A layered event-driven perception-response architecture for physical world agents, chip, system and method

Abstract

The application relates to a layered event-driven perception-response architecture for physical world agents, a chip, a system and a method, and belongs to the field of intelligent robot control. The architecture comprises a nanosecond-level L1 pure hardware fuse layer, a microsecond-level L2 storage-computing integrated instinct mapping layer and a millisecond-level L3 hardware differential equation deduction layer. The three layers are cooperated through independent hardware channels, and the system is in a dormant state and reduces power consumption when there is no physical event. A double-layer structure of a general L2 rule library and an individual system parameter vector is matched to generate a physical parameter library, and an automatic generation scheme of a rule library based on a digital model of a physical law is provided. The defects of high response delay, large continuous power consumption, poor environmental adaptability and software interference of safety protection of an existing architecture are solved, and the application is suitable for intelligent agents such as biped robots and industrial robot arms which need real-time physical interaction.

Description

Technical Field

[0001] This invention belongs to the field of intelligent robot control, specifically relating to a hierarchical event-driven perception-response architecture, chip, system, and method for intelligent agents in the physical world. Background Technology

[0002] Currently, embodied intelligent robots (such as legged robots, industrial robotic arms, and autonomous vehicles) mainly rely on a serial computing paradigm of "sensors → analog-to-digital conversion → general-purpose processors (GPUs / CPUs) → neural network inference → actuators." Although this paradigm has made some progress in structured environments, it still suffers from the following structural defects when facing unstructured physical environments, and these defects cannot be fundamentally solved by simply increasing computing power:

[0003] First, there is a mismatch between response latency and the timescale of physical events. The occurrence windows of critical events in the physical world, such as collisions, slips, and falls, are typically on the order of microseconds to milliseconds, while the end-to-end inference latency of traditional neural networks is generally on the order of tens to hundreds of milliseconds. Limited by a fixed-period sampling mechanism, traditional agents are in a "blind" state between two sampling intervals, making decisions based on outdated state information. By the time the computation result is output, the response window of the physical event has often closed, preventing the agent from truly existing within the time flow of the physical world.

[0004] Second, the energy consumption bottleneck caused by continuous polling. Traditional perceptual computing pipelines perform continuous sampling and processing based on a fixed clock cycle. Regardless of whether the physical scene changes, the computing system maintains full speed. The frequent transfer of data between memory and processor consumes the vast majority of the total energy, severely limiting the agent's endurance and continuous working time.

[0005] Third, there is a contradiction between the storage of control parameters and physical uncertainties. Existing control systems typically store precise numerical thresholds as control parameters. When actual environmental parameters deviate from preset values ​​due to changes in friction coefficients, load fluctuations, or other factors, the system needs to be recalculated. Before the calculation is completed, the risk of loss of control is extremely high. Furthermore, existing robots primarily undergo kinematic parameter calibration (such as DH parameter identification) or sensor calibration before leaving the factory, lacking systematic calibration of physical interaction parameters. For example, key physical parameters such as joint friction-velocity curves, motor torque-current mapping, and foot contact force-deformation relationships are usually not systematically identified and stored at the factory. Simultaneously, existing technologies lack a "personal coefficient" mechanism to characterize the differences in physical parameters between different individuals of the same model. Individual differences caused by manufacturing tolerances are mostly handled through online compensation during operation, rather than being quantified into reusable coefficients at the factory. Simulation technology is also mostly used for training control strategies rather than generating an initial physical rule base, resulting in physical parameter identification often lagging behind product delivery and relying on on-site user debugging.

[0006] Fourth, the delays and limitations of safety protection mechanisms. Existing hardware comparators are mostly used for overcurrent protection, typically monitoring only the absolute threshold of a single sensor signal. They lack the ability to receive multimodal sensor signals and perform dynamic constraint judgments based on physical laws (such as energy conservation and momentum conservation). Furthermore, safety protection loops often communicate with the main control system via a shared data bus or software protocol stack. Their response delay is affected by bus arbitration and software instruction cycles, making it impossible to achieve physically unbypassable nanosecond-level protection.

[0007] Fifth, there is a lack of coordination among heterogeneous computing resources. Although independent hardware comparators, in-memory computing chips, and FPGA hardware solvers exist in existing technologies, there is still no system solution that integrates these three components heterogeneously in a hierarchical manner according to the time urgency of physical events and coordinates them through a unified event-driven bus.

[0008] In summary, the problems of high response latency, high continuous power consumption, poor adaptability to unknown physical scenarios, security protection affected by software latency, and lack of systematic physical law parameter library in factory configuration are the fundamental reasons restricting the safe and efficient deployment of embodied intelligence in unstructured real-world environments. Summary of the Invention

[0009] The purpose of this invention is to address the structural defects of existing serial computing paradigms for embodied intelligent robots, such as response latency mismatch with physical events, high energy consumption of continuous polling, inability of precise thresholds to adapt to physical uncertainties, interference of security protection by software latency, lack of a systematic physical law parameter library at the factory, and lack of heterogeneous computing resource collaboration. This invention provides a layered event-driven perception-response architecture, chip, system, and method for intelligent agents in the physical world. Through hardware layered heterogeneous integration, event-driven scheduling, regionalized expression of physical laws, two-layer calibration of the factory parameter library, and individual-group parallel evolution, this invention solves the problems that cannot be overcome by simply increasing computing power at the hardware level.

[0010] To achieve the above objectives, the technical solution of the present invention is: a hierarchical event-driven perception-response architecture for intelligent agents in the physical world, comprising:

[0011] The first response layer L1 is a pure hardware logic circuit consisting of an analog comparator and a power drive circuit directly connected. Its input is connected to the output of the force / torque sensor and the inertial measurement unit of the intelligent agent, and its output is connected to the enable terminal or the current limiting terminal of the corresponding motor driver.

[0012] The second response layer L2 is a hardware mapping circuit composed of a memory computing array. Its input is connected to an event-driven sensor and an analog signal output sensor. The memory computing array stores a stable domain boundary defined by multi-dimensional parameters, which is expressed in the form of a set of input-output electrical parameters.

[0013] The third response layer L3 consists of a hardware-based differential equation solver, which is connected to L2 via an independent hardware interrupt signal line. Under normal circumstances, it is in a sleep state with clock gating and power gating.

[0014] The original sensor signal enters L1 and L2 simultaneously. L1 performs protection judgment independently of L2 and L3. When L2 matches successfully, the wake-up signal of L3 is not triggered.

[0015] Furthermore, the core architecture of this invention is divided into three heterogeneous response units based on the time urgency of physical events. Each layer uses a computing medium that matches the response time and collaborates through an event-driven mechanism.

[0016] The first response layer, L1, is a pure hardware logic circuit consisting of an analog comparator and a power drive circuit directly connected. Its inputs are directly connected to the analog outputs of the joint force / torque sensors and the inertial measurement unit (IMU) of the intelligent agent; its outputs are directly connected to the enable or current limiting terminals of the corresponding joint motor driver, without analog-to-digital conversion, instruction cycles, or software intervention. The L1 storage element contains preset physical safety boundary thresholds such as maximum joint torque, maximum angular velocity, minimum ground clearance, and maximum acceleration. When the sensor signal exceeds the threshold, L1 directly triggers a motor drive protection command within nanoseconds.

[0017] The second response layer, L2, consists of a hardware mapping circuit formed by a memory-based computing array. Its inputs are connected to the asynchronous event stream outputs of event-driven sensors (such as event cameras and piezoelectric tactile sensors) and the outputs of analog signal output sensors. Each physical law mapping parameter stored in the memory-based computing array is not a single numerical threshold, but rather a stable domain boundary defined by multiple dimensions. This is ultimately compiled and burned into a set of input-output electrical parameters, including: the input current range after sensor signal conversion, the output current range corresponding to the actuator drive signal, and the conductance matrix or lookup table in the memory-based computing array that implements the input-output mapping. For example, the stable domain corresponding to "stable walking on wet and slippery surfaces" is jointly defined by the lower limit of the plantar friction coefficient, the upper limit of stride length, the upper limit of the forward shift of the center of gravity, the allowable offset range of the plantar pressure center, the ankle joint stiffness adjustment range, and the hip joint torque reduction coefficient.

[0018] The third response layer, L3, consists of a hardware-based differential equation solver. It is connected to L2 via a hardware interrupt signal line independent of the data bus and shared memory. Under normal circumstances, it is in a clock-gated and power-gated sleep state, and only starts when L2 triggers a wake-up signal. After being woken up, L3 performs sparse sampling on the current input event stream and explores new parameter combination points in the parameter space near the known stability region boundary.

[0019] Signal flow logic: The original sensor signals enter L1 and L2 simultaneously and in parallel. L1 performs protection judgment independently of L2 and L3. L2 completes the stability domain matching in microseconds. When the matching is successful, the L3 wake-up signal is not triggered, and the drive parameters are directly output. L3 is only woken up through an independent hardware interrupt line when L2 fails to match or when there is an output conflict.

[0020] Furthermore, the L2 internally includes multiple parallel matching channels pre-splitting according to physical interaction type, a cerebellum hardware weighted summation module, and a hardware protection zone; the parallel matching channels include visual optical flow channels, plantar contact channels, grasping channels, and trunk posture channels; the hardware protection zone has independent analog signal input pins and independent physical signal output pins, the independent analog signal input pins are directly connected to the analog output terminals of each joint sensor, and the independent physical signal output pins are directly connected to the enable and current limiting terminals of each joint motor driver, and the signal transmission path does not pass through analog-to-digital conversion and data bus.

[0021] Furthermore, the parallel matching channels are pre-splittered into four independent physical channels based on the physical interaction type. Each channel has independent input pins, in-memory compute subarrays, and output circuitry. The layout design uses physical isolation between word lines and bit lines to avoid crosstalk.

[0022] Visual optical flow channel: The input is a digital event stream from the event camera, which is converted into an internal parallel signal by the digital interface and then injected into the storage operator array. It stores visually related physical laws such as optical flow dilation rate-avoidance action and optical flow breakage mode-step gait mapping, and outputs motion intention parameters such as target movement direction, speed, and expected contact force.

[0023] Foot contact channel: The input charge signal from the foot piezoelectric tactile sensor is converted into a current signal by the charge amplifier and transconductance amplifier and then injected into the storage operator array to store the foot interaction law such as pressure change rate-gait adjustment and friction cone constraint, and output the foot contact deviation correction parameter;

[0024] Grip channel: Inputs charge signals from the palm piezoelectric tactile sensor, and processes them in the same way as the foot channel. It stores grip interaction rules such as slippage tendency-grip force compensation and contact vibration frequency-object elastic modulus mapping, and outputs grip deviation correction parameters.

[0025] Trunk attitude channel: The input is an analog voltage signal from an IMU with an analog output buffer. After being amplified by a buffer amplifier, it is injected into the storage operator array to store trunk attitude laws such as angular velocity change rate-attitude compensation and external force disturbance mode-joint stiffness adjustment. The output is trunk attitude correction parameters.

[0026] Cerebellum hardware weighted summation module: Multi-channel output fusion is achieved by summing currents in the analog domain: the motion intent parameters output by the visual channel are used as the reference current, and the deviation correction parameters output by the tactile and posture channels are used as the correction current. The two types of currents are naturally summed at the fusion node according to Kirchhoff's current law, and the final driving parameters are generated after current-to-voltage conversion. The weight coefficients of each channel can be dynamically adjusted by the upper brain according to the task scenario through the gain adjustment of the transconductance amplifier.

[0027] Independent hardware protection zone: It has a signal path completely independent of the main matching channel: the independent analog signal input pins are directly connected to the analog output terminals of the torque sensors, angular velocity sensors, angle sensors, IMU, and plantar pressure sensors of each joint. The signal does not go through analog-to-digital conversion, data bus, or preprocessing of the main matching channel; the threshold storage area is composed of one-time programmable (OTP) memory units or electronic fuses, storing hard safety thresholds related to energy conservation and momentum conservation; the independent physical signal output pins are directly connected to the enable and current limit terminals of the motor drivers of each joint. The constraint judgment of the protection zone runs in parallel and independently with L1, and the output signal has a higher priority than the drive parameters of the main matching channel.

[0028] Furthermore, L2 also includes a temporary rule cache area for storing parameter combination points successfully derived by L3; L3 internally includes a path management unit, 2 to 5 parallel derivation pipelines and a hardware constraint unit. After each integration step, the hardware constraint unit loads safety constraint parameters from the threshold storage area of ​​the L2 protected area and performs synchronous verification. When any path is determined to violate the safety constraint in any step, the derivation pipeline of that path is immediately terminated.

[0029] Furthermore, the L2 temporary rule cache, independent of the preset rule protection area and the OTA-updated group rule area, stores new parameter combinations successfully derived from L3, including triggering conditions, mapping relationships, and control parameters. It also stores visual experience (new mappings between event camera optical flow patterns and motion intentions) and tactile experience (new mappings between pressure change patterns and gait / grip corrections). This cache is only valid locally and is not shared with other agents via OTA. When the cloud distributes group consensus rules corresponding to temporary rules via OTA, the temporary rules are overwritten or merged by the cloud rules. Temporary rules that have not been invoked for a long time naturally decay in weight and are automatically cleared when their weight falls below a threshold.

[0030] L3 hardware simulation architecture: Internally includes a path management unit, 2-5 parallel simulation pipelines, a hardware constraint unit, and an evaluation and selection unit.

[0031] Path Management Unit: Controls the generation, termination, and resource reallocation of candidate paths. Upon activation, it first loads all safety constraint parameters (joint torque meltdown threshold, power limit, friction cone conservatism coefficient, support polygon boundary, kinetic energy-torque derating mapping table, etc.) from the L2 protected area.

[0032] Parallel simulation pipeline: Each pipeline independently contains a slope calculation unit and a state update unit. A fourth-order Runge-Kutta (RK4) pipeline architecture is used to numerically solve the ordinary differential equations of robot dynamics. The state space equation is dX / dt=f(X,U), which includes the translational dynamics equation of the center of mass (m·dv / dt=F_gravity+F_contact+F_external) and the rotational dynamics equation about the center of mass (I·dω / dt=τ_contact+τ_external-ω×(I·ω)). Among them, the state vector X contains the angles, angular velocities, position and velocity of each joint, U is the control input vector, i.e., joint torque; m is the total mass of the robot, v is the velocity vector of the center of mass, and the terms on the right side of the translational dynamics equation of the center of mass are the resultant forces of gravity, ground contact force and external force, respectively; I is the inertia tensor, ω is the angular velocity vector, and τ_contact and τ_external are the torque vectors of the corresponding forces.

[0033] Hardware constraint unit: After each integration step is completed, a safety check is performed synchronously instead of waiting for the entire path derivation to be completed: check in turn whether the joint torque is within the melting threshold, whether the centroid projection is within the supporting polygon, whether the plantar tangential force meets the friction cone constraint, and whether the joint power is within the power limit. If any item fails, the path derivation is terminated immediately, and the released computing resources are reallocated to new candidate paths.

[0034] Evaluation and selection unit: Calculate the comprehensive score J = w1·S + w2·E + w3·M for all candidate paths that pass the safety filter, where S is the attitude stability index (the degree to which the centroid projection deviates from the center of the supporting polygon), E is the energy consumption index (the integral of the square of the joint torque), and M is the joint torque margin index (the ratio of the peak joint torque to the rated torque). The default weights are w1=0.5, w2=0.2, and w3=0.3, which can be adjusted by the upper-level brain. Finally, the path with the smallest J value is selected as the optimal strategy output.

[0035] Furthermore, the analog comparator input of L1 is connected to the analog voltage output of the sensor, and the reference voltage is connected to the output of a resistor divider network configured by a one-time programmable memory cell or an electronic fuse. The comparator output directly drives the gate of the power MOSFET. The threshold parameter of L1 is stored in the one-time programmable memory cell, and after writing, it is permanently locked by the fuse through the fuse programming enable.

[0036] Furthermore, L1 is powered by an auxiliary power domain independent of the system's main power supply, ensuring that the system enters the working state immediately upon power-up. Its internal circuitry is divided into four levels:

[0037] Input buffer and filtering stage: The analog voltage output by the sensor first enters the voltage follower to provide high input impedance and avoid load effect; the output is divided into two paths, one of which goes directly into the comparator, and the other is filtered by a low-pass filter with a cutoff frequency matching the effective bandwidth of the sensor to remove high-frequency noise and prevent false triggering.

[0038] Threshold Comparison Stage: Employs a comparator chip with ultra-fast propagation delay, such as the ADI ADCMP580 / ADCMP581 series (typical propagation delay of 180 picoseconds) or the TI TLV3601 (propagation delay of approximately 2.5 nanoseconds). The inverting input of the comparator is connected to the filtered sensor signal, and the non-inverting input is connected to the reference voltage. The output level flips in nanoseconds when the signal exceeds the threshold.

[0039] Reference voltage generation and threshold storage stage (hardware protection zone): A stable reference voltage is provided by a high-precision bandgap reference voltage source and connected to a resistor voltage divider network configured by OTP or electronic fuse. After the threshold is written through the JTAG / SPI interface during factory calibration, the fuse programming enable fuse permanently locks the reference voltage. After the chip is packaged, the programming interface pin is internally pulled low or left floating and cannot be accessed from the outside. If the safety threshold needs to be modified, the chip hardware must be replaced.

[0040] Power drive and output stage: The comparator output is directly connected to the gate of the N-channel power MOSFET. The drain of the MOSFET is connected to the motor driver enable pin, and the source is grounded. During normal operation, the MOSFET is turned on and pulls the enable pin low. During a fault, it is turned off and pulls the enable pin high. The delay of the entire link is controlled in the order of nanoseconds to tens of nanoseconds.

[0041] The present invention also provides a chip that implements a hierarchical event-driven perception-response architecture for physical world intelligent agents as described above. The chip can be any physical form of a single chip, a multi-chip module, a system-in-package, or a board-level system.

[0042] Furthermore, the chip can be implemented in any of the following physical forms: a single chip integrating all three layers of functions, a multi-chip module (MCM), a system-in-package (SiP), or a board-level system. All forms follow the above-mentioned three-layer event-driven collaborative logic, and the independent interrupt lines, independent signal paths, and hardware protection zone isolation designs between each layer remain unchanged.

[0043] The present invention also provides an intelligent agent system, comprising:

[0044] The chip described above;

[0045] The sensor module includes at least an event camera and a piezoelectric tactile sensor, with its output connected to the input of chip L2.

[0046] The analog signal output type sensor module includes at least a joint torque sensor, an IMU with analog output buffer, and the output terminal is connected to the input terminal of chip L1 and the independent analog input pin of the L2 protection zone.

[0047] The actuator module includes at least a joint motor and its driver, with the enable terminal and current limiting terminal of the driver connected to the output terminal of chip L1 and the independent physical signal output pin of the L2 protected area, respectively.

[0048] Communication module: Connected to the chip, it is used to transmit data with the cloud server, including uploading physical anomaly event characteristics and individual experience rules, and receiving group consensus rules and updated general L2 rule bases issued by OTA.

[0049] This invention also provides a method for physical interaction using the hierarchical event-driven perception-response architecture for physical world intelligent agents described above, comprising:

[0050] The raw sensor signal enters both L1 and L2 simultaneously;

[0051] L1 independently determines whether the sensor signal exceeds the safety threshold within nanoseconds; if it does, it directly triggers the protection action.

[0052] L2 performs hardware matching of the current state parameters with the preset stable domain boundary within microseconds. If the current state parameters fall within the stable domain, it directly outputs the drive parameters and keeps L3 in sleep mode.

[0053] If the current state parameters exceed all known stability regions or a conflict occurs, L2 will wake up L3 via a separate hardware interrupt signal line;

[0054] L3 performs sparse sampling of the current event stream in milliseconds, loads safety constraint parameters from the L2 protected area, performs safety checks synchronously at each step of the simulation and immediately terminates unsafe paths, and selects the optimal path and outputs driving parameters through safety filtering and comprehensive scoring.

[0055] L3 stores the successful parameter combinations from this simulation into the temporary rule cache of L2.

[0056] Furthermore, the method follows the event triggering process described below:

[0057] The raw sensor signal enters L1 and L2 simultaneously. When there is no change in physical state, the event camera does not output data. L2 is in a silent listening state, and L3 remains in sleep mode. The total power consumption of the system approaches zero.

[0058] L1 independently determines whether the sensor signal exceeds the safety threshold within nanoseconds. If it does, it directly pulls down the enable terminal of the corresponding motor driver to cut off or limit the drive current without waiting for the processing results of L2 and L3.

[0059] L2 completes hardware matching of the current state parameters with the preset stability domain boundary within microseconds: The input signal is processed in parallel across multiple channels. After each channel outputs a suggested parameter range, a range comparison array is used to determine if the multi-channel outputs are consistent. If they are consistent, the current execution value is maintained; otherwise, the lower edge of the intersection is used for execution, and subsequently, the value is gradually increased to approximate the upper edge of the intersection based on a safety margin. For example, if the visual channel suggests a stride range of [28cm, 32cm] and the foot contact channel suggests a stride range of [25cm, 29cm], then 25cm is taken as the current execution value. After execution, if the safety margin is sufficient, it is gradually increased to 29cm. Upon successful matching, the driving parameters are directly output, and L3 remains in sleep mode.

[0060] If the current state parameters exceed all known stability regions, or the difference between the output parameters of different channels exceeds the preset conflict threshold, L2 wakes up L3 through an independent hardware interrupt signal line;

[0061] L3 performs sparse sampling of the current event stream within milliseconds, loads all safety constraint parameters from the L2 protected area, starts 2 to 5 parallel inference pipelines, and performs safety checks synchronously after each integration step. Unsafe paths are immediately terminated and resources are reallocated. All paths that pass the safety filter are selected as the optimal path and output driving parameters through comprehensive scoring.

[0062] L3 stores the successful parameter combination from this simulation into the temporary rule cache of L2 as local individual experience, and stores the feature vector and response strategy of this abnormal event in a non-volatile cache for later uploading to the cloud; after completion, L3 automatically returns to the dormant state.

[0063] This invention also provides a method for generating a physical law parameter library for an AI agent in the physical world at the time of its manufacture, including:

[0064] Establish the nominal physical parameter model of the corresponding intelligent agent;

[0065] Execute standardized test sequences in the physics simulation engine and generate a general physical law parameter library based on the simulation results;

[0066] Perform factory calibration tests on each intelligent agent, measure the deviation between the actual physical response and the simulation expected value, and quantify the deviation into an individual coefficient vector;

[0067] The general physical law parameter library is merged with the individual coefficient vector to generate a personalized physical law parameter library unique to the corresponding intelligent agent, and then burned into the local storage unit of the corresponding intelligent agent.

[0068] Furthermore, a two-layer structure of "general L2 rule base + individual coefficient vector" is adopted to decouple model universality from individual specificity. The steps are as follows:

[0069] Establish a nominal physical parameter model for this type of intelligent agent: Based on design parameters (joint motor specifications, reducer transmission ratio, link mass and inertia, foot / hand sensor characteristics, etc.), construct a nominal model containing complete physical interaction parameters, including joint friction force-velocity characteristic curve, motor current-output torque mapping relationship, foot contact force-deformation characteristic curve, hand gripping force-drive signal mapping relationship, and the correspondence between the overall center of mass position and attitude, etc.

[0070] Simulation generates a general L2 rule library: A digital twin model of this model is built in the physical simulation engine, and standardized test sequences covering the entire working range are executed (full-range excitation trajectory of each joint, contact test between the foot and different ground materials, grip test between the hand and different object materials, and static / dynamic balance test of the whole machine). Physical parameter mapping relationships are extracted from the simulation data to generate a general L2 rule library. The general library is divided into three categories: ① Invariant constraints (expression of energy conservation and momentum conservation in joint space), which are locked in the protected area OTP after factory calibration and do not participate in OTA updates; ② Parametric physical laws (friction cone constraints, center of gravity projection constraints, weight-torque relationship), which are mapped to the analog comparator reference current range of the protected area and are also locked and not updated via OTA; ③ Variable mode response templates (optical flow expansion-avoidance mapping, step geometry-gait mapping, etc.), which are stored in the main matching area in-memory subarray and can be updated via OTA.

[0071] Factory calibration generates individual coefficient vectors: Four types of calibrations are performed on the agent on a standardized test bench: ① Joint power calibration: Each joint is driven to perform a sinusoidal scanning motion, and the motor current and angular velocity are recorded to identify the actual power limit; ② Kinetic energy derating calibration: Each joint performs an emergency stop at different initial angular velocities to identify the individualized kinetic energy-torque derating mapping parameters; ③ Gravity term calibration: Each joint is slowly moved under no-load and rated load conditions to identify the individualized gravity term compensation curve; ④ Friction cone calibration: The sole of the foot is placed on various standard ground materials to measure the maximum static friction force in each direction to identify the individualized conservative friction coefficient. The ratio of the measured value to the simulated nominal value is used as the individual coefficient, and each coefficient corresponds to a parameter mapping relationship in a general library.

[0072] Generate a personalized L2 rule base: Multiply each parameter mapping relationship in the general L2 rule base by the corresponding individual coefficient to obtain a personalized rule base adapted to the physical characteristics of the intelligent agent, and burn it to the local storage unit;

[0073] Updates during operation: After the agent is put into operation, it continuously monitors the deviation between the actual physical response and the predicted value of the rule base. When the deviation exceeds the preset threshold, the corresponding individual coefficient component is updated. The updated individual coefficients are uploaded to the cloud periodically to optimize the general L2 rule base of this model and distributed to the entire group via OTA.

[0074] This invention also provides a method for generating a personalized L2 rule base for an AI agent in the physical world based on a digital model of physical laws, including:

[0075] A digital model of physical laws is constructed, which is trained by the AI ​​large model on a training set containing mathematical descriptions of classical physical laws, material physical property data and known physical experimental data.

[0076] Construct a target description library for the target intelligent agent, wherein each target parameter in the target description library includes an action type identifier, a range of target values, and a weight of target priority;

[0077] The standard parameter model of the target intelligent agent and the sensing parameter range of each L2 channel are input into the physical law digital model;

[0078] The physical law digital model, based on the joint input of the standard parameter model, the target description library, and the perception parameter range, selects the optimal solution from the candidate execution parameter schemes that satisfy the hard constraints through a multi-objective optimization function, and generates the corresponding L2 rule base parameters. The hard constraints include at least joint angle limit constraints, joint torque limit constraints, and center of gravity projection constraints.

[0079] The L2 rule base parameters are combined with the individual coefficient vector of the target agent to generate a personalized L2 rule base unique to the agent, which is then burned into the agent's local storage unit.

[0080] Furthermore, the method for generating a personalized L2 rule base for AI agents in the physical world automates the generation of the rule base through a digital model of physical laws. The steps are as follows:

[0081] Constructing a digital model of physical laws: Employing a Physical Information Neural Network (PINN) architecture, the model is trained on a training set containing mathematical descriptions of classical physical laws (Newtonian mechanics, friction laws, elastic collisions, fluid mechanics, etc.), material physical property data (friction coefficient, elastic modulus, density, etc.), descriptions in human physical interaction experience language, and known physical experimental data (force-displacement curves, velocity-time curves, etc.). The model comprises a neural network backbone, a physical information layer (encoding physical laws such as Newton-Euler dynamics equations into loss function constraints to ensure that the output satisfies physical consistency), and a decision layer (processing multi-objective optimization output).

[0082] Construct a target description library: covering target parameters for four types of actions. Each parameter includes an action type identifier, parameter name and unit, target value range (minimum, maximum, typical, allowable deviation), priority weight, and coupling constraints: ① Movement type: Walking (cadence 1.0~1.5Hz, stride length 0.3~0.5m, leg lift height 0.05~0.10m), Running (cadence 2.5~3.5Hz, stride length 0.6~1.0m, leg lift height 0.10~0.20m, airtime ≤0.1s), Climbing (cadence 0.8~1.2Hz, stride length 0.2~0.4m, trunk forward lean 10°~20°), Descending (cadence 0.8~1.2Hz, stride length 0.2~0.4m, trunk backward lean 5°~15°); ② Manipulation type: Grasping (end-effector arrival speed ≤0.5m / s, grip strength 0.5~5 0.0N, contact force accuracy ≤0.1N), placement (end descent speed ≤0.3m / s, release force ≤0.1N, positioning accuracy ≤1mm), lifting (speed 0.1~0.3m / s, lifting force 10~50N, posture deviation ≤1°); ③ Posture type: standing (center of gravity height design value ±0.01m, trunk verticality ≤0.5°, center of gravity projection at the center of the support surface ±0.02m), squatting (knee joint angle 60°~120°, trunk forward tilt ≤15°), sitting posture (uniform hip pressure, trunk verticality 0°±2°); ④ Response type: obstacle avoidance (deflection angle 10°~45°, acceleration 0.5~2.0m / s²), emergency stop (deceleration ≤3.0m / s², center of gravity offset after emergency stop ≤ support surface boundary -0.02m), posture recovery (recovery time ≤0.5s, deviation ≤0.5°);

[0083] Input the standard parameter model of the intelligent agent, including mechanical structure parameters (link length, mass, center of mass position, moment of inertia), actuator parameters (motor torque-speed characteristics, maximum torque, maximum speed, response time constant), transmission system parameters (reduction ratio, transmission efficiency, backlash), contact surface parameters (foot / hand contact area, material hardness, friction coefficient range), and sensor parameters (visual resolution / field of view / measurement range, tactile sensitivity / measurement range, IMU accuracy / range).

[0084] Input the range of sensing parameters for each L2 channel: including the visual channel (object category recognition, distance measurement range, velocity measurement range, field of view range), the tactile channel (force resolution range, pressure change rate range, material type, slip detection sensitivity), the inertial channel (attitude angle range, angular velocity range, acceleration range), and the force channel (joint torque range, end contact force range that can be measured).

[0085] Multi-objective optimization to select the optimal solution: Among all candidate solutions that meet the hard constraints, the physical law digital model selects the optimal solution through a multi-objective optimization function. The hard constraints include: joint angles within mechanical limits, joint torques within the rated range, joint velocities within the rated range, the center of gravity projection within the supporting polygon, and the plantar tangential force satisfying the friction cone constraint (|F_t|≤μ·F_n, where F_t is the plantar contact surface tangential force, F_n is the plantar contact surface normal force (normal pressure), and μ is the static friction coefficient). If any constraint is not met, the candidate solution is eliminated. The optimization dimensions include: ① Energy consumption optimization: Total energy consumption E_total=∫(ΣU_j·I_j)dt+∫f_friction(ω_j,τ_j)dt+P_cooling·t, where U_j is the driving voltage of the j-th joint motor, I_j is the phase current of the j-th joint motor, f_friction is the friction / damping loss function, ω_j is the angular velocity of the j-th joint, τ_j is the output torque of the j-th joint, and P_cooling is the power consumption of the thermal management / heat dissipation system. The goal is to minimize E_total; ② Stability optimization: Stability index Stability=α·d_ZMP+β·d_COM +γ·min(margin_j)-δ·Δθ, where d_ZMP is the distance between the zero-moment point and the center of the supporting polygon, d_COM is the distance between the centroid projection and the boundary of the supporting polygon, margin_j is the joint moment margin, Δθ is the attitude angle deviation, and α, β, γ, and δ are the corresponding weights. The objective is to maximize Stability; ③ Time efficiency optimization: Task execution time t_total, the objective is to minimize t_total, and the constraints are Stability ≥ minimum stability Stability_min and E_total ≤ maximum total energy consumption E_max; ④ Smoothness optimization: Smoothness = ∫(Σ(d3 θ_j / dt 3 ) 2 )dt,d 3 θ_j / dt 3 Let θ_j be the angular acceleration of the j-th joint, and θ_j / be the angular position of the j-th joint. The objective is to minimize Smoothness. The overall objective function is F = w1·E_total + w2·(1 / Stability) + w3·t_total + w4·Smoothness, with weights adjusted according to task type: Emergency tasks w1=0.15, w2=0.25, w3=0.50, w4=0.10; Long-term inspection w1=0.50, w2=0.30, w3=0.05, w4=0.15; Precision operation w1=0.15, w2=0.40, w3=0.05, w4=0.40; Emergency obstacle avoidance w1=0.05, w2=0.55, w3=0.30, w4=0.10. Optimization algorithms can include Sequential Quadratic Programming (SQP), Particle Swarm Optimization, NSGA-II, or Linear Quadratic Regulators.

[0086] Generate and burn a personalized L2 rule base: Optimize the output and the motor drive values ​​(torque value, position value, speed value) corresponding to each group of perception inputs, and combine them with the individual coefficient vector to generate a personalized L2 rule base, which is divided into a main matching area (input parameter range - output execution parameter mapping table) and a protection area (safety boundary parameters - protection trigger conditions), and burn it to the agent's local storage unit.

[0087] This method also includes the following preferred designs: ① Offline pre-computation: The physical law digital model pre-computes the entire working space of the agent in the offline stage to generate a global optimization mapping table. Parameters can be obtained directly from the table at runtime without online optimization; ② Uncertainty output: When generating the rule base, the uncertainty range or confidence interval of each parameter is output simultaneously to guide the initial step size setting for subsequent individual coefficient fine-tuning and avoid exceeding the safety boundary; ③ Multi-environment adaptation: Working environment parameters (temperature, humidity, altitude, etc.) are received to generate multiple versions of the rule base adapted to different environments. The agent automatically switches when entering the corresponding environment.

[0088] This invention also provides a physical world AI agent personalized L2 rule base generation system for realizing a method for generating a personalized L2 rule base for a physical world AI agent, comprising:

[0089] The system consists of the following modules, each with a function corresponding to the steps described above:

[0090] Physical law digital model generation module: used to construct physical law digital models based on the above PINN architecture, and load training datasets to complete model training;

[0091] The target description library storage module is used to store target description parameters for movement, operation, posture, and response actions, and supports calling the corresponding parameters according to the task type.

[0092] Intelligent agent parameter input module: used to receive the standard parameter model of the target intelligent agent and the range of sensing parameters of each L2 channel, and input the parameters into the physical law digital model;

[0093] Multi-objective optimization module: used to load hard constraints, execute multi-objective optimization algorithms, and select the optimal L2 rule base parameters from candidate solutions that meet the constraints;

[0094] Personalized rule base generation module: used to combine L2 rule base parameters with individual coefficient vectors to generate a personalized L2 rule base that includes the main matching area and the protection area;

[0095] Rule base burning module: Used to burn personalized L2 rule bases to the local storage unit of the agent, supporting batch burning and single-unit burning.

[0096] Compared with the prior art, the present invention has the following beneficial effects:

[0097] This invention addresses the structural deficiencies of existing technologies at their root through mechanisms such as a fully hardware-layered design, event-driven scheduling, regionalized representation of physical laws, and dual-layer parameter library calibration. The specific effects are as follows:

[0098] 1. Response delay matches physical time flow: The delay of the L2 perception-decision link is compressed to the microsecond level, and thousands of parameter corrections can be completed within a few milliseconds window of a single motor action, eliminating the sampling blind control period of the traditional serial paradigm. The intelligent agent can synchronize with the physical reality in real time.

[0099] 2. Significantly reduced energy consumption: Following the principle of "no calculation without event change", the sensor does not output data when there is no physical change, L2 is silent and L3 is in sleep mode, and the total power consumption of the system approaches zero, solving the energy consumption bottleneck of continuous polling;

[0100] 3. Strong adaptability to physical uncertainty: L2 adopts multi-dimensional stable domain matching instead of single threshold judgment, without the need for precise calculation of environmental parameters, only the judgment of whether the state is within the safe domain; the temporary rule cache realizes individual evolution of "learning upon first collision and instinct upon second encounter", and cloud-based group evolution realizes continuous optimization of the rule base, adapting to unstructured environments;

[0101] 4. Physical safety protection cannot be bypassed: L1 adopts pure hardware direct connection to achieve nanosecond-level fuse protection, L2 protection zone has an independent signal path and dynamically executes energy and momentum conservation constraints, L3 deduction process step-level safety verification, the three layers of protection are completely independent of the software system, avoiding the delay risks caused by bus arbitration and instruction cycle;

[0102] 5. Standardized factory configuration: Through a two-layer structure of "general L2 rule base + individual coefficient vector", the identification of physical parameters is moved from on-site debugging to the factory stage. Each intelligent agent has an initial instinctive rule base that matches its own hardware characteristics when it leaves the factory, which solves the problem of insufficient compensation for individual differences.

[0103] 6. Low engineering implementation cost: The rule base generation method based on physical law digital model realizes automated and standardized output, without relying on human experience, and has zero runtime dependency. Compared with general models such as NVIDIA Cosmos and DeepMind GeminiRobotics, it focuses more on the generation of executable rules for specific intelligent agents, resulting in lower deployment costs. Attached Figure Description

[0104] Figure 1 This is a schematic diagram of the overall structure of the hierarchical event-driven perception-response architecture for intelligent agents in the physical world according to the present invention.

[0105] Figure 2 This is a flowchart of the signal triggering process between the three layers L1, L2, and L3.

[0106] Figure 3 This is a structural diagram of the L2 internal multi-channel parallel matching, the cerebellar hardware weighted summation module, and the independent input-output pathways of the protected area.

[0107] Figure 4 A schematic diagram illustrating the layered storage content of an L2 in-memory SRAM array.

[0108] Figure 5 This is a schematic diagram of an embodiment of the multi-chip module of the present invention.

[0109] Figure 6 This is a schematic diagram of the L1 single-channel signal processing link circuit.

[0110] Figure 7 This is a diagram of an L1 multi-channel parallel architecture and a priority arbitration circuit.

[0111] Figure 8 This is a schematic diagram of an embodiment of the chip of the present invention in the adaptive gait control of a legged robot.

[0112] Figure 9 This is a pipeline architecture diagram for the L3 hardware-based differential equation solver.

[0113] Figure 10 The schematic diagram of the L3 hardware constraint unit and evaluation selection unit circuit.

[0114] Figure 11 This is a schematic diagram of the system for generating a database of physical law parameters for factory shipments and for individual calibration. Detailed Implementation

[0115] The following is in conjunction with the appendix Figures 1 to 11 The technical solution of the present invention will be described in detail below.

[0116] Example 1: Gait Adaptive Control of a Legged Robot on Unknown Terrain

[0117] like Figure 1 As shown, the robot's piezoelectric tactile sensors on its feet and its visual event camera on its head perceive changes in the physical world in real time. When the robot approaches a step, the visual event camera captures the longitudinal break pattern of the optical flow field. This event flow enters the L2 visual optical flow channel, where the pre-set step geometry feature-gait parameter mapping relationship is activated at the hardware level. Specifically, the event flow is converted by a digital interface and enters the word line of the in-memory computing subarray. The word line voltage is multiplied by the stored conductance lookup table in the analog domain. The bit line output current is summed by a current adder to generate motion intention parameters—including leg lift height and forward shift of the center of gravity. When the foot actually contacts the step surface, the piezoelectric tactile sensor detects a slight deviation between the actual contact force and the visually predicted expected contact force. This tactile signal is converted into a current signal by a charge amplifier and a transconductance amplifier, and then enters the in-memory computing subarray of the L2 foot contact channel. The bit line output current is processed by a differential amplifier to generate foot contact deviation correction parameters. After successful L2 rule matching, the visual and tactile channels each output a suggested parameter range, rather than a single value. For example, the visual channel suggests a stride range of [28cm, 32cm], while the plantar contact channel suggests a stride range of [25cm, 29cm] based on the current friction. The system first uses a range comparator to determine if the two are inconsistent, taking the lower edge of the intersection, 25cm, as the current stride execution value. After execution, the tactile system continuously monitors, and if the safety margin is sufficient, gradually increases the stride to approach 29cm. After the parameters are coordinated and the specific execution values ​​are determined, the cerebellar hardware weighted summation module uses the visual motor intention as a benchmark and the tactile correction parameters as correction values, and generates the final drive command in the analog domain through current-weighted summation. Figure 2 As shown, the perception-decision process is completed in microseconds, without the need for any CPU or AI model intervention.

[0118] Taking a direct-drive torque motor as an example, the complete closed-loop response time is approximately 5 to 20 milliseconds. Compared to the traditional solution (GPU inference latency of 30 to 100 milliseconds) plus the same motor execution latency, resulting in a total response time of 35 to 120 milliseconds, this invention reduces the total response time by approximately 2 to 6 times. More importantly, within the 10-millisecond time window during which the motor completes one response, L2 has already completed thousands of state checks and parameter corrections, achieving continuous control of "execution and correction simultaneously," completely eliminating the blind control state of tens of milliseconds between two perception samples in the traditional solution. In the traditional solution, if the robot suddenly slips, it may have already moved forward 4 to 12 centimeters during the blind control period, and the fall will be irreversible; in this solution, the correction command is issued before the robot moves forward less than 0.001 millimeters, preventing the fall before it even occurs.

[0119] If the robot encounters a ground material it has never seen before, none of the channels in L2 can match the current plantar pressure change pattern with any preset stability domain. L2 outputs a hardware interrupt to wake up L3. L3 performs sparse sampling of the current event stream, extracts key features, loads safety constraint parameters from the L2 protected area, and infers multiple gait adjustment strategies in parallel in the parameter space near the known stability domain boundary. Each candidate path is synchronously verified for safety constraints by the hardware constraint unit after each step of integration; unsafe paths are immediately terminated. The paths that pass the safety filter are selected as the optimal strategy through comprehensive scoring. L3 stores the new parameter combination that was successfully inferred in this instance into the temporary rule cache of L2 as its individual experience. When the robot walks on the same terrain again, L2 prioritizes retrieving the individual experience rule from the temporary cache and directly outputs the appropriate gait parameters without waking up L3 again. At the same time, L3 stores the feature vector of this abnormal event and the executed strategy, which is uploaded to the cloud by the local brain when idle. The cloud-based AI system refines existing stable domain boundaries from massive amounts of anomalous data uploaded from various terminals, or extracts the shared structure of multiple related stable domains into a more general stable domain skeleton, which is then distributed to the entire population's L2 level via OTA (Over-The-Air). Figure 8 As shown, the temporary rules of the local machine are subsequently overridden or merged by the group consensus rules issued by the cloud, completing the evolutionary loop from individual experience to group consensus.

[0120] Example 2: Specific Implementation Methods for Enforcing Energy and Momentum Conservation Constraints in the L2 Protected Area

[0121] In this embodiment, the protected area of ​​the L2 Instinct Mapping Core not only stores static safety thresholds, but also includes independent analog signal input pins, a dedicated analog comparator array, and an in-memory computing subarray, which are used to perform dynamic judgments of energy conservation constraints and momentum conservation constraints in microseconds.

[0122] The protected area has an independent input signal source. The protected area connects directly to the following sensor outputs via independent analog signal pins, without preprocessing through analog-to-digital converters, data buses, or the main matching channel: analog voltage outputs of torque sensors for each joint motor; analog voltage outputs of angular velocity sensors for each joint (preferably using tachogenerators or rotary transformer decoders with analog outputs); analog voltage outputs of angle sensors for each joint (preferably using potentiometer-type absolute angle sensors or magnetic encoders with analog outputs); analog outputs of attitude angles, angular velocities, and accelerations from the inertial measurement unit (using IMU models with analog output buffers); and analog voltage outputs of plantar pressure sensors.

[0123] The energy conservation constraint is enforced as follows: The energy conservation subarray within the protected area contains multiple sets of analog multipliers and analog comparators. The multipliers multiply the joint torque signal and the joint angular velocity signal to obtain the current signal corresponding to the real-time mechanical power. The comparators compare this real-time power current with a continuous current range consisting of the upper and lower reference currents of the OTP (Over-The-Point) reference current. When the real-time power current falls within the range, the system operates normally; when the real-time power current exceeds the upper reference current, the comparator outputs an over-limit signal, directly pulling down the current limit terminal of the corresponding motor driver through an independent physical circuit, limiting the motor current within a safe range. Simultaneously, the kinetic energy-torque derating mapping table outputs a derating coefficient based on the absolute value of the joint angular velocity through an in-memory lookup table. This coefficient is applied to the maximum allowable torque value, achieving automatic torque derating under high-speed motion.

[0124] The execution method of momentum conservation constraints. The momentum conservation subarray of the protected area contains multiple sets of analog comparators to perform the following constraint judgments: The maximum joint torque comparator compares the real-time torque signal with the torque fuse threshold stored in the OTP. If the threshold is exceeded, the enable terminal of the motor driver is directly pulled low; The friction cone constraint comparator compares the plantar tangential force signal with the product of the normal force and the conservative friction coefficient. If the threshold is exceeded, the attitude adjustment signal is triggered; The support polygon boundary comparator calculates the centroid projection position based on the attitude angle provided by the IMU and compares it with the support polygon boundary stored in the OTP. When the centroid projection is close to the boundary, the attitude correction signal is output.

[0125] The constraint judgment of the protected area operates independently and in parallel with the L1 instinctive fuse layer. L1 is responsible for the most urgent nanosecond-level hardware protection (such as absolute maximum torque exceeding limits), while the protected area is responsible for microsecond-level constraint judgment that requires dynamic calculation (such as power limit, kinetic energy derating, and friction cone constraints). The output of the protected area and the output of L1 have the same physical direct connection characteristics, both directly connected to the enable and current limit terminals of the motor driver. When the constraint signal of the protected area conflicts with the drive parameters of the L2 main matching channel, the priority arbitration circuit inside the motor driver ensures that the protection signal takes precedence over the drive parameters.

[0126] A concrete example. Consider a robot's knee joint experiencing an accidental lateral impact. At the moment of impact, the joint torque sensor detects a sharp increase in torque. This signal enters the momentum conservation subarray of the L2 protected area via an independent analog circuit. Within microseconds, the maximum torque comparator determines that the torque is approaching the fuse threshold stored in the OTP (Override Tolerance) and immediately pulls down the enable terminal of the knee joint motor driver via an independent physical circuit, cutting off the drive current. Simultaneously, the L2 main matching channel may be processing visual or tactile event streams and has not yet detected the impact. The direct-connect protection loop of the protected area independently completes the safety protection action without relying on the main channel or any software.

[0127] Example 3: Adaptive Force Control Grasping of Industrial Robotic Arms

[0128] The robotic arm's end effector integrates piezoelectric tactile sensors and a miniature event camera. When the robotic arm grasps an object of unknown material and weight, the piezoelectric tactile sensor detects the rate of pressure change the instant the fingers contact the object. If the rate of change indicates that the object may slip, the L2 gripping channel maps the deviation signal into an increment of gripping force within microseconds. Simultaneously, the visual event camera captures the minute displacement optical flow of the object during the grasping process, and the L2 visual channel maps the optical flow pattern into end-effector trajectory correction parameters. L1 sets the maximum gripping force threshold to prevent the object from being crushed. If the object's material and surface properties are completely unknown, L2 matching fails. L3 infers new parameter combinations near the stability domain boundary of the currently known gripping strategies, finds a gripping strategy suitable for the new object, and stores the new strategy in the L2 temporary buffer, achieving individual evolution of "learning on the first grasp and instinct on the second grasp."

[0129] Example 4: Suppression of Sudden Gust Disturbances in Quadcopter UAVs

[0130] The UAV integrates an inertial measurement unit (IMU model with analog output buffer) and multiple miniature event cameras. When the UAV encounters a sudden crosswind gust while hovering, the IMU detects the sudden change in fuselage angular velocity within microseconds. The L2 trunk channel maps the rate of change of angular velocity to differential thrust compensation commands. The downward-facing event camera detects global optical flow offset, and the L2 vision channel maps the optical flow offset to horizontal position correction commands. L1 limits the maximum tilt angle and maximum motor current. If the gust characteristics do not match the preset stability domain, L3 is activated to explore new anti-disturbance strategies near the parameter space boundary of the known attitude control strategy and stores the new strategies in the L2 temporary buffer. When the UAV encounters a similar gust again, it directly calls upon individual experience without re-deductive analysis.

[0131] Example 5: Circuit principle of L1 intrinsic fuse layer

[0132] like Figure 6As shown, in this embodiment, the L1 instinctive fuse layer consists of multiple sets of parallel signal processing links, each set of links corresponding to a sensor signal that needs to be monitored. Each set of links contains the following four levels of circuitry.

[0133] The first stage is the input buffer and filtering stage. The analog voltage signal output by the sensor first enters a voltage follower, providing high input impedance to avoid loading the sensor signal source. The output of the voltage follower is divided into two paths: one path goes directly to the next stage comparator, and the other path passes through a low-pass filter with a cutoff frequency set above the effective bandwidth of the sensor to filter out high-frequency noise and transient glitches, preventing false triggering. The voltage follower can be a general-purpose operational amplifier (such as TI's OPA series), and the low-pass filter can be a first-order RC passive filter structure.

[0134] The second stage is the threshold comparison stage. It employs a comparator chip with ultrafast propagation delay, such as Analog Devices' ADCMP580 / ADCMP581 series (typical propagation delay 180 picoseconds) or Texas Instruments' TLV3601 (propagation delay approximately 2.5 nanoseconds). The inverting input of the comparator is connected to the filtered sensor signal, and the non-inverting input is connected to a reference voltage. When the sensor signal voltage exceeds the reference voltage, the comparator output immediately flips; the flipping delay is limited only by the physical characteristics of the comparator itself.

[0135] The third stage is the reference voltage generation and threshold storage stage (hardware protected area). A high-precision bandgap reference voltage source (such as Analog Devices' ADR series) generates a stable reference voltage, which is connected to a resistor divider network configured with a one-time programmable memory cell or an electronic fuse. During factory calibration, the voltage is written to the OTP array via a JTAG or SPI interface. After the write is complete, the programming enable fuse is blown, and the reference voltage is permanently locked. The output of the voltage divider network is the safety threshold reference voltage, which cannot be modified through any software interface or OTA.

[0136] The fourth stage is the power drive and output stage. The comparator output is connected to the gate of an N-channel power MOSFET, the drain of the MOSFET is connected to the enable pin of the motor driver, and the source is grounded. During normal operation, the MOSFET is turned on, and the enable pin is pulled low; during a fault, the MOSFET is turned off, the enable pin is pulled high, and the driver is immediately disabled. The delay of the entire signal link is only the comparator propagation delay plus the MOSFET switching delay, with the total delay controlled in the order of nanoseconds to tens of nanoseconds.

[0137] All circuits in L1 are powered by an auxiliary power domain independent of the system's main power supply, such as... Figure 7 As shown, each joint channel operates completely independently and in parallel, ensuring that the system enters working state the instant it is powered on.

[0138] Example 6: Implementation of the L3 Optimal Solution Selection System

[0139] like Figure 10 As shown, after L3 generates 2 to 5 candidate solutions, the optimal solution is determined through the following hardware-based evaluation and selection mechanism. Unlike the prior art's method of deduction followed by filtering, this invention performs security checks simultaneously during the deduction process.

[0140] Before initiating the simulation, L3 loads all safety constraint parameters from the threshold storage area of ​​the L2 protected area into the hardware constraint unit. The path management unit generates candidate control inputs, and each candidate path is executed in parallel in an independent simulation pipeline. Each simulation pipeline contains a slope calculation unit, a state update unit, and a hardware constraint unit.

[0141] After each integration step, the hardware constraint unit synchronously verifies: whether the joint torque is within the fusing threshold, whether the centroid projection is within the supporting polygon, whether the plantar tangential force satisfies the friction cone constraint, and whether the joint power is within the power limit. If any of these fail, the path is marked as infeasible and terminated immediately, and its computational resources are released. The path management unit can generate new candidate paths to replace the eliminated paths.

[0142] Paths that pass the safety filter are considered in the comprehensive score: J = 0.5·S + 0.2·E + 0.3·M. The path with the lowest J value is selected for execution. If all paths fail the safety filter, L3 returns to the L2 protected area to implement a safety shutdown strategy.

[0143] Example 7: Specific Implementation of Continuous Region Mapping of Physical Law Parameters

[0144] In this embodiment, the physical parameter of plantar friction coefficient is used as an example to illustrate how it is stored and matched in the form of a continuous region in L2 hardware.

[0145] The plantar friction coefficient is represented as a continuous safety range in the friction cone constraint subarray of the L2 protected area. The friction cone constraint comparator in this subarray contains a continuous range defined by two reference currents. The lower reference current corresponds to a conservative friction coefficient (e.g., 0.20), generated by a resistor divider from the reference voltage provided by the OTP memory cells. The upper reference current corresponds to an economic friction coefficient (e.g., 0.60), generated by a resistor divider from the reference voltage provided by another set of OTP memory cells. The plantar tangential force signal is converted into a current signal by a transconductance amplifier and then fed into the comparator along with the current signal obtained by multiplying the normal force signal by the current friction coefficient. When the equivalent current of the tangential force falls between the lower and upper reference currents, the comparator outputs a safety signal. When the equivalent current of the tangential force is lower than the lower reference current, it indicates insufficient friction, and the comparator outputs an anti-slip protection signal. When the equivalent current of the tangential force is higher than the upper reference current, it indicates sufficient friction, and the system can appropriately reduce the plantar contact force to save energy.

[0146] Within the safe range, the comparator's output is not a simple binary signal, but a differential signal that continuously varies with the position of the equivalent tangential force current within the range. When the equivalent tangential force current is at the center of the range, the differential output is zero, and the system maintains the current strategy. As the equivalent tangential force current gradually approaches the lower limit, the differential output continuously increases in the negative direction, triggering a gradual anti-slip adjustment. As the equivalent tangential force current gradually approaches the upper limit, the differential output continuously increases in the positive direction, triggering a gradual energy-saving adjustment. This continuous and gradual response mechanism ensures a smooth and natural transition for the robot between different friction conditions, eliminating abrupt changes in action caused by discrete threshold switching.

[0147] For power constraints on joint torque and angular velocity, the upper power limit is also expressed as a continuous range rather than a single threshold. When the real-time power is below 80% of the upper power limit, the system operates normally. When the real-time power enters the range of 80% to 100% of the upper power limit, the derating factor decreases linearly from 1.0 to 0.5, and the motor current is smoothly limited. When the real-time power exceeds 100%, the derating factor jumps to 0, and the motor is immediately shut down. This piecewise continuous response curve balances the timeliness of safety protection with the smoothness of action adjustment.

[0148] Example 8: Generation of the physical law parameter library and configuration of the two-layer structure of the intelligent agent at the time of manufacture

[0149] like Figure 11 As shown in the figure, this embodiment describes in detail how to generate an initial L2 rule base and configure individual coefficients before the agent leaves the factory.

[0150] Step 1: Establish a nominal physical parameter model. Based on the design parameters of this intelligent agent (including joint motor specifications, reducer transmission ratio, link mass and inertia, foot and hand sensor characteristics, etc.), a digital twin model is built in the physical simulation engine, containing complete physical interaction parameter mapping relationships.

[0151] Step Two: Execute standardized test sequences. The simulation system runs standardized test sequences covering the full range of joint motion, contact with different ground materials, and grasping different objects. It systematically traverses the parameter space, records stability indicators under various parameter combinations, and extracts the safe and stable domain boundaries. The specific method is consistent with the aforementioned "Simulation Scanning and Stability Domain Extraction".

[0152] Step 3: Generate a general physical law parameter library for this model based on simulation results, and simultaneously extract the hard safety thresholds of the protected area. Two types of parameters are extracted from the simulation test data. The first type is the hard safety thresholds of the protected area, including the absolute boundaries of physical laws such as the upper limit of torque fusing, the upper limit of power, the conservative friction coefficient, and the boundary of the supporting polygon for each joint. These thresholds are written into the OTP storage unit of the protected area and locked at the factory, and do not participate in OTA updates. The second type is the physical law parameter library of the main matching area, which is the suggested parameter range that can be updated via OTA, including joint speed-friction force mapping, motor current-output torque mapping, contact force-deformation mapping, grip force-object weight mapping, posture-center of gravity position mapping, optical flow dilatation rate-avoidance action mapping, step geometry-gait parameter mapping, etc. These mapping relationships are stored in the conductance lookup table of the in-memory computing subarray in the main matching area in the form of a set of input-output electrical parameter ranges, forming an evolutionary general L2 rule library. The hard thresholds of the protected area and the parameter ranges of the main matching area are from the same source—both from simulation scanning and physical law verification—but are stored separately and updated independently.

[0153] Step 4: Perform factory calibration to generate individual coefficient vectors. On a standardized test bench, perform joint power calibration, kinetic energy derating calibration, gravity term calibration, and friction cone calibration on the agent. Compare the measured values ​​with the simulated nominal values ​​to calculate the individual coefficient vector. For example, knee joint torque coefficient = measured maximum safe torque / simulated nominal maximum safe torque; plantar friction coefficient = measured conservative friction coefficient / simulated nominal conservative friction coefficient. Each coefficient in the individual coefficient vector corresponds to a specific parameter mapping relationship in the general L2 rule base.

[0154] Step 5: Merge and Burn. Merge the general L2 rule base and the individual coefficient vector to generate a personalized L2 rule base (using the base value × coefficient method), and burn it into the agent's L2 storage unit. At the same time, the original data of the general L2 rule base and the individual coefficient vector are also stored in local non-volatile storage for subsequent OTA updates.

[0155] Step Six: Deployment and Continuous Updates. After deployment, the intelligent agent continuously monitors deviations during actual operation, updates individual coefficient vectors, and uploads them to the cloud, forming a closed loop of individual experience and group evolution.

[0156] Example 9: Quantization Coding Scheme for Electrical Parameter Mapping

[0157] This embodiment describes in detail how the parameters in the physical law parameter library are quantified into a set of input-output electrical parameters that can be executed by the L2 in-memory computing array.

[0158] Input-side encoding scheme. The analog voltage V_sense output by the sensor is converted into an input current I_in by a transconductance amplifier, with the conversion relationship being I_in = gm × V_sense. The gain gm of the transconductance amplifier is determined by factory calibration. The calibration method is as follows: apply a known standard voltage V_cal to the input of the transconductance amplifier, measure the output current I_cal, and calculate gm = I_cal / V_cal. The calibrated gm value is stored in L2 non-volatile memory as part of the individual coefficients. The multidimensional vector representation of multiple sensor signals is as follows: the voltage signal of each sensor is converted into an independent current component by an independent transconductance amplifier and injected into different word lines of the in-memory array. In the analog domain, the current of each word line is multiplied by the conductance matrix of the corresponding bit line at the intersection point. The currents at all intersection points naturally converge and sum on the bit line according to Kirchhoff's current law.

[0159] Output-side encoding scheme. The bit-line converged current I_out is converted into a drive voltage V_drive by a current-to-voltage converter, with the conversion relationship V_drive = R_feedback × I_out, where the feedback resistor R_feedback is calibrated according to the actuator input voltage range. For the L2 protected area, the upper limit reference current I_ref_high and lower limit reference current I_ref_low of the current comparator are generated by the bandgap reference voltage source V_ref through a resistor divider network, with the relationship I_ref_high = V_ref / R_high and I_ref_low = V_ref / R_low. R_high and R_low are determined by the resistor network configured by the OTP and are permanently locked after factory calibration.

[0160] Conductivity mapping scheme. The conductivity value G_ij at each intersection point in the in-memory computing array is set according to the desired input-output relationship, which is G_ij = I_out_i / V_in_j. Here, I_out_i is the desired current component of the i-th bit line output, and V_in_j is the voltage input to the j-th word line. For linear mapping, the conductivity value is constant; for nonlinear mapping, the conductivity value is stored in adjacent SRAM cells in the form of a lookup table. When the input signal falls between adjacent lookup table entries, the output signal obtains a continuous value through linear interpolation in the analog domain, ensuring that the mapping relationship is continuously differentiable across the entire parameter space.

[0161] Example 10: Circuit Principle of L2 Electrical Parameter Matching

[0162] like Figure 4 As shown, this embodiment describes how an L2 in-memory computing array achieves electrical parameter matching, in conjunction with a specific circuit structure.

[0163] Word line drive circuit. Each word line input is connected to a voltage buffer, which converts the input current I_in into a word line voltage V_word = I_in × R_word through a precision resistor R_word. The voltage buffer ensures that the word line voltage is not affected by the subsequent array load. Multiple word lines are simultaneously subjected to voltage, forming an input vector.

[0164] In-memory computing array. The array consists of N word lines and M bit lines intersecting, with each intersection point containing an SRAM cell storing a weight value. The output of the SRAM cell controls the conduction level of a transmission gate, and the conductance value G_ij of the transmission gate is proportional to the stored weight value. When the word line voltage V_word_j is applied to the j-th word line, the current flowing out of the j-th intersection point on the i-th bit line is I_ij = V_word_j × G_ij. The sum of the currents at all intersection points on the i-th bit line is I_out_i = Σ(V_word_j × G_ij), which is the product of the input vector and the conductance matrix.

[0165] Current comparators. The output current I_out of each bit line enters a dual-limit current comparator. The upper limit reference current I_ref_high and the lower limit reference current I_ref_low of the comparator constitute a continuous safe current range. When I_out falls within the range, the comparator outputs a high level, indicating a successful match. When I_out exceeds the range, the comparator outputs a low level, indicating a failed match. The comparator uses a differential pair structure, with a response time of approximately 200 nanoseconds, consistent with the aforementioned L2 time budget.

[0166] Example 11: Specific Implementation of the L3 Hardware Differential Equation Solver

[0167] like Figure 5As shown, this embodiment further describes the specific digital circuit implementation of the L3 hardware-based differential equation solver.

[0168] like Figure 9 As shown, the floating-point arithmetic unit contains a set of parallel floating-point multipliers and adders, using the IEEE 754 single-precision floating-point format. Each component of the state vector X_n is assigned an independent multiplier and adder. Taking the calculation of k1 as an example, k1_i = f_i(X_n, U), where f_i is the derivative function of the i-th state component, specifically derived from the translational dynamics equations of the robot's center of mass and the rotational dynamics equations about the center of mass. The multiplier and adder adopts a three-stage pipeline structure: the first stage performs multiplication, the second stage performs addition, and the third stage outputs the result. The floating-point multiplier and adder can be integrated using mature IP cores.

[0169] State Update Unit. The state update unit calculates the weighted sum of the four slopes k1 to k4, multiplies it by the step size h / 6, and then adds it to the current state X_n. Specifically, it uses an adder tree to calculate k1+2k2+2k3+k4, multiplies it by h / 6 using a floating-point multiplier, and finally adds it to X_n using a floating-point adder to obtain X_{n+1}.

[0170] The path management unit is a finite state machine. The core of the path management unit is a finite state machine that controls the generation, termination, and resource reallocation of candidate paths. The state machine includes the following states: Idle state, waiting for the L2 wake-up signal; Initialization state, loading safety constraint parameters from the L2 protected area and generating 2 to 5 sets of initial control inputs U; Deduction state, where each candidate path performs RK4 integration in parallel; Verification state, where the hardware constraint unit performs safety verification synchronously after each integration step; if verification fails, the path is terminated and resources are released; Evaluation state, after all path deductions are completed, a comprehensive score is calculated for the paths that pass the safety filter; Selection state, selecting the path with the smallest J value for output; Completion state, storing the result in the L2 temporary buffer and returning to the idle state. The state machine uses binary encoding, implements the state registers with flip-flops, and implements state transitions with combinational logic.

[0171] Example 12: Weighting Circuit of Cerebellum Hardware Weighted Summation Module

[0172] like Figure 3 As shown in the figure, this embodiment describes in detail the analog domain weighted fusion circuit of the cerebellum hardware weighted summation module.

[0173] A weighted summation circuit using transconductance amplifiers is used. The motion intent voltage V_vision output from the visual channel is converted into a current I_vision by the first transconductance amplifier, with the conversion relationship I_vision = gm_vision × V_vision. The deviation correction voltage V_tactile output from the tactile channel is converted into a current I_tactile by the second transconductance amplifier, with the conversion relationship I_tactile = gm_tactile × V_tactile. The two current lines are directly connected at the fusion node and naturally summed according to Kirchhoff's current law: I_fusion = I_vision + I_tactile. The total fused current I_fusion is converted into the final drive voltage V_drive = R_feedback × I_fusion by a current-to-voltage converter.

[0174] Dynamic adjustment of weights. The weights of each channel are adjusted via the gain (gm) of the transconductance amplifier. The gain of the transconductance amplifier is dynamically set by the upper-level brain via a digitally controlled potentiometer or digital-to-analog converter (DAC). When the robot is navigating unfamiliar terrain, the upper-level brain can increase the gain (gm_tactile) of the tactile channel, increasing the weight of tactile correction and prioritizing safety. When the robot is moving at high speed in a familiar environment, the upper-level brain can increase the gain (gm_vision) of the visual channel, increasing the weight of visual prediction and prioritizing efficiency.

[0175] Example 13: Specific Algorithm and Criteria for Stability Region Extraction

[0176] This embodiment details the specific algorithms and criteria for extracting multidimensional stable domain boundaries in a simulated physical world.

[0177] Stability Criteria. For legged robots, the following quantitative indicators are used to determine the stability of each simulation test. The posture stability index is S = max(d_c / R_support), where d_c is the maximum distance the centroid projection deviates from the center of the supporting polygon, and R_support is the radius of the circumcircle of the supporting surface. When S ≤ 1.0, the robot is considered stable, as the centroid projection always falls within the supporting surface. When S > 1.0, the robot is considered unstable, and it has tipped over. The energy consumption index E is used to assess energy consumption levels and is not a hard safety criterion. The safety criterion uses a binary judgment. The robot is considered safe if all of the following conditions are met: no tipping occurred (S ≤ 1.0 always holds); no slippage occurred (the tangential force F_t on the foot always satisfies |F_t| ≤ μ × F_n); no joint torque exceeded the limit (the torque of each joint is always within the fuse threshold set in the L2 protection zone); and no power exceeded the limit (the real-time power of each joint is always within the upper power limit).

[0178] Boundary extraction algorithm. A binary search method is used to search for safe and hazardous boundaries along each dimension of the parameter space. The specific steps are as follows: Select the target parameter dimension to be extracted, such as the plantar friction coefficient, and fix other parameters at their nominal values. Set an initial search interval [P_min, P_max] for the target parameter, where P_min is the known safe value and P_max is the known hazardous value. Take the midpoint P_mid of the interval for simulation testing, run a standardized test sequence, and record stability indicators. If the robot is stable at P_mid, reduce the search interval to [P_mid, P_max]; if it is unstable, reduce the search interval to [P_min, P_mid]. Repeat the above steps until the interval width is less than the preset accuracy ε; the midpoint at this point is the boundary value of the stable domain in that dimension. Repeat the above process for all parameter dimensions to obtain the set of boundary points of the multidimensional stable domain. Store the boundary point set as the upper and lower limit reference currents of the L2 protection zone comparator.

[0179] Example 14: Multiple ways to define the input-output range of a physical law parameter library

[0180] This embodiment describes several specific ways of defining the 'input range - output range' in the L2 physical law parameter library. These methods are physically unified—they all express the laws of the physical world as parameter ranges that can be executed by hardware—but different forms of expression are used due to differences in sensor type, signal form, and law complexity. Regardless of the method used, they are ultimately compiled into a set of electrical parameters that can be executed by the L2 in-memory computing array.

[0181] The first type: Direct mapping of analog electrical signals. This is suitable for sensors that output analog signals, such as tactile sensors, torque sensors, and angular velocity sensors. The input range is defined as the range of analog current or voltage after the sensor signal is converted by the front-end circuit, and the output range is defined as the range of analog current or voltage driving the actuator. For example, in tactile-grip force compensation mapping, the input range is the current range of the piezoelectric tactile sensor after charge amplification and transconductance amplification, corresponding to the physical range of the pressure change rate; the output range is the range of the grip force compensation motor drive current, corresponding to the physical range of the grip force increment. This type of mapping directly corresponds to the conductance value stored in the in-memory array. The input current generates a voltage on the word line, which is multiplied by the conductance matrix to obtain the output current; all calculations are completed in the analog domain.

[0182] The second type: Digital event stream parameter range. This is suitable for sensors that output digital event streams, such as event cameras. The input range is defined as the interval of feature parameters extracted in real time from the event stream, such as the range of optical flow dilatation rate or the range of step edge depth values ​​in optical flow fracture modes; the output range is still defined as the analog drive current interval. The input parameters for this type of mapping are extracted in real time from the digital event stream, converted into internal parallel signals via the digital interface of the L2 vision optical flow channel, and then entered into the in-memory computing subarray for matching.

[0183] The third type: Directly defined range of physical quantities. Applicable to safety constraints in L2 protected areas. The input range is directly defined by physical quantities, such as the ratio range of plantar tangential force to normal force, the range of distances between the centroid projection and the center of the support surface, etc.; the output is a binary state signal, such as protection triggered or normal. The range of physical quantities mapped in this type is compiled at the factory into a reference current or reference voltage range stored in the OTP, and the analog comparator makes continuous range judgments.

[0184] The fourth type: Jointly defined range of multi-dimensional parameters. This is suitable for complex laws involving the coordination of multiple physical quantities. The input range is the joint region of the multi-dimensional parameter space. For example, the law of "stable walking on wet and slippery ground" is jointly defined by the range of foot friction coefficient, stride length, center of gravity forward shift, and foot pressure center offset. All parameters must fall within their respective ranges for the system to determine that the current state is within the stable domain. The output range is the allowable range of joint torques, angles, stiffness, etc. In hardware, the multi-dimensional stable domain is simultaneously verified by multiple parallel comparator arrays. Each comparator is responsible for judging the range of one dimension, and the outputs of all comparators are logically ANDed.

[0185] The fifth type: Time series parameter range. Suitable for detecting trends rather than static values. The input range is the interval of the rate of change within a time window. For example, the range of pressure change rate over a certain time period is used to detect slippage trends, and the range of angular velocity change rate over a certain time period is used to predict attitude instability. The output is a response curve that changes continuously over time.

[0186] Regardless of the definition method used, they are ultimately compiled into a set of electrical parameters that can be executed by the L2 in-memory computing array, including the input current range, output current range, conductance matrix or lookup table, and the reference current range of the protection comparator.

[0187] Example 15: A Personalized L2 Rule Base Generation Method Based on a Digital Model of Physical Laws

[0188] Step 1: Construct a digital model of physical laws (offline training phase).

[0189] By utilizing large AI models (language models or multimodal models) and training them on massive amounts of physics knowledge data, a general "digital model of physical laws" can be established. The training data includes at least: mathematical descriptions of classical physical laws (Newtonian mechanics, friction laws, elastic collisions, fluid mechanics, etc.); material physical property data (friction coefficient, elastic modulus, density, thermal conductivity, etc.); linguistic descriptions of human physical interaction experience; and known physical experimental data (force-displacement curves, velocity-time curves, stress-strain curves, etc.).

[0190] In this embodiment, the physical law digital model can adopt a Physical Information Neural Network (PINN) architecture, embedding physical laws as constraints into the neural network loss function, so that the physical parameters output by the model naturally satisfy physical consistency. The PINN architecture includes at least: a neural network backbone for fitting physical mapping relationships; a physical information layer that encodes physical laws (such as Newton-Euler dynamics equations) as loss function constraints; and a decision layer for processing multi-objective optimization outputs.

[0191] Step 2: Construct a target description library for intelligent agents.

[0192] A target description library is constructed to describe the various actions performed by the target intelligent agent, serving as the "target constraints" for the calculation rule base of the digital model of physical laws. The target description library includes at least:

[0193] (1) Description of moving action targets: walking (step frequency 1.0-1.5Hz, stride length 0.3-0.5m, leg lift height 0.05-0.10m); running (step frequency 2.5-3.5Hz, stride length 0.6-1.0m, leg lift height 0.10-0.20m, air time ≤0.1s); climbing (step frequency 0.8-1.2Hz, stride length 0.2-0.4m, trunk forward lean angle 10°-20°); descending (step frequency 0.8-1.2Hz, stride length 0.2-0.4m, trunk backward lean angle 5°-15°).

[0194] (2) Description of operational targets: grasping (end-point arrival speed ≤ 0.5 m / s, gripping force adjustment range 0.5-5.0 N, contact force control accuracy ≤ 0.1 N); placing (end-point descent speed ≤ 0.3 m / s, release force ≤ 0.1 N, positioning accuracy ≤ 1 mm); lifting (lifting speed 0.1-0.3 m / s, lifting force range 10-50 N, posture stability deviation ≤ 1°).

[0195] (3) Description of postural action targets: standing (center of gravity height design value ±0.01m, trunk verticality ≤0.5°, center of gravity projection position support surface center ±0.02m); squatting (knee joint angle 60°-120°, trunk forward tilt angle ≤15°); sitting (even distribution of pressure on buttocks, trunk verticality 0°±2°).

[0196] (4) Description of response action targets: obstacle avoidance (deflection angle 10°-45°, acceleration 0.5-2.0m / s²); emergency stop (deceleration ≤3.0m / s², after emergency stop the center of gravity projection offset ≤ support surface boundary -0.02m); attitude recovery (recovery time ≤0.5s, after recovery the deviation ≤0.5°).

[0197] Each target parameter in the target description library includes: action type identifier, target parameter name and unit, target value range (minimum, maximum, typical, allowable deviation), target priority weight (safety priority / accuracy priority / speed priority / energy efficiency priority), and coupling constraints between target parameters.

[0198] Step 3: Input the standard parameter model of the intelligent agent.

[0199] The physical design parameters of the target intelligent agent are input into the digital model of physical laws. The standard parameter model of the intelligent agent includes at least: mechanical structure parameters (length, mass, center of mass position, and moment of inertia of each joint link); actuator parameters (torque-speed characteristic curves, maximum torque, maximum speed, and response time constant of each joint motor); transmission system parameters (reducer transmission ratio, transmission efficiency, and backlash); contact surface parameters (foot / hand contact area, material hardness, and friction coefficient range); and sensor parameters (visual system resolution, field of view, and measurement range; tactile sensor sensitivity and measurement range; inertial measurement unit accuracy and range).

[0200] Step 4: Input the sensing parameter range for each L2 channel.

[0201] The perception parameter ranges of each channel of the agent's L2 dynamic rule base are input into the digital model of physical laws. The perception parameter ranges include at least: visual channel (set of recognizable object categories, distance measurement range, velocity measurement range, field of view range); tactile channel (range of resolvable forces, range of pressure change rate, set of material categories, slip detection sensitivity); inertial channel (range of perceptible attitude angles, range of angular velocity, range of acceleration); force channel (range of measurable joint torques, range of end-effector contact forces).

[0202] Step 5: Multi-objective optimization and solution selection.

[0203] The physical law digital model, based on the joint input of the agent's standard parameters, the target description library, and the range of perception parameters, selects the optimal solution from all candidate execution parameter schemes that satisfy the basic physical constraints through a multi-objective optimization function. The optimization function includes at least the following optimization dimensions:

[0204] (1) Energy consumption optimization. Define the total energy consumption function of the agent performing this action, expressed as the integral of the power consumption of each joint motor: E_total=∫(ΣU_j·I_j)dt+∫(Σf_friction(ω_j,τ_j))dt+P_cooling×t. The optimization objective is to minimize E_total.

[0205] (2) Stability Optimization. The comprehensive stability index of the agent is defined as: Stability = α·d_ZMP + β·d_COM + γ·min(margin_j) - δ·Δθ, where d_ZMP is the distance between the zero-moment point and the center of the supporting polygon, d_COM is the distance between the centroid projection and the boundary of the supporting polygon, margin_j is the torque margin of each joint, and Δθ is the attitude angle deviation. The optimization objective is to maximize Stability.

[0206] (3) Time efficiency optimization. Define the task execution time t_total, the optimization objective is to minimize t_total, and the constraints are Stability≥Stability_min and E_total≤E_max.

[0207] (4) Smoothness optimization. Define the motion trajectory smoothness as Smoothness = ∫(Σ(d 3 θ_j / dt 3 ) 2 The optimization objective is to minimize Smoothness.

[0208] The comprehensive objective function is: minimizeF = w1·E_total + w2·(1 / Stability) + w3·t_total + w4·Smoothness, where each weight coefficient is set in real time by the brain A according to the task type, working environment and the current state of the agent or preset by the system.

[0209] Hard constraints (candidate solutions are eliminated if any one of them is not met): Joint angles are within the mechanical limit range θ_min≤θ_j≤θ_max; Joint torques are within the rated range of the motor |τ_j|≤τ_max_j; Joint speeds are within the rated range of the motor |ω_j|≤ω_max_j; Center of gravity projection is within the supporting polygon P_COM∈polygon; Foot contact force is within the friction cone |F_t|≤μ·F_n.

[0210] The multi-objective optimization can employ one or a combination of the following algorithms: gradient-based optimization methods (such as sequential quadratic programming (SQP); metaheuristic algorithms (such as particle swarm optimization, NSGA-II); and methods based on linear quadratic regulators.

[0211] Step 6: Generate a personalized L2 rule base.

[0212] After multi-objective optimization and screening, the physical law digital model outputs the motor drive values ​​corresponding to each set of sensing input parameters after multi-objective optimization, in the format: [sensing parameter index] → [motor 1 torque value, motor 1 position value, motor 1 speed value; motor 2 torque value, motor 2 position value, motor 2 speed value;...].

[0213] The above outputs are combined with individual coefficient vectors (joint friction coefficients, sensor sensitivity coefficients, etc.) to generate a unique personalized L2 rule base for the agent, which is then burned into its local storage unit. The personalized L2 rule base includes at least:

[0214] (1) Main matching area: contains a mapping table of input parameter range and output value, in the format of [input parameter range] → [output execution parameter].

[0215] (2) Protection zone: includes safety boundary parameters and protection triggering conditions, in the format of [safety parameters] → [protection action parameters], used for L2 hard boundary check and L1 instinctive circuit breaker threshold reference.

[0216] Step 7: Validation and evolution of the rule base.

[0217] After the agent is put into operation in the real environment, the rule base is verified and evolved in the following ways: (1) The rule base is fine-tuned online through individual coefficient vectors to compensate for the deviation between the output of the digital model of physical laws and the actual physical execution; (2) Through the group consensus mechanism, the fine-tuning data of multiple agents of the same model are collected to the cloud and new general rules are extracted; (3) The new rules verified by the group consensus are distributed to all agents of the same model through OTA update.

[0218] In a preferred embodiment, the physical law digital model pre-computes the entire workspace of the agent during the offline stage to generate a global optimization mapping table. During runtime, the pre-optimized parameters are directly retrieved from the table without the need for online optimization calculations.

[0219] In another preferred embodiment, the weight coefficients of different optimization dimensions are set in real time by Brain A according to the task type: w1=0.15, w2=0.25, w3=0.50, w4=0.10 for emergency tasks; w1=0.50, w2=0.30, w3=0.05, w4=0.15 for long-term inspections; w1=0.15, w2=0.40, w3=0.05, w4=0.40 for precision operations; and w1=0.05, w2=0.55, w3=0.30, w4=0.10 for emergency obstacle avoidance.

[0220] In another preferred embodiment, when generating the rule base, the physical law digital model also outputs the uncertainty range or confidence interval of each parameter to guide the initial step size setting for subsequent individual coefficient fine-tuning, ensuring that the online fine-tuning process will not exceed the safety boundary due to an excessively large initial step size.

[0221] In another preferred embodiment, the physical law digital model also accepts input of the agent's working environment parameters (working temperature range, humidity range, altitude, etc.) to generate a multi-version rule base adapted to different environmental conditions. The agent automatically switches to the corresponding version when entering different environments.

[0222] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A hierarchical event-driven perception-response architecture for intelligent agents in the physical world, characterized in that, include: The first response layer L1 is a pure hardware logic circuit consisting of an analog comparator and a power drive circuit directly connected. Its input is connected to the output of the force / torque sensor and the inertial measurement unit of the intelligent agent, and its output is connected to the enable terminal or the current limiting terminal of the corresponding motor driver. The second response layer L2 is a hardware mapping circuit composed of a memory computing array. Its input is connected to an event-driven sensor and an analog signal output sensor. The memory computing array stores a stable domain boundary defined by multi-dimensional parameters, which is expressed in the form of a set of input-output electrical parameters. The third response layer L3 consists of a hardware-based differential equation solver, which is connected to L2 via an independent hardware interrupt signal line. Under normal circumstances, it is in a sleep state with clock gating and power gating. The original sensor signal enters L1 and L2 simultaneously. L1 performs protection judgment independently of L2 and L3. When L2 is successfully matched, the wake-up signal of L3 is not triggered.

2. The hierarchical event-driven perception-response architecture for physical world intelligent agents according to claim 1, characterized in that, The L2 internally includes multiple parallel matching channels pre-splitting according to physical interaction type, a cerebellum hardware weighted summation module, and a hardware protection zone; the parallel matching channels include visual optical flow channels, plantar contact channels, grasping channels, and trunk posture channels; the hardware protection zone has independent analog signal input pins and independent physical signal output pins. The independent analog signal input pins are directly connected to the analog output terminals of each joint sensor, and the independent physical signal output pins are directly connected to the enable and current limiting terminals of each joint motor driver. The signal transmission path does not pass through analog-to-digital conversion and data bus.

3. The hierarchical event-driven perception-response architecture for physical world intelligent agents according to claim 2, characterized in that, The L2 also includes a temporary rule cache area for storing parameter combination points successfully derived by the L3. The L3 includes a path management unit, 2 to 5 parallel derivation pipelines and a hardware constraint unit. After each integration step, the hardware constraint unit loads safety constraint parameters from the threshold storage area of ​​the L2 protected area and performs synchronous verification. When any path is determined to violate the safety constraint in any step, the derivation pipeline of that path is immediately terminated.

4. The hierarchical event-driven perception-response architecture for physical world intelligent agents according to claim 1, characterized in that, The input terminal of the analog comparator of L1 is connected to the analog voltage output of the sensor, and the reference voltage terminal is connected to the output terminal of the resistor voltage divider network configured by a one-time programmable memory cell or an electronic fuse. The comparator output directly drives the gate of the power MOSFET. The threshold parameter of L1 is stored in the one-time programmable memory cell, and after writing is completed, it is permanently locked by the fuse through the fuse programming enable.

5. A chip, characterized in that, The hierarchical event-driven perception-response architecture for physical world intelligent agents as described in any one of claims 1 to 4 is implemented, wherein the chip is any physical form of a single chip, a multi-chip module, a system-in-package, or a board-level system.

6. An intelligent agent system, characterized in that, include: The chip according to claim 5; The sensor module includes at least an event camera and a piezoelectric tactile sensor; An analog signal output sensor module, comprising at least a torque sensor and an inertial measurement unit; The actuator module includes at least a joint motor and its driver, with the enable terminal and current limiting terminal of the driver connected to the L1 output terminal and the L2 protection zone independent physical signal output pin of the chip, respectively. The communication module connects to the chip and is used for data transmission with the cloud server.

7. A method for physical interaction using a hierarchical event-driven perception-response architecture for physical world intelligent agents as described in any one of claims 1 to 4, characterized in that, include: The raw sensor signal enters both L1 and L2 simultaneously; L1 independently determines whether the sensor signal exceeds the safety threshold within nanoseconds; if it does, it directly triggers the protection action. L2 performs hardware matching of the current state parameters with the preset stable domain boundary within microseconds. If the current state parameters fall within the stable domain, it directly outputs the drive parameters and keeps L3 in sleep mode. If the current state parameters exceed all known stability regions or a conflict occurs, L2 will wake up L3 via a separate hardware interrupt signal line; L3 performs sparse sampling of the current event stream in milliseconds, loads safety constraint parameters from the L2 protected area, performs safety checks synchronously at each step of the simulation and immediately terminates unsafe paths, and selects the optimal path and outputs driving parameters through safety filtering and comprehensive scoring. L3 stores the successful parameter combinations from this simulation into the temporary rule cache of L2.

8. A method for generating a physical law parameter database for an AI agent in the physical world at the time of its manufacture, characterized in that, include: Establish the nominal physical parameter model of the corresponding intelligent agent; Execute standardized test sequences in the physics simulation engine and generate a general physical law parameter library based on the simulation results; Perform factory calibration tests on each intelligent agent, measure the deviation between the actual physical response and the simulation expected value, and quantify the deviation into an individual coefficient vector; The general physical law parameter library is merged with the individual coefficient vector to generate a personalized physical law parameter library unique to the corresponding intelligent agent, and then burned into the local storage unit of the corresponding intelligent agent.

9. A method for generating a personalized L2 rule base for an AI agent in the physical world based on a digital model of physical laws, characterized in that, include: A digital model of physical laws is constructed, which is trained by the AI ​​large model on a training set containing mathematical descriptions of classical physical laws, material physical property data and known physical experimental data. Construct a target description library for the target intelligent agent, wherein each target parameter in the target description library includes an action type identifier, a range of target values, and a weight of target priority; The standard parameter model of the target intelligent agent and the perception parameter range of each L2 channel are input into the physical law digital model; The physical law digital model, based on the joint input of the standard parameter model, the target description library, and the perception parameter range, selects the optimal solution from the candidate execution parameter schemes that satisfy the hard constraints through a multi-objective optimization function, and generates the corresponding L2 rule base parameters. The hard constraints include at least joint angle limit constraints, joint torque limit constraints, and center of gravity projection constraints. The L2 rule base parameters are combined with the individual coefficient vector of the target agent to generate a personalized L2 rule base unique to the agent, which is then burned into the agent's local storage unit.

10. A physical world AI agent personalized L2 rule base generation system implementing the method of claim 9, characterized in that, include: A physical law digital model generation module is used to construct the physical law digital model; A target description library storage module is used to store the target description library; The agent parameter input module is used to receive the standard parameter model and the range of sensing parameters for each L2 channel; The multi-objective optimization module is used to select the optimal solution through a multi-objective optimization function and generate the corresponding L2 rule base parameters. A personalized rule base generation module is used to combine the L2 rule base parameters with the individual coefficient vector to generate a personalized L2 rule base. The rule base burning module is used to burn the personalized L2 rule base into the local storage unit of the intelligent agent.

Images