decision-making as described in claim 6, characterized in that: The structured skill template described in step S5 includes at least: Experimental parameters, operating procedures, experience in handling anomalies, and success rate of reproduction; The skill templates must be approved and confirmed by authorized researchers before being stored in the project skill library or organization skill library according to their applicable scope, or published to an external capability reuse platform for retrieval, access, or trading after being anonymized.

10. The collaborative control method for a blackout experiment based on multi-agent collaborative decision-making as described in claim 6, characterized in that: For existing skill templates, the system calculates their value using a skill reuse value index formula and dynamically sorts and recommends them based on this index. ④ In formula ④, As a skill reuse value index, The frequency with which this skill is invoked by other teams or agents on the platform. The success rate of reproduction is calculated based on feedback data. The average recurrence period.

Images