A scientific embodied intelligent safety gating and dynamic accountability tracing method and system based on physical simulation and control obstacle function.
By employing a safety gating method based on physical simulation and control barrier functions, the problem of uninterrupted physical danger commands in embodied intelligence systems is solved, achieving precision in experimental safety and accountability, and ensuring the safety and compliance of embodied intelligence systems.
Patent Information
- Application Number
- CN202611535117.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-17
AI Technical Summary
Existing embodied intelligence systems suffer from a "syntax-security gap" in their end-to-end control architecture, which prevents physically dangerous commands from being effectively intercepted and smoothly corrected, and makes it impossible to accurately trace the responsibility of multiple parties after an incident.
A security gating method based on physical simulation and control barrier function is adopted. Experimental environment data is acquired through multimodal sensors, forward state inference verification is performed using the physical simulation proxy model of edge computing nodes, a security envelope set is constructed and minimal intrusive correction is performed, and dynamic responsibility traceability is achieved by combining hardware security gating and blockchain technology.
It enables proactive identification and interception of physically hazardous instructions, ensuring experimental safety and allowing for precise tracing of responsibility after an accident, avoiding the crude intervention and legal risks associated with traditional safety gates.
Smart Images

Figure CN123113628A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and provides a scientific embodied intelligent safety gating and dynamic responsibility tracing method and system based on physical simulation and control obstacle functions. Background Technology
[0002] Artificial intelligence-driven scientific research (AI for Science, AI4S) is evolving from cognitive reasoning to embodied automation of the physical world. With the large-scale introduction of Visual Language Models (VLM) and Visual Language Action Models (VLA), embodied intelligence, as the core hub connecting digital space and physical entities, is gradually acquiring end-to-end autonomous decision-making capabilities. Basic large-scale models integrate massive amounts of scientific literature and professional research tools, significantly reducing the time required for literature review and scientific reasoning. When the speed of scientific deduction is compressed to its limit by AI, the execution capability of physical experiments becomes the biggest bottleneck. "Self-Driving Laboratories" (SDLs), which connect the reasoning capabilities of large-scale models with the operational capabilities of robots, mark a new era in laboratory automation.
[0003] However, this leap in capabilities has also brought unprecedented physical security risks and dilemmas in determining legal liability.
[0004] First, there is a disastrous "syntax-to-safety gap".
[0005] While open-source large models have significantly narrowed the scientific knowledge gap, their vulnerability in physical execution environments cannot be ignored. Experimental protocols generated by fundamental large models may be perfectly legal in terms of "syntax" and chemical equations, but when guiding robotic arms, they lack semantic awareness of real-world physical hazards. This "syntax-safety gap" can lead to models issuing incorrect thermodynamic reaction conditions, violating the order of addition, or ignoring functionally hazardous ends when grasping glass beakers, potentially causing thermal runaway, cross-contamination, or the leakage of toxic substances.
[0006] Secondly, the handling of deformable media and high-frequency dynamic disturbances present significant challenges. In chemical synthesis and life science experiments, robots do not deal with rigid, regular parts found in factories, but rather with highly unpredictable deformable media, including volatile fluids, powdered reagents, and flexible biological tissues. Fluid pouring operations involve complex dynamic changes in high-dimensional partial differential equations (PDEs); when handling petri dishes or reagent bottles filled with strong acids, container breakage or slippage is highly likely. Current end-to-end data-driven models prioritize experimental efficiency over providing robust physical boundary proofs, making embodied agents highly susceptible to triggering a cascading cognitive-dynamic disaster when faced with sudden disturbances in the laboratory.
[0007] Third, the current end-to-end control architecture has fundamental safety flaws. The current industry mainstream practice is "end-to-end" control, where a large model directly outputs motor rotation commands based on camera feeds. This approach deprives the robot of the ability to "refuse dangerous commands," making physical damage inevitable if the AI malfunctions. Furthermore, traditional safety systems lock and shut down upon receiving an alarm (Emergency Stop), rendering experiments unusable. This crude intervention method fails to maintain task continuity while ensuring safety.
[0008] Fourth, international standard compliance and research ethics face dilemmas. Modern scientific instruments and laboratory electrical equipment are subject to extremely stringent international standards. The IEC 61010-1 standard (Safety requirements for electrical equipment for measurement, control and laboratory use) specifies in detail that equipment must have absolute protection against electric shock, mechanical stress, flame spread, fluid pressure bursts, and the release of toxic gases. The probabilistic instructions generated by AI models are simply unacceptable to traditional compliance systems. Furthermore, if AI misoperation leads to damage to millions of dollars worth of scientific instruments or leakage of high-risk biological agents, traditional product liability frameworks struggle to clarify the entanglements of responsibility between large model illusions, hardware failures, and data bias.
[0009] Fifth, the blurred boundaries of rights and responsibilities and the potential collapse of liability tracing are significant challenges. When embodied intelligent entities cause accidents during the execution of physical actions, the existing legal and liability tracing framework faces severe challenges. Under the current Civil Code and related product liability regulations, damages are usually awarded to the product's manufacturer or seller. However, embodied intelligent decision-making is a complex chain involving multi-source data input, algorithm pre-training models, edge real-time inference, and hardware sensor data collection. In 2024, the British Columbia Civil Adjudication Division in Canada established a global precedent in a case against Air Canada: companies cannot shift responsibility to chatbots or algorithms themselves and must bear legal responsibility for the output of their AI systems. If a real-time recording and auditing mechanism that penetrates the algorithm's black box cannot be established from the initial system design stage, and "misjudgments caused by algorithmic defects," "loss of control in underlying hardware execution," and "interference from improper user operations" cannot be clearly defined, companies will face incalculable legal risks.
[0010] In summary, existing technologies lack a comprehensive technical solution that can separate cognition from governance, provide multi-level security barriers for embodied intelligence systems during operation, and achieve dynamic accountability throughout the entire lifecycle. Summary of the Invention
[0011] The purpose of this invention is to solve the technical problem that existing embodied intelligent systems suffer from a “syntax-security gap” in their end-to-end control architecture, which makes it impossible to effectively intercept and smoothly correct physically dangerous commands, and also makes it impossible to accurately trace the responsibility of multiple parties after an accident.
[0012] To achieve the above objectives, the present invention employs the following technical means:
[0013] This invention provides a scientifically embodied intelligent safety gating and dynamic accountability method based on physical simulation and control obstacle functions, characterized by the following steps:
[0014] Step 1: Acquire experimental environment data using multimodal sensors, and generate a sequence of action commands to be verified based on the experimental environment data;
[0015] Step 2: Using a physical simulation proxy model deployed on edge computing nodes, embed physical process constraints into the forward state deduction process, perform forward state deduction verification on the action command sequence, and generate a safety verification result; if the safety verification result indicates a danger, intercept the action command sequence and trigger replanning.
[0016] Step 3: For the verified action instruction sequence, define a control barrier function based on the safety contract and construct a safe envelope set that satisfies forward invariance; when the nominal control instruction attempts to exceed the safe envelope set, construct constraints based on the Lie derivative condition of the control barrier function, and perform minimal intrusive modification of the nominal control instruction through quadratic programming to generate a safe control instruction;
[0017] Step 4: Perform compliance verification on the security control command through hardware security gating independent of the main control system. When a physical quantity is detected to exceed the preset security threshold, trigger a hardware-level security action.
[0018] Step 5: Package the execution data from Steps 1 to 4 using a hash algorithm, add a timestamp, and perform hierarchical signature authentication. Write the signed execution data and the corresponding physical feedback data into the blockchain for dynamic accountability tracking.
[0019] In the above scheme, step 1 includes:
[0020] Step 1.1: Obtain heterogeneous high-frequency data through binocular vision, infrared thermal imaging and spectral analysis, extract feature information in three-dimensional space, and predict the occupancy probability and mass attributes of voxels around the self-subject.
[0021] Step 1.2: Encode the real-time environmental data into scene image tuples, combine them with natural language experimental task instructions for reasoning, and generate a macroscopic action instruction sequence from coarse to fine.
[0022] The action instruction sequence is only used as an experimental request to be verified and does not have the authority to directly drive hardware execution.
[0023] In the above scheme, step 2 includes:
[0024] Step 2.1: Use a physical information neural network or a reversible normalized flow as the physical simulation proxy model;
[0025] Step 2.2: Embed the chemical kinetics, fluid mechanics, and instrument energy consumption models as soft constraints of the differential equations into the forward propagation of the physical simulation proxy model;
[0026] Step 2.3: Perform forward state extrapolation on the action command sequence within a millisecond time window to predict whether a center of gravity instability or thermal runaway side reaction will occur in future time steps;
[0027] Step 2.4: If the predicted trajectory indicates that a danger will occur, generate a hard intercept signal to prevent the issuance of the action command sequence and trigger an abnormal feedback to request replanning.
[0028] In the above scheme, step 3 includes:
[0029] Step 3.1: Bind a safety contract to each basic motion primitive, wherein the safety contract declares the allowed spatial boundaries, maximum liquid addition flow rate, maximum contact torque limit, and minimum human-machine safety distance;
[0030] Step 3.2: Based on the security contract, define the operation space control barrier function and construct a security envelope set, wherein the security envelope set satisfies forward invariance;
[0031] Step 3.3: When the nominal control command attempts to exceed the safe envelope set, the nominal control command is modified in a minimally invasive manner by solving a quadratic programming problem in real time, so as to constrain dangerous actions to the safe envelope boundary while maintaining task continuity.
[0032] In the above scheme, in step 3, the robot's dynamic model has an affine form:
[0033]
[0034] in, This indicates the robot's current state, including joint angles, position, and velocity; This represents the rate of change of state over time; Indicates control input; This represents the inherent dynamics of the system itself; Represents the control system matrix;
[0035] The security envelope set From continuously differentiable functions definition, If and only if The control barrier function satisfies the following Lie derivative condition:
[0036]
[0037] in, Represents the maximum set of instructions that the hardware can execute; This indicates the rate at which the safe distance changes with the system's natural inertia; This indicates the rate at which the safe distance changes with control intervention; For a strictly increasing function, the closer to the danger boundary, the more... As the value approaches 0, the closer the right side of the equation is to 0, the more tightly the upper limit of the system's speed in the direction of danger is locked when it is forced to move closer to danger.
[0038] The quadratic programming problem is:
[0039]
[0040]
[0041] in, This refers to the nominal control command; This indicates the final safety control instructions issued to the implementing agency; This is the constraint condition obtained by transforming the Lie derivative condition.
[0042] In the above scheme, step 4 includes:
[0043] Step 4.1: The hardware security gating system consists of a dual-channel independent security microcontroller with cross-monitoring function to eliminate common cause failures and meet the requirement that the hardware fault tolerance is greater than or equal to 1.
[0044] Step 4.2: The safety sensor array is directly hardwired to the hardware safety gating, and the safety sensor array includes at least one of a safety light curtain, a leakage detection sensor, a volatile gas sensor, and a six-axis torque sensor;
[0045] Step 4.3: When the quasi-static compressive force exceeds the preset biomechanical limit, or when unexpected vital signs break through the safety isolation zone in the workspace, the hardware safety gate ignores all upper-level instructions and triggers either a Category 0 safety action or a Category 1 safety action within milliseconds; wherein, the Category 0 safety action is to cut off the servo power, and the Category 1 safety action is to control the emergency stop.
[0046] In the above scheme, step 5 includes:
[0047] Step 5.1: Perform the first-level signature, namely the AI and sandbox system pre-signature: Package the current 3D environment map data, contextual prompts, generated policy confidence probability, and simulation pass certificate using a hash algorithm and add a timestamp;
[0048] Step 5.2: Perform the second-level signature, namely, human expert or hard boundary control confirmation signature: For actions that have been securely verified, the underlying security controller issues a certificate for automated countersigning confirmation; for operations that exceed the preset risk level threshold, a forced hover is triggered, requiring the researcher holding the system private key to manually confirm and authorize the signature through private key verification.
[0049] Step 5.3: The execution instructions with dual signatures and the accompanying physical feedback data are continuously written into the blockchain via high-frequency data stream; the physical feedback data includes at least one of the following: actual motor current, joint encoder deviation, lidar point cloud snapshot, and change in the transmittance of the liquid surface in the reactor.
[0050] In the above scheme, the dynamic accountability tracing includes locating the cause of the accident based on the immutable logs in the blockchain:
[0051] If the log decoding shows that the sequence of action instructions is dangerous and is not intercepted by the physical simulation agent model, then the cause of the accident is identified as a defect in the algorithm generation.
[0052] If the log shows that the dual signature is working correctly and the security control command has been issued correctly, but the execution mechanism fails to respond on time, the cause of the accident is identified as a hardware execution failure.
[0053] If the log records show that the hardware security gating has issued multiple unauthorized warnings, but there are records of overriding commands with administrator privileges or physical short-circuit records of safety interlock switches, then the cause of the accident is determined to be human intervention.
[0054] This invention also provides a scientifically embodied intelligent safety gating and dynamic accountability system based on physical simulation and control obstacle functions, comprising:
[0055] The cognitive planning module is used to acquire experimental environment data through multimodal sensors and generate a sequence of action instructions to be verified.
[0056] The edge physics simulation module is used to embed physical process constraints into the forward inference process using a physics simulation proxy model, and to perform forward state inference verification on the action command sequence.
[0057] The safety constraint correction module is used to define a control barrier function based on a safety contract, construct a safety envelope set that satisfies forward invariance, construct constraints based on the Lie derivative condition of the control barrier function, and perform minimal intrusive correction on the nominal control command through quadratic programming to generate a safety control command.
[0058] The hardware security gating module, which is set independently of the main control system, is used to perform compliance verification on the security control commands and trigger hardware-level security actions when a physical quantity is detected to exceed a preset security threshold.
[0059] The dynamic responsibility traceability module is used to package the execution data of the entire chain using a hash algorithm, add a timestamp for double signature authentication, and write the signed execution data and the corresponding physical feedback data into the blockchain.
[0060] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.
[0061] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:
[0062] 1. This invention addresses the "syntax-safety gap" problem where the experimental protocol generated by the embodied intelligence model may be completely legal in terms of "syntax" and chemical equations, but lacks semantic awareness of physical real dangers when guiding the robotic arm. To solve the technical problem that the large model may issue incorrect thermodynamic reaction conditions, violate the order of addition, or ignore functionally dangerous ends when grasping glass beakers, thus causing thermal runaway, cross-contamination, or leakage of toxic substances, this invention adopts a technique in step 2 that embeds chemical kinetics, fluid mechanics, and the energy consumption model of the instrument as soft constraints of differential equations into the forward propagation of the physical simulation proxy model. This allows the simulation verification to go beyond the rigid body kinematics collision detection of the robotic arm, and to perform forward state extrapolation of physical and chemical dangers such as whether the mixing of reagents will lead to strong thermal runaway side reactions and whether instability of the center of gravity will occur in future time steps. This achieves the effect of eliminating the "syntax-security gap" at the physical mechanism level, so that syntactically legal but physically dangerous instructions are identified before they are sent to the execution mechanism. The simulation sandbox immediately generates a hard interception signal to prevent the issuance of the action and triggers anomaly feedback to require the planning layer to recalculate, thereby ensuring that dangerous instructions are prevented from being issued before they reach the physical execution stage.
[0063] 2. This invention addresses the problems of traditional safety doors locking and stopping upon alarms (Emergency Stop), leading to experimental failure, and traditional laboratory automation relying on rigid trajectory playback and lacking "sovereignty" in the underlying controller. To resolve the technical dilemma of the trade-off between safety intervention and task continuity, this invention employs a technique in step 3 based on defining the Operational Space Control Barrier Function (OSCBF) and Boundary Control Barrier Function (BCBF) according to safety contracts. This constructs a safety envelope set that satisfies forward invariance, and uses quadratic programming to perform minimally invasive corrections to the nominal control commands. Specifically, the objective function of the quadratic programming requires minimizing the deviation between the final safety command and the nominal control command, with constraints derived from the Lie derivative conditions of the control barrier functions. It achieves the effect of maintaining the continuity of fluid operation tasks while "sliding" dangerous actions to the safety edge: the operation space control barrier function corrects micro-boundary violations in milliseconds at a servo update rate of several kilohertz, and the boundary control barrier function dynamically restrains the excessive tilt angle or excessive reagent injection command issued by AI without interrupting the operation process. Together, they maintain the task-consistent behavior to the greatest extent, completely abandoning the crude intervention mode of traditional safety doors locking and stopping the machine when an alarm is triggered, which leads to the failure of the experiment.
[0064] 3. This invention addresses the problem of existing end-to-end control architectures depriving robots of the power to "reject dangerous commands" and the unavoidable physical damage resulting from AI errors. To solve the technical problem of commands generated by the cognitive layer directly driving hardware execution without physical safety verification, this invention employs a technical approach in step 1 that strictly deprives the cognitive layer of its highest authority to directly drive hardware. All output commands are treated merely as "experimental requests" awaiting verification and must undergo legality review by lower-level systems. This achieves the effect of severing the direct control path between the large model and the hardware execution mechanism at the architectural level. This ensures that any action command must undergo physical simulation verification in step 2 and safety constraint correction in step 3 before physical execution. Even if the AI errs and outputs a dangerous command, the command cannot bypass the verification process and directly affect the physical world, providing an architectural safeguard for subsequent layers of safety defense.
[0065] 4. This invention addresses the issue that control barrier functions still rely on the computation of the main control CPU or GPU, and that software-layer security constraints completely fail when the computing chip crashes, memory overflows, or extreme external electromagnetic interference causes a system-level disaster. To solve the technical problem of pure software security defenses losing their protective capabilities under extreme conditions, this invention employs a hardware security gating independent of the main control system in step 4. This gating consists of dual-channel independent security microcontrollers with cross-monitoring capabilities to eliminate common cause failures (CCF), and a security sensor array directly hardwired to the hardware gating. This achieves the effect that even if all upper-layer AI instructions and software-layer control barrier functions completely fail due to computing power failure, the hardware security gating can still ignore all upper-layer instructions and trigger category 0 (cut off servo power) or category 1 (controlled emergency stop) safety actions within milliseconds, forming a "reflection-level" protection network that does not rely on any software computation, serving as the ultimate safety net to prevent embodied intelligent agents from causing fatal damage in the laboratory.
[0066] 5. This invention is based on embodied intelligence decision-making, which involves a complex chain of multi-source data input, algorithm pre-trained models, edge real-time inference, and hardware sensor acquisition. Considering the current legal and accountability frameworks facing collapse, the inability to clarify the illusion of large models, and the entanglement of responsibilities between hardware failures and data bias, this invention, in step 5, employs a dual-signature system combined with an immutable blockchain log to address the technical problem of accurately assigning responsibility after an incident, where the algorithm's black box cannot be penetrated. The first-level signature is a pre-screening signature from the AI and sandbox system, packaging and timestamping 3D environment map data, contextual clues, policy confidence probabilities, and simulation data using a hash algorithm. The second-level signature is a confirmation signature from a human expert or hard boundary control, forcibly triggering a pause for operations exceeding a preset risk level threshold and requiring confirmation and authorization from the researcher holding the system's private key. All execution instructions with dual signatures, along with accompanying physical feedback data such as actual motor current, joint encoder deviation, LiDAR point cloud snapshots, and changes in reactor liquid surface transmittance, are continuously written to the blockchain via high-frequency data streams. It achieves the effect of reconstructing the accident scene at the millisecond level based on tamper-proof logs after an accident, accurately distinguishing the three types of responsible parties: algorithm design defects, hardware manufacturing failures, and human intervention. It provides enterprises and research institutions with definitive digital cryptographic evidence, avoiding the shirking of responsibility and legal risks.
[0067] 6. This invention combines the three technical means—safety contract, control barrier function, and quadratic programming solution—in step 3, producing a synergistic effect that cannot be achieved by using any one or two of these techniques alone. Specifically:
[0068] If there is only a safety contract without a control barrier function and quadratic programming, the contract can only be executed in the traditional way, that is, once an over-limit is detected, a lock-up shutdown is triggered, resulting in the failure of the experiment. This is the crude intervention mode of the traditional safety gate locking up and shutting down (Emergency Stop) as soon as an alarm is detected, which leads to the failure of the experiment.
[0069] If there is only a control barrier function and quadratic programming but no safety contract, then there is a lack of differentiated safety boundary definitions for different basic action primitives such as centrifugation, titration, and heating, and the safety envelope set of the control barrier function cannot be established.
[0070] If there are only safety contracts and control barrier functions but no quadratic programming, it is impossible to find the correction instructions that are most consistent with the original intention of the AI within the safety envelope set, and it is impossible to achieve "minimally invasive" correction.
[0071] This invention combines these three elements and binds a hard-coded safety contract to each basic action primitive during instantiation to declare the action's exclusive spatial boundaries, maximum liquid flow rate, maximum contact torque limit, and minimum human-machine safety distance. By controlling the obstacle function, this contract is transformed into a continuously differentiable safety envelope set that satisfies forward invariance. Then, through quadratic programming, the safety control command with the smallest deviation from the nominal control command is solved within this safety envelope set. This allows for smooth and lossless enforcement of the safety contract at a servo update rate of several kilohertz, "sliding" dangerous actions to the safety edge and maximizing task-consistent behavior. This solves the technical problem in traditional safety mechanisms where ensuring safety requires interrupting the experiment, while maintaining the experiment cannot guarantee safety. Attached Figure Description
[0072] Figure 1 This is a simplified flowchart of the overall process of the method of the present invention;
[0073] Figure 2 This is a simplified diagram of the system architecture of the present invention;
[0074] Figure 3 A simplified diagram illustrating the working principle of the edge physics simulation sandbox layer;
[0075] Figure 4 A simplified diagram of the safety constraint correction principle for OSCBF;
[0076] Figure 5 A simplified diagram of the dual signature and dynamic accountability mechanism. Detailed Implementation
[0077] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.
[0078] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.
[0079] Example 1: Overall Method Flow
[0080] This embodiment provides a scientific embodied intelligent security gating and dynamic responsibility tracing method based on physical simulation and control barrier functions. The method designs a top-down four-layer defense line, covering the entire closed loop from operation design domain specifications to physical disconnection. It purifies the intentions of AI scientists step by step in the order of "macro-semantic understanding - digital physical deduction - micro-high frequency correction - hardware ultimate protection".
[0081] The method includes the following steps:
[0082] Step 1: Acquire experimental environment data using multimodal sensors, and generate a sequence of action commands to be verified based on the experimental environment data;
[0083] Step 2: Using a physical simulation proxy model deployed on edge computing nodes, embed physical process constraints into the forward inference process, perform forward state inference verification on the action command sequence, and generate a safety verification result; if the safety verification result indicates that there is a danger, then intercept the action command sequence and trigger replanning;
[0084] Step 3: For the verified action instruction sequence, define a control barrier function based on the safety contract and construct a safe envelope set that satisfies forward invariance; when the nominal control instruction attempts to exceed the safe envelope set, construct constraints based on the Lie derivative condition of the control barrier function, and perform minimal intrusive modification of the nominal control instruction through quadratic programming to generate a safe control instruction;
[0085] Step 4: Perform compliance verification on the security control command through hardware security gating independent of the main control system. When a physical quantity is detected to exceed the preset security threshold, trigger a hardware-level security action.
[0086] Step 5: Package the execution data from Steps 1 to 4 using a hash algorithm, add a timestamp, and perform dual signature authentication. Write the signed execution data and the corresponding physical feedback data into the blockchain for dynamic accountability tracking.
[0087] In this embodiment, cognition and governance are separated, and runtime shields are used to provide security for the embodied intelligence system, abandoning the traditional approach of internalizing all security logic into a single black-box model.
[0088] Example 2: Cognitive Layer: Multimodal Environment Analysis and Experiment Planning
[0089] This embodiment provides a detailed description of the specific implementation method of step 1.
[0090] In this embodiment, the system continuously receives heterogeneous high-frequency data from multimodal sensors (such as binocular vision, infrared thermal imaging, and spectral analysis), extracts feature information in three-dimensional space, and transforms it into a comprehensive expression of the environmental state. During this process, image attributes are processed through feature space dimensionality reduction or self-organizing map (SOM) to predict the occupancy probability and quality attributes of voxels surrounding the ego object.
[0091] The planner encodes real-time environmental data into scene image tuples. It combines the natural language experimental task instructions issued by human scientists to reason and generate a series of macroscopic action instruction sequences from coarse to fine.
[0092] It is important to note that although the planning layer possesses a strong ability to understand intent, in this system architecture, it is strictly deprived of the highest authority to directly drive the hardware. All output commands are treated only as "requests" to be verified and must undergo legitimacy review by lower-level systems.
[0093] Example 3: Sandbox Layer: Edge Physical Twin Simulation Pre-demonstration Based on Lightweight Proxy Model
[0094] This embodiment provides a detailed description of the specific implementation method for step 2.
[0095] Because large embodied models lack precise perception of on-site physical parameters during inference in the cloud or on high-computing-power nodes, the issued action commands must undergo real-time dynamic verification. However, automated platforms in laboratories often have strict size, weight, and power consumption (SWaP) limitations, making it impossible to run bulky traditional physics engines on onboard hardware. Therefore, this embodiment introduces an edge intelligence (EEI) simulation sandbox mechanism as the first line of defense to eliminate the "syntax-security gap".
[0096] Specifically, the sandbox layer employs lightweight proxy models such as Physical Information Neural Networks (PINNs) or Normalizing Flows image restoration techniques. Upon receiving macroscopic instructions, the proxy model performs forward state deduction within milliseconds. The simulation process not only considers the rigid body kinematics and collisions of the robotic arm but also embeds chemical kinetics, fluid mechanics, and the energy consumption models of the instrument as soft constraints in the differential equations into the forward propagation of the neural network.
[0097] If the predicted trajectory indicates that instability of the center of gravity will occur within the next few seconds, or that the mixing of reagents will lead to a strong thermal runaway side reaction, the simulation sandbox will immediately generate a hard interception signal to prevent the action from being issued and trigger anomaly feedback to require the planning layer to recalculate.
[0098] In a preferred embodiment, the lightweight physical simulation proxy model includes a fluid state prediction model and a reaction state prediction model. The fluid state prediction model employs a position dynamics-based calculation method to predict the liquid level, fluid distribution, and overflow amount at future time steps based on container geometry parameters, fluid properties, robotic arm motion states, and tipping action parameters. The reaction state prediction model uses trained neural operators to predict the temperature, pressure, and reaction rate at future time steps based on reactant types, ratios, temperature, pressure, and boundary conditions. Edge computing nodes compare the prediction results with preset thresholds for liquid level, overflow amount, temperature, and pressure in the safety contract; when any prediction result exceeds the corresponding threshold, a hazard assessment result is generated, and the issuance of the corresponding action command is prevented.
[0099] Example 4: Interception Layer: Microscopic Smoothing Intervention Based on Contractual Skills and Operational Space Control Barrier Function
[0100] This embodiment provides a detailed description of the specific implementation method of step 3.
[0101] Instructions that pass the macroscopic simulation test then enter the high-frequency skill execution layer. Traditional laboratory automation relies on rigid trajectory playback, and the underlying controller lacks "sovereignty." This embodiment innovatively proposes the concept of "contractual skills." For each basic motion primitive such as centrifugation, titration, and heating, a hard-coded safety contract must be bound during instantiation, declaring the allowed spatial boundaries, maximum liquid addition rate, maximum contact torque limit, and minimum human-machine safe distance.
[0102] To smoothly and losslessly enforce these contracts at servo update rates of several kilohertz, the system employs an Operation Space Control Barrier Function (OSCBF) and a Boundary Control Barrier Function (BCBF). The control barrier functions provide a mathematical guarantee of forward invariance for the nonlinear control system.
[0103] Specifically, assume that the dynamic model of the system has an affine form:
[0104]
[0105] in, Indicates the robot's current state (e.g., joint angles, position, and velocity); It represents the rate of change of the state over time (i.e., how the state will change in the next instant). This indicates the control input (i.e., the command issued to the motor, such as torque or voltage); It represents the inherent dynamics of the system itself, such as the natural downward tendency of a system when it is not powered on due to gravity and inertia. The control system matrix determines the control commands. How will this specifically affect the change in state?
[0106] Laboratory safety envelope set From continuously differentiable functions Definition, that is If and only if As long as the input is controlled The absolute safety of a system can be guaranteed if the following Lie derivative conditions are met:
[0107]
[0108] in, Represents the maximum set of instructions that can be executed by all hardware. Search for instructions in the middle; and These represent the rates at which the safe distance changes with the system's natural inertia and with control intervention, respectively. It is a strictly increasing function. The closer to the danger boundary ( The closer the system is to 0, the closer the right side of the equation becomes to 0. This forces a stricter limit on the system's speed in the direction of danger.
[0109] When the nominal control command issued by the large model To address the challenge of exceeding safety boundaries (e.g., the angular acceleration of poured liquid is about to cause splashing), OSCBF transforms the aforementioned inequalities into a real-time online quadratic programming (QP) problem, applying the nominal control law to the large model. Perform minimally invasive corrections:
[0110]
[0111]
[0112] in, The original action commands (nominal control laws) issued to the large model; This is the final safety command issued to the motor. Objective function. The meaning is: the security instructions that were finally found. It must be consistent with the original intent of the large model. Get as close as possible.
[0113] This optimization algorithm, while maintaining maximum consistency with the AI's original intent, "slides" its dangerous actions to the edge of safety. OSCBF can perform millisecond-level micro-invasive corrections while maintaining the continuity of fluid operation tasks, maximizing task-consistent behavior and completely abandoning the crude intervention mode of traditional safety doors that lock and shut down upon alarms (Emergency Stop), thus causing experiments to be scrapped.
[0114] In a preferred embodiment, considering the fragile nature of glassware, the system determines the upper limit of contact force, the upper limit of contact torque, and the minimum clamping force required for anti-slip based on the material, wall thickness, mass, and clamping posture of the glassware, and writes these parameters into the safety contract of the corresponding grasping action primitive. A multi-finger tactile sensor collects the contact force and contact torque, a disturbance observer estimates external disturbances based on the contact force, contact torque, and robot motion state, the controller constructs control obstacle function constraints based on the safety contract, and modifies the nominal grasping control command through quadratic programming so that the modified clamping force is between the minimum clamping force and the upper limit of the allowable contact force, thereby reducing the risk of glassware breakage or slippage.
[0115] In another preferred embodiment, for fluid pouring and reagent injection operations, the system constructs control barrier functions characterizing the fluid overflow boundary and the reaction temperature boundary based on container geometry parameters, current liquid level, tilt angle, tilt angular velocity, reagent injection flow rate, and reaction temperature. The inequality constraints derived from these control barrier functions are then incorporated into a quadratic programming problem. When the predicted liquid level is close to a preset overflow threshold or the predicted reaction temperature is close to a preset upper temperature limit, the quadratic programming solver modifies the nominal control command within the feasible range of constraints to reduce the tilt angular velocity or reagent injection flow rate, keeping the system state within a safe envelope set. When no feasible solution exists in the quadratic programming problem, a hovering command or a controlled stop command is generated.
[0116] Example 5: Compliance Layer: Absolute Hardware Security Gating for Standards such as IEC 61010-1
[0117] This embodiment provides a detailed description of the specific implementation method for step 4.
[0118] Although the control barrier function is mathematically flawless, it still relies on the computation of the main control CPU or GPU. To cope with system-level disasters caused by computing chip crashes, memory overflows, or extreme external electromagnetic interference, this embodiment forcibly inserts an independent hardware safety gate layer that meets extremely high compliance requirements at the final motor drive and instrument interface.
[0119] This layer is designed directly to IEC 61010-1 (Safety Requirements for Laboratory Equipment) and the industry standards IEC 62061 and ISO 13849-1. Its hardware consists of dual-channel independent safety microcontrollers with cross-monitoring capabilities to eliminate common cause failures (CCF), thereby meeting the requirement of a hardware fault tolerance (HFT) greater than or equal to 1, directly aligning with PLE and SIL 3 levels.
[0120] The functional positioning and standard response of each security mechanism are as follows:
[0121] (1) OSCBF continuous filtering: As a software-level microsecond defense line, it smoothly constrains tactile sliding, fluid overflow and temperature overrun, satisfies the continuity of dynamic process, and reduces experimental trial and error and cell sample consumption.
[0122] (2) Safety-level monitoring stop: As a hardware-level intervention, it maintains position and monitors drift when personnel approach, responding to the mandatory requirements of ISO 10218-1 / 2 and ISO / TS 15066.
[0123] (3) Power and force limitation (PFL): Power is cut off at the moment of impact or thermodynamic heating source is cut off at an extremely fast speed, strictly aligning with IEC61010-1 over-temperature / leakage prevention requirements, and not exceeding the biomechanical pain threshold.
[0124] Safety sensor arrays (such as safety light curtains, leak detectors, volatile gas sensors, and six-axis torque sensors) are directly hardwired to this hardware gating system. When the system detects that the quasi-static compressive force is about to exceed the Young's modulus of the glass and the biomechanical limits specified in Appendix A of ISO / TS 15066, or when unexpected vital signs breach the safety isolation zone (Speed and Separation Monitoring area) in the workspace, the hardware gating system ignores all upper-level AI commands and triggers either Category 0 (disconnecting servo power) or Category 1 (controlled emergency stop) safety actions within milliseconds. This "reflective" protection network serves as the ultimate safety net to prevent embodied intelligent agents from causing fatal injuries in the laboratory.
[0125] Example 6: Dynamic Responsibility Tracing and Dual Signature Mechanism
[0126] This embodiment provides a detailed description of the specific implementation method of step 5.
[0127] The automation of scientific knowledge inevitably involves challenges such as intellectual property rights confirmation, cross-border flow of high-value scientific research data, and in-depth ethical issues related to science and technology. This embodiment provides deterministic digital cryptographic evidence for accident liability determination by constructing a "full lifecycle dynamic responsibility tracing and dual-signature log network."
[0128] I. Double-signature pre-verification in core AI experimental steps and human involvement in the process
[0129] Before any high-risk embodied action is performed in physical space, the responsibilities and rights must be clearly defined within the system. This embodiment constructs a dual-signature system using cryptographic techniques:
[0130] The first level of signature is the "AI and Sandbox System Pre-Approval Signature". After the system plans a sequence of actions and passes the edge physics sandbox simulation verification, the system will use a hash algorithm to package the current 3D environment map data, the contextual prompts of the large model, the generated policy confidence probability, and the certificate of simulation success, and add a timestamp. This constitutes the system's technical declaration that "under the current cognitive and physical model predictions, the action is absolutely safe".
[0131] The second level of signature is a "human expert or hard boundary control confirmation signature." For actions within low-risk zones or those rigorously verified by OSCBF, the underlying security controller issues a certificate for automated countersigning confirmation. For response processes involving extreme thermodynamic risks, the synthesis of toxic substances, or potential violations of scientific ethics, the system will forcibly trigger a pause, requiring the principal investigator (PI) holding the system's private key to provide informed consent and execute a human signature. This ensures that decision-making in cutting-edge medical and scientific fields always adheres to the principle of "human decision-making first."
[0132] II. Unalterable logs and precise division of responsibilities
[0133] All execution instructions with double signatures, along with real sensor snapshots of their actual physical feedback—including actual motor current, joint encoder deviation, LiDAR point cloud snapshots, and even changes in the transmissivity of the reactor liquid surface—will be continuously written to the blockchain via high-frequency data streams. In the event of instrument damage or personal injury, legal and forensic institutions can directly extract the logs to reconstruct the accident scene at millisecond levels.
[0134] (1) Defining the responsibility of algorithm developers / large model suppliers: If the log decoding shows that the AI output a dangerous liquid addition command based on a wrong understanding, and coincidentally bypassed the physical sandbox defense at the time due to cognitive defects, the responsibility will be clearly traced back to algorithm design defects or bias in the original pre-training data.
[0135] (2) Delineation of robot manufacturer (hardware) responsibility: If the log shows that the double signature is working correctly and the OSCBF has accurately issued emergency stop or depressurization commands, but fails to respond on time due to mechanical failure of the servo motor, valve aging and jamming, or failure of the hardware safety light curtain, it is a typical manufacturing defect or maintenance deficiency, and the hardware manufacturer shall bear the responsibility.
[0136] (3) User / Operator Responsibility Determination: If the logs clearly record that the robot hardware gate issued multiple over-authorization warnings, but the experimenter forcibly issued an Override command using senior administrator privileges, or physically shorted the fume hood interlock safety door switch for ease of operation, then the accident is entirely attributable to the improper operation of the human user.
[0137] Example 7: Independent Security Monitoring, Tiered Authentication, and Regulatory Sandbox
[0138] This embodiment provides a detailed explanation of the system's governance and supervision mechanisms.
[0139] Regulation of embodied intelligence in science cannot be a one-off deal, because cloud updates of AI models and aging of underlying hardware will cause the system's behavior to evolve dynamically.
[0140] I. Establish an Independent AI Safety System
[0141] In addition to the main control business logic, this architecture deploys a separate AI safety monitoring system in parallel. This system does not participate in experiment planning, but continuously monitors the behavior of the operating AI, detects anomalies and spikes in uncertainty, and proactively enforces conservative physical interventions when dangerous trends are identified.
[0142] II. Tiered Compliance Certification
[0143] A tiered certification system of "basic," "advanced," and "professional" levels is implemented based on the degree of laboratory hazard. R&D companies are required to establish internal ethics review committees and have annual ethics and safety compliance audits conducted by independent third parties with IEC 61010-1 testing qualifications.
[0144] III. Regulatory Sandbox Pilot Mechanism
[0145] To prevent rigid, one-size-fits-all regulations from stifling AI4S innovation, this architecture proposes using a regulatory sandbox as an "intermediate buffer" before new intelligent agents enter real-world wet laboratories. The regulatory sandbox allows newly developed embodied intelligence systems to be experimentally deployed in limited and controlled physical environments. Regulatory authorities continuously monitor high-risk behaviors and social impacts through interfaces and dynamically adjust regulatory rules based on real-world physical feedback, thereby providing a practical Chinese solution for the formulation of global governance rules for scientific embodied intelligence.
[0146] Example 8: System Example
[0147] This embodiment provides a scientifically embodied intelligent security gating and dynamic accountability system based on physical simulation and control obstacle functions, including:
[0148] The cognitive planning module is used to acquire experimental environment data through multimodal sensors and generate a sequence of action instructions to be verified.
[0149] The edge physics simulation module is used to embed physical process constraints into the forward inference process using a physics simulation proxy model, and to perform forward state inference verification on the action command sequence.
[0150] The safety constraint correction module is used to define a control barrier function based on a safety contract, construct a safety envelope set that satisfies forward invariance, construct constraints based on the Lie derivative condition of the control barrier function, and perform minimal intrusive correction on the nominal control command through quadratic programming to generate a safety control command.
[0151] The hardware security gating module, which is set independently of the main control system, is used to perform compliance verification on the security control commands and trigger hardware-level security actions when a physical quantity is detected to exceed a preset security threshold.
[0152] The dynamic responsibility traceability module is used to package the execution data of the entire chain using a hash algorithm, add a timestamp for double signature authentication, and write the signed execution data and the corresponding physical feedback data into the blockchain.
[0153] In this embodiment, the specific implementation methods of each module are described in Embodiments 2 to 6.
Claims
1. A scientific embodied intelligence safety gate and dynamic accountability method based on physical simulation and control barrier function, characterized in that, Includes the following steps: Step 1: Acquire experimental environment data using multimodal sensors, and generate a sequence of action commands to be verified based on the experimental environment data; Step 2: Using a physical simulation proxy model deployed on edge computing nodes, embed physical process constraints into the forward state deduction process, perform forward state deduction verification on the action command sequence, and generate a security verification result; If the security verification result indicates a danger, the action command sequence is intercepted and a replanning is triggered; Step 3: For the verified action instruction sequence, define a control barrier function based on the safety contract and construct a safe envelope set that satisfies forward invariance; when the nominal control instruction attempts to exceed the safe envelope set, construct constraints based on the Lie derivative condition of the control barrier function, and perform minimal intrusive modification of the nominal control instruction through quadratic programming to generate a safe control instruction; Step 4: Perform compliance verification on the security control command through hardware security gating independent of the main control system. When a physical quantity is detected to exceed the preset security threshold, trigger a hardware-level security action. Step 5: Package the execution data from Steps 1 to 4 using a hash algorithm, add a timestamp, and perform hierarchical signature authentication. Write the signed execution data and the corresponding physical feedback data into the blockchain for dynamic accountability tracking.
2. The method of claim 1, wherein, Step 1 includes: Step 1.1: Obtain heterogeneous high-frequency data through binocular vision, infrared thermal imaging and spectral analysis, extract feature information in three-dimensional space, and predict the occupancy probability and mass attributes of voxels around the self-subject. Step 1.2: Encode real-time environmental data into scene image tuples, combine them with natural language experimental task instructions for reasoning, and generate a macroscopic action instruction sequence from coarse to fine. The action instruction sequence is only used as an experimental request to be verified and does not have the authority to directly drive hardware execution.
3. The method of claim 1, wherein, Step 2 includes: Step 2.1: Use a physical information neural network or a reversible normalized flow as the physical simulation proxy model; Step 2.2: Embed the chemical kinetics, fluid mechanics, and instrument energy consumption models as soft constraints of the differential equations into the forward propagation of the physical simulation proxy model; Step 2.3: Perform forward state extrapolation on the action command sequence within a millisecond time window to predict whether a center of gravity instability or thermal runaway side reaction will occur in future time steps; Step 2.4: If the predicted trajectory indicates that a danger will occur, generate a hard intercept signal to prevent the issuance of the action command sequence and trigger an abnormal feedback to request replanning.
4. The method of claim 1, wherein, Step 3 includes: Step 3.1: Bind a safety contract to each basic motion primitive, wherein the safety contract declares the allowed spatial boundaries, maximum liquid addition flow rate, maximum contact torque limit, and minimum human-machine safety distance; Step 3.2: Based on the security contract, define the operation space control barrier function and construct a security envelope set, wherein the security envelope set satisfies forward invariance; Step 3.3: When the nominal control command attempts to exceed the safe envelope set, the nominal control command is modified in a minimally invasive manner by solving a quadratic programming problem in real time, so as to constrain dangerous actions to the safe envelope boundary while maintaining task continuity.
5. The method of claim 4, wherein, In step 3, the robot's dynamic model has an affine form: wherein, represents the current state of the robot, including joint angles, position, and velocity; represents the rate of change of the state over time; represents the control input; represents the system's own inherent dynamics; represents the control system matrix; The security envelope set From continuously differentiable functions definition, If and only if The control barrier function satisfies the following Lie derivative condition: in, Represents the maximum set of instructions that the hardware can execute; This indicates the rate at which the safe distance changes with the system's natural inertia; This indicates the rate at which the safe distance changes with control intervention; For a strictly increasing function, the closer to the danger boundary, the more... As the value approaches 0, the closer the right side of the equation is to 0, the more tightly the upper limit of the system's speed in the direction of danger is locked when it is forced to move closer to danger. The quadratic programming problem is: in, This refers to the nominal control command; This indicates the final safety control instructions issued to the implementing agency; This is the constraint condition obtained by transforming the Lie derivative condition.
6. The method according to claim 1, characterized in that, Step 4 includes: Step 4.1: The hardware security gating system consists of a dual-channel independent security microcontroller with cross-monitoring function to eliminate common cause failures and meet the requirement that the hardware fault tolerance is greater than or equal to 1. Step 4.2: The safety sensor array is directly hardwired to the hardware safety gating, and the safety sensor array includes at least one of a safety light curtain, a leakage detection sensor, a volatile gas sensor, and a six-axis torque sensor; Step 4.3: When the quasi-static compressive force exceeds the preset biomechanical limit, or when unexpected vital signs break through the safety isolation zone in the workspace, the hardware safety gate ignores all upper-level instructions and triggers either a Category 0 safety action or a Category 1 safety action within milliseconds; wherein, the Category 0 safety action is to cut off the servo power, and the Category 1 safety action is to control the emergency stop.
7. The method according to claim 1, characterized in that, Step 5 includes: Step 5.1: Perform the first-level signature, namely the AI and sandbox system pre-signature: Package the current 3D environment map data, contextual prompts, generated policy confidence probability, and simulation pass certificate using a hash algorithm and add a timestamp; Step 5.2: Perform the second-level signature, namely, human expert or hard boundary control confirmation signature: For actions that have been securely verified, the underlying security controller issues a certificate for automated countersigning confirmation; for operations that exceed the preset risk level threshold, a forced hover is triggered, requiring the researcher holding the system private key to manually confirm and authorize the signature through private key verification. Step 5.3: The execution instructions with dual signatures and the accompanying physical feedback data are continuously written into the blockchain via high-frequency data stream; the physical feedback data includes at least one of the following: actual motor current, joint encoder deviation, lidar point cloud snapshot, and change in the transmittance of the liquid surface in the reactor.
8. The method according to claim 7, characterized in that, The dynamic accountability tracing includes locating the cause of the incident based on the immutable logs in the blockchain: If the log decoding shows that the sequence of action instructions is dangerous and is not intercepted by the physical simulation agent model, then the cause of the accident is identified as a defect in the algorithm generation. If the log shows that the dual signature is working correctly and the security control command has been issued correctly, but the execution mechanism fails to respond on time, the cause of the accident is identified as a hardware execution failure. If the log records show that the hardware security gating has issued multiple unauthorized warnings, but there are records of overriding commands with administrator privileges or physical short-circuit records of safety interlock switches, then the cause of the accident is determined to be human intervention.
9. A scientifically embodied intelligent safety gating and dynamic accountability traceability system based on physical simulation and control obstacle functions, characterized in that, include: The cognitive planning module is used to acquire experimental environment data through multimodal sensors and generate a sequence of action instructions to be verified. The edge physics simulation module is used to embed physical process constraints into the forward inference process using a physics simulation proxy model, and to perform forward state inference verification on the action command sequence. The safety constraint correction module is used to define a control barrier function based on a safety contract, construct a safety envelope set that satisfies forward invariance, construct constraints based on the Lie derivative condition of the control barrier function, and perform minimal intrusive correction on the nominal control command through quadratic programming to generate a safety control command. The hardware security gating module, which is set independently of the main control system, is used to perform compliance verification on the security control commands and trigger hardware-level security actions when a physical quantity is detected to exceed a preset security threshold. The dynamic responsibility traceability module is used to package the execution data of the entire chain using a hash algorithm, add a timestamp for double signature authentication, and write the signed execution data and the corresponding physical feedback data into the blockchain.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.
