A world model based embodied intelligence control method

Abstract

The application relates to the technical field of neural network models, and discloses a somatic intelligent control method based on a world model, which collects multi-modal sensing data and equipment states, generates a unified state package containing equipment fingerprints, instantiates a single-equipment digital twin based on an equipment cluster template, decomposes a total task into a verifiable sub-target sequence with preconditions and postconditions and a rollback node by a high-level task world model, generates corresponding skill parameters by a middle-level skill world model, makes a short-time prediction and generates an action block by a low-level dynamics world model, verifies logical preconditions and geometric accessibility conditions, safety constraint conditions and uncertainty conditions in sequence, and executes an action after all conditions are satisfied; execution feedback is collected, a sub-task postcondition is verified, the next sub-target is promoted if the verification result is positive, and the digital twin parameters are corrected if the verification result is negative. The application solves the problems of limited end-side computing power and error accumulation caused by long tasks, and achieves real-time controllability, long-time stability and self-adaptive correction.

Description

Technical Field

[0001] This invention relates to the field of neural network model technology, and in particular to an embodied intelligent control method based on a world model. Background Technology

[0002] World models are essentially "internal simulators" for embodied intelligent agents, used to model how environmental states evolve under the influence of actions, thereby supporting forward prediction, counterfactual inference, planning, and decision-making. A recent review further summarizes embodied world models as a three-axis system: functionally, they can be divided into task-coupled and general-purpose models; temporally, they can be divided into sequential simulation and differential prediction; and spatially, they can be divided into global latent, token sequences, spatial grids, and decomposed rendering representations. The review also points out that its open challenges include: a lack of unified data, a lack of evaluation systems that emphasize physical consistency rather than pixel fidelity, the contradiction between computational efficiency and real-time control, and the accumulation of long-term time-series errors.

[0003] For example, Chinese patent CN119623536B discloses a task planning method based on an embodied intelligent world model using visual reasoning. It provides the following technical solution: combining the language processing capabilities of a large language model with advanced visual perception technology, enabling the intelligent agent to pre-reasonize environmental states and simulate possible actions using a world model with limited environmental visual information. This allows for effective decision-making and rapid adaptation to constantly changing environmental states in complex environments, significantly improving the agent's task execution efficiency and success rate in complex scenarios. However, this task planning method based on an embodied intelligent world model relies on a large language model and a diffusion model for high-level visual reasoning and state prediction. It lacks mechanisms for underlying execution loops, device-specific difference compensation, edge computing power constraint adaptation, and mass production consistency, making it difficult to solve the physical interaction and real-time control problems in actual deployment. Summary of the Invention

[0004] This invention addresses the problems of limited computing power on the edge, error accumulation due to long tasks, and poor mass production consistency in existing technologies. It proposes an embodied intelligent control method based on a world model, achieving the goals of mass production capability, real-time controllability, long-term stability, and adaptive correction.

[0005] Furthermore, this invention aims to construct an embodied intelligence platform that integrates cloud-edge-device collaboration and merges device fingerprints with a hierarchical world model, in order to solve the challenges of real-time, stable, and adaptive execution in mass production scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An embodied intelligent control method based on a world model includes: Collect multimodal sensor data and device status, generate a unified status package containing device fingerprints, and instantiate a single device digital twin based on a device cluster template; The overall task is decomposed into a verifiable sequence of sub-objectives with preconditions, postconditions, and backtracking nodes by the high-level task world model; the mid-level skill world model generates corresponding skill parameters; and the low-level dynamics world model makes short-term predictions and generates action blocks. After verifying the logical preconditions, geometric reachability conditions, safety constraints, and uncertainty conditions in sequence, the action is issued and executed once all conditions are met. Collect execution feedback, verify the post-conditions of the subtask, and if they pass, advance to the next sub-objective; otherwise, roll back to the most recently verified node and correct the digital twin parameters.

[0007] Preferably, in the process of generating a unified state package containing device fingerprints, a device cluster template for similar devices is established, standardized calibration data and operating data of single devices are collected, individual feature parameters of joint transmission visual time delay temperature drift and aging are extracted, a residual fingerprint relative to the cluster template is formed, and a three-layer world model and a digital twin module are injected to complete the correction of state prediction and action generation.

[0008] Preferably, in the instantiation of a single device digital twin based on a device cluster template, a dual mechanism of offline mass production template closed loop and online device synchronization closed loop is adopted. The offline closed loop updates the template by combining device cluster partitioning and cluster-level simulation training with batch data backhaul. The online closed loop continuously corrects the residual parameters of a single device by observing the residual in real time.

[0009] Preferably, the decomposition of the total task by the high-level task world model includes adapting different implementation methods according to task type. For open semantic tasks, a language model combined with a symbolic constraint solver is used. For industrial process tasks, a graph neural network combined with a task state machine is used. For tasks with strong rules and requiring causal interpretability, an embodied task knowledge graph and a causal reasoning module are used.

[0010] Preferably, the rollback to the most recently verified node includes: firstly rolling back to the starting node of the current subtask; if the preceding state is distorted, rolling back to the direct predecessor node; if the local task chain fails, rolling back to the root node of the high-level replanning; after rolling back, first restoring the state snapshot and re-sensing before continuing execution.

[0011] Preferably, in the low-level dynamic world model, short-time prediction and action block generation are performed using a dual-channel encoding method. The low-frequency smooth principal components of the action sequence are extracted through frequency domain transformation or smooth curve fitting. The event components are obtained by sparse detection of abrupt changes in the reconstructed residuals. The end-side only decodes the preceding actions required in the current time domain.

[0012] Preferably, the verification of the uncertainty condition also includes a dynamic precision scheduling step, where multiple precision versions of the model are deployed on the edge, and the precision is switched according to the task risk level, latency budget and uncertainty level. The confidence level is output through the confidence mapping head jointly trained with the quantization model. If the confidence level is lower than the threshold, the precision is increased or the data is offloaded to the edge cloud.

[0013] Preferably, in the process of falling back to the most recently verified node and correcting the digital twin parameters, when the uncertainty does not reach the shutdown threshold, a micro-probing action mechanism is triggered to automatically generate a trial action, update the pose, contact, friction or clamping state using its observation results, and re-execute the verification after the state correction is completed.

[0014] Preferably, the actions are issued and executed after all the conditions are met. If the logical preconditions are not met, the system returns to the high-level task world model for task rearrangement or state restoration. If the geometric reachability conditions are not met, the system triggers re-sensing or path replanning. If the dynamic and safety constraints are not met, the system reduces the speed, shortens the step size, or switches to a conservative control mode. If the uncertainty conditions are not met, the system increases the model accuracy or requests replanning on the edge or cloud side.

[0015] Preferably, the safety constraints are determined based on the total risk value obtained by weighted fusion of the collision contact pose execution task and the model risk components.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0017] 1. This invention uses a parameterized system that combines equipment cluster templates with individual equipment residual fingerprints to separate and manage the common priors of similar equipment from the individual deviations of a single equipment. With the help of an online synchronous correction mechanism, the digital twin model can continuously track the wear, temperature drift, and assembly deviations of real equipment. This transforms the traditional one-time simulation migration into a self-correcting closed loop that runs through the entire life cycle of the equipment, ensuring the stability of the strategy and consistency of the execution of batch-delivered robots under complex working conditions.

[0018] 2. This invention constructs a heterogeneous computing architecture with three-level collaboration between cloud, edge, and device. Long-term simulation and retraining are placed in the cloud, task planning and event memory are placed on the edge, and only short-term dynamic prediction and security closure are left on the device. Supplemented by dynamic precision switching, dual-channel encoding of action blocks, and quantization reliability calibration mechanisms, the lightweight device model can meet the high real-time control closed-loop requirements on a low-power controller, while seamlessly requesting upper-layer computing power to take over in highly uncertain tasks.

[0019] 3. This invention employs a hierarchical physical agent architecture that coordinates a three-layer world model: a high-level task, a mid-level skill, and a low-level dynamics. The high-level layer is responsible for constructing sub-target trees with preconditions and postconditions; the mid-level layer is responsible for skill parameterization and termination verification; and the low-level layer is responsible for millisecond-level contact prediction and execution permission checks. Combined with event memory compression and a verified node rollback mechanism, the system does not need to start from scratch in the event of a partial failure; it can simply roll back to the most recently trusted state to continue execution, significantly improving the success rate and robustness of multi-stage complex operations. Attached Figure Description

[0020] Figure 1 This is a flowchart of the long task execution process of the present invention.

[0021] Figure 2 This is a flowchart of the cloud-edge-device collaboration and dynamic precision scheduling process of the present invention. Detailed Implementation

[0022] See Figure 1 and Figure 2 As shown, an embodied intelligent control method based on a world model includes: Collect multimodal sensor data and device status, generate a unified status package containing device fingerprints, and instantiate a single device digital twin based on a device cluster template; The overall task is decomposed into a verifiable sequence of sub-objectives with preconditions, postconditions, and backtracking nodes by the high-level task world model; the mid-level skill world model generates corresponding skill parameters; and the low-level dynamics world model makes short-term predictions and generates action blocks. After verifying the logical preconditions, geometric reachability conditions, safety constraints, and uncertainty conditions in sequence, the action is issued and executed once all conditions are met. Collect execution feedback, verify the post-conditions of the subtask, and if they pass, advance to the next sub-objective; otherwise, roll back to the most recently verified node and correct the digital twin parameters.

[0023] The platform of this invention is generally divided into three layers. The first layer is the cloud-based training and world model center; the second layer is the edge-based collaborative control center; and the third layer is the end-side embodied execution device. The three layers are not in a fixed master-slave relationship, but rather dynamically divided according to task complexity, uncertainty, real-time budget, and security level.

[0024] The Cloud-based Training and World Model Center is primarily responsible for: aggregating real robot data, simulation data, open-source embodied data, and synthetic world data; training or fine-tuning the basic world model; generating extreme conditions, long-tail conditions, and counterfactual training data; forming a device cluster template library; and distilling, trimming, and quantizing edge and end-side models.

[0025] The device cluster template library is formed, which includes three parts: the template library construction process, the update mechanism, and the parameter types within the library. 1. Template Library Construction Process First, the robot equipment of the same or similar models is divided into clusters. The cluster division is based on at least the mechanical structure, number and layout of joints, actuator type, reducer type, end tool type, sensor configuration, camera mounting method, controller version, cable and transmission layout, etc.

[0026] After the cluster division is completed, multi-source data corresponding to each device cluster is collected. The multi-source data includes at least factory calibration data, kinematic calibration data, dynamic identification data, visual calibration data, time delay measurement data, grasping and contact test data, historical fault data, and mass production quality inspection data.

[0027] Subsequently, statistical modeling, cluster fitting, or distribution estimation are performed on the parameters of multiple sample devices within the same device cluster to obtain the common template parameters of the device cluster, and the device cluster template entries are written into the template library.

[0028] For each device cluster, a template record structure of "basic template + parameter upper and lower bounds + typical operating condition label + applicable task label" is established for subsequent calls.

[0029] 2. Template library update mechanism The template library is not generated statically all at once, but is continuously maintained through a combination of offline and online updates.

[0030] Offline updates refer to periodically summarizing calibration data, operation logs, repair data, fault data, and typical task execution data of newly added mass-produced equipment, re-clustering or re-estimating parameter distributions of equipment clusters, and updating the corresponding templates. Online updates refer to the process of correcting the statistical boundaries, prior distributions, or representative values ​​of equipment cluster templates during equipment deployment and operation, based on information such as long-term observation residuals, performance drift, temperature drift, wear changes, capture success rate changes, and control delay changes of individual devices.

[0031] When a device deviates from the existing template threshold for an extended period, a new cluster generation mechanism can be triggered, which will either assign the device to a new device cluster or create a new template branch, thereby preventing the original template from becoming distorted.

[0032] Device clustering refers to classifying robot devices in mass production scenarios into several categories based on their common characteristics that significantly impact control, perception, and sim2real, with each category corresponding to a device cluster.

[0033] The classification dimensions include at least the following: mechanical structure characteristics: number of joints, linkage size, range of motion, and installation method; actuator characteristics: motor model, reducer batch, driver version, and gripper or end effector type; sensor characteristics: camera model, force / tactile sensor model, and IMU configuration; vision installation scheme: camera installation position, viewing angle, and calibration method; tool head scheme: different tool end configurations such as suction cups, grippers, welding guns, and tightening heads; and control link characteristics: controller version, communication bus type, and latency level.

[0034] The above division enables robots within the same device cluster to share similarities in kinematics, dynamics, perception characteristics, and execution links, thereby sharing a set of cluster-level prior parameters and simulation models.

[0035] Cluster-level simulation training refers to building a corresponding cluster-level simulation model for a cluster of devices, rather than for a single device, and then performing batch training on that model.

[0036] The specific process may include: establishing a cluster-level template model: based on the common parameters of the device cluster, establishing corresponding kinematic, dynamic, visual, and time-delay simulation models; setting the intra-cluster parameter distribution: setting the intra-cluster fluctuation range for parameters such as joint zero bias, friction coefficient, hysteresis, camera extrinsic parameters, gripper elasticity, and execution delay; performing distributed sampling simulation: sampling from the parameter distribution of the device cluster during training to generate multiple sets of simulation instances of "different individuals in the same cluster"; training the world model and control strategy in the simulation: enabling the model to learn to stably complete the task under changes in parameters within the device cluster; outputting a cluster-level shared model: training to obtain a cluster-level model that is universally applicable to the device cluster, and then superimposing single-device residual corrections during deployment.

[0037] The template library adopts a version management mechanism, which records the template version number, update time, applicable device range, parameter changes and performance evaluation results for each update, in order to support rollback, traceability and batch deployment.

[0038] 3. Parameter types within the template library The parameters in the template library must include at least one or more of the following: Kinematic parameters: joint zero position, link length, installation deviation, DH parameter, end-effector coordinate system deviation; Dynamic parameters: mass, moment of inertia, coefficient of friction, damping coefficient, stiffness coefficient, backlash parameter, compliance parameter; Execution chain parameters: motor constant, drive gain, reduction ratio, gripper closure characteristics, tool end elasticity parameters; Visual parameters: camera intrinsic parameters, extrinsic parameters, distortion parameters, camera mounting pose deviation, multi-camera synchronization deviation; Contact and sensing parameters: force / torque sensor zero drift, tactile sensitivity, contact threshold, slip detection threshold; Delay parameters: sensor sampling delay, bus communication delay, control link delay, actuator response delay; Environmental adaptability parameters: temperature drift parameters, humidity effect parameters, surface friction condition labels, load range parameters; Reliability and lifespan parameters: wear coefficient, historical failure frequency, drift trend parameters, and health indicators; Task-related parameters: applicable skill set, recommended control mode, typical action block length, and allowed speed / force threshold range.

[0039] This design is compatible with real-world datasets across different embodiments, such as OpenX-Embodiment, as well as environment data generated by world-based models like Cosmos and Genie2. Publicly available information shows that OpenX-Embodiment has aggregated over 1 million real robot trajectories, 22 robot embodiments, and 60 datasets; Cosmos provides video cleaning, pre-trained world-based models, and tokenizers for physical AI; and Genie2 can generate interactive 3D environments from cue images.

[0040] The edge-side collaborative control center is primarily responsible for: long task decomposition; skill selection; event memory and context management; uncertainty assessment; device-level residual identification; and mesoscale replanning when edge-side computing power is insufficient or uncertainty increases. The edge does not directly send each motor control variable; instead, it sends constraint-checked sub-objectives, skill identifiers, and action block parameters to the edge.

[0041] The edge-side embodied execution device is primarily responsible for: multimodal perception; real-time state fusion; short-term dynamics prediction; action block decoding; real-time servo control; security shielding; and online digital twin synchronization. The edge only handles the tasks that must be completed in the closed loop at the millisecond to tens of millisecond level, thus enabling the world model to run under limited computing power.

[0042] Specific indicators in uncertainty assessment should include at least the following: Perceived uncertainty: target detection confidence, segmentation confidence, pose estimation error, depth / point cloud variance, target tracking stability; State uncertainty: prediction variance, observation residual, and filter covariance of key state quantities in the unified state package; Dynamic uncertainties include: deviation between model-predicted trajectory and actual feedback trajectory, contact force prediction error, slip probability, and collision risk value. Task uncertainty: probability of completing the current subtask, probability of satisfying the postcondition, skill success rate, and ambiguity in task decomposition; Model uncertainty: Output dispersion, confidence entropy, and output bias after quantization for different models / different sampling results; Execution uncertainties include: control delay jitter, actuator response deviation, clamping stability, and end-effector positioning error. Environmental uncertainties include: the magnitude of changes in illumination, the degree of shading, the intensity of external disturbances, and the amount of workpiece position drift.

[0043] This invention designs a unified embodied state package, denoted as: Z_t equals {I_t,P_t,F_t,J_t,G_t,M_t,D_i,U_t,C_t}. Wherein: I_t: visual observation, including RGB, depth, semantic segmentation, and keyframe features; P_t: geometric and contact observation, including point cloud, pose, contact point, and occupied voxel; F_t: force / torque / tactile observation; J_t: joint state, including angle, angular velocity, current, temperature, and encoder bias; G_t: task graph and logic state graph; M_t: event memory token; D_i: device fingerprint of the i-th device; U_t: uncertainty vector; C_t: security constraints, process constraints, and computing power budget.

[0044] The device fingerprint D_i is one of the key innovations of this invention, and includes at least one or more of the following types of parameters: Joint zero bias, friction curve, backlash curve, transmission smoothness, camera intrinsic and extrinsic parameters, gripper elasticity coefficient, tactile calibration coefficient, actuator delay, network jitter spectrum, temperature drift parameter, aging parameter.

[0045] The steps to implement device fingerprinting are as follows: First, divide robots of the same or similar models into equipment clusters. Then, based on the calibration data, motion data, force data and historical operation data of the batch prototypes, establish a common parameter template for the equipment cluster, denoted as D_cluster.

[0046] The device cluster template includes at least kinematic parameter templates, dynamic parameter templates, visual parameter templates, execution chain delay templates, and end-effector parameter templates.

[0047] For a single robot to be deployed, a standardized data collection process is executed to obtain the device's raw feature data.

[0048] The original feature data includes at least: joint reciprocating motion data under no-load conditions, drive current data under different speeds and loads, end-effector pose measurement data, visual calibration images, gripper opening and closing test data, tactile sensing test data, communication link delay data, temperature rise process data, and aging operation data.

[0049] The original feature data is identified and fitted to extract the individual feature parameters of the single device.

[0050] The extraction results should include at least one or more of the following: Joint zero bias, friction curve, backlash curve, transmission smoothness, camera internal and external parameter deviation, gripper elasticity coefficient, tactile calibration coefficient, actuator delay, network jitter spectrum, temperature drift parameter, and aging parameter.

[0051] The extracted individual feature parameters are encoded into a single-device fingerprint Di.

[0052] The D_i can be represented as a parameter vector, feature matrix, lookup table, implicit embedding vector, or a combination thereof.

[0053] A single device fingerprint is represented as: D_i equals D_cluster plus ΔD_i; where D_cluster is the device cluster template, and ΔD_i is the single device residual parameter relative to the device cluster template.

[0054] The single-device fingerprint D_i is injected into the high-level task world model, the mid-level skill world model, the low-level dynamics world model, and the digital twin module to correct state prediction, action generation, contact determination, and latency compensation.

[0055] For example: use joint zero bias and transmission compliance to correct end-effector pose prediction; use friction curves and backlash curves to correct dynamic prediction; use camera intrinsic and extrinsic parameter deviations to correct visual pose estimation; use execution chain delay and network jitter spectrum to correct control timing.

[0056] During robot operation, the residuals between actual observations and model predictions are continuously collected, and ΔD_i is updated online based on these residuals. If a long-term stable deviation is observed, the single-device fingerprint is updated; if the deviation exceeds the current device cluster boundary, re-clustering or the creation of a new device cluster template is triggered.

[0057] This enables a device fingerprint mechanism that combines "cluster-level sharing + individual correction + online evolution".

[0058] Robots of the same model can first share a device cluster template, and then add individual device fingerprints, thus taking into account both batch deployment and individual differences.

[0059] To facilitate unified state exchange between different modules, this invention establishes a unified state bus. This bus transmits Z_t in timestamp-aligned state frames and allows simultaneous reading of the following three types of world models: The high-level mission world model reads G_t", M_t", U_t", and C_t; The mid-level skill world model reads I_t", P_t", G_t", Di", and M_t; The low-level dynamic world model reads I_t", P_t", F_t", J_t", and D_i.

[0060] In this way, logical planning, visual perception, contact control, and equipment difference compensation can work together based on a unified state at the same moment.

[0061] This invention employs a hierarchical physical agent consisting of a three-layer world model, device fingerprinting, and online residual synchronization. The high-level task world model is used to handle long-term objectives, task graph construction, sub-objective decomposition, and failure rollback.

[0062] Its inputs are: user language instructions or task scripts; current logic state diagram G_t; event memory M_t; safety / process constraints C_t.

[0063] Its output is: the subtask sequence T equals {g_1,g_2,...,g_n}; the precondition Pre(g_k) for each subtask; the postcondition Post(g_k) for each subtask; the set of allowed skills Skill(g_k) for each subtask; and the backoff node Back(g_k) for failure.

[0064] The high-level task world model does not directly output motor control quantities, but rather outputs a verifiable task graph. For example, for a long task such as "grabbing a plug, inserting it into the socket, and then closing the cover," the high-level model can be decomposed into: "Position Plug" → "Grab Plug" → "Move to Socket" → "Precise Alignment" → "Micro Insertion" → "Detected in Position" → "Release Grippers" → "Close Cover".

[0065] Each step defines preconditions and postconditions. For example, the postcondition for "micro-insertion" is not simply a change in video, but rather "pose error less than the threshold + insertion force curve within the target range + plug root occlusion relationship conforming to the insertion state".

[0066] The high-level task world model preferably adopts one of the following implementation methods: language model + symbolic constraint solver; graph neural network + task state machine; task planning Transformer; embodied task knowledge graph and causal reasoning module.

[0067] The specific technical details of the high-level mission world model can be implemented separately according to mission type: 1. Language model + symbolic constraint solver It is suitable for open semantic tasks, natural language instruction tasks, and multi-constraint tasks, such as "tidy up the desktop and put the red cup on the designated tray" and "take materials first and then assemble, and do not enter the restricted area".

[0068] The language model is responsible for parsing natural language instructions into candidate subtask sequences, target objects, operation order, and constraints; the symbolic constraint solver is responsible for performing precondition checks, resource constraint checks, timing constraint checks, and security constraint solving on the candidate subtask sequences, and outputting a task graph that satisfies the constraints.

[0069] Its core fusion mechanism is as follows: the language model generates task candidates, and the symbolic solver performs executability screening and constraint correction on the candidate results.

[0070] 2. Graph Neural Network + Task State Machine It is suitable for industrial tasks with fixed processes, fixed workstations, and clear state transitions, such as assembly, loading and unloading, sorting, palletizing, and inspection. The graph neural network is used to model the relationship graph between workpieces, tools, workstations, fixtures, and robots; the task state machine is used to define discrete task stages such as "to be gripped - gripped - aligned - assembled - inspected".

[0071] Its core fusion mechanism is as follows: the graph neural network outputs the current relation state and node priority, and the task state machine triggers task phase transitions based on the relation state.

[0072] 3. Task Planning Transformer It is suitable for complex tasks with long time sequences, multiple steps, and context dependencies, such as "opening the door - retrieving the item - moving - placing - closing the door" and "multi-step flexible assembly".

[0073] The task planning Transformer receives historical task sequences, current environment status, completed nodes, and failure records, and predicts subsequent subtask sequences, rollback nodes, and branch paths.

[0074] Its core fusion mechanism is to use the long context modeling capability of Transformer to encode historical execution information, current state and target task in a unified manner, and output a global task plan.

[0075] 4. Embodied Task Knowledge Graph and Causal Reasoning Module It is suitable for tasks with strong rules, dependencies on causal relationships, and the need for explanation, such as plugging, locking, equipment maintenance, and kitchen operations. The knowledge graph stores the relationships of "object-attribute-operation-result-constraint"; the causal reasoning module deduces the next task and the cause of failure based on "what state change will be caused by performing a certain operation".

[0076] Its core integration mechanism is as follows: the knowledge graph provides structured task priors, and the causal reasoning module derives the task path and backtracking path based on the current state and the target state.

[0077] The mid-level skill world model is used to transform high-level sub-goals into executable skill parameters. A skill is a reusable unit of operation, such as grasping, placing, rotating, opening a door, wiping, inserting, tightening, delivering, or inspecting.

[0078] The inputs to the mid-level skill world model are: current subtask g_k; visual geometric state I_t", "P_t; current logical state G_t; device fingerprint Di; and historical event memory M_t.

[0079] Its output is: skill number s_k; skill parameters α_k, such as grab point, end pose, approach direction, contact force threshold, action block length; expected observation template Ω_k; termination condition ξ_k; failure repair strategy R_k.

[0080] In this invention, the mid-level skill world model includes at least two key sub-modules: (a) Skill parameterization generator Based on the current subtask and environmental state, skill parameters are generated. For example, for "grabbing a cup", the skill parameters include: approach direction, gripping width, gripper speed, allowable tilt angle, and force threshold.

[0081] (b) Skill Termination and Verifier Determine whether the skill has met the termination conditions. For example, "secured", "inserted", "rotated into position", "target area wiped and covered".

[0082] Unlike existing solutions, this invention requires each skill to have an observation template and a failure recovery strategy. For example, the failure recovery strategy for the "insertion" skill could be: first perform a slight retraction → perform a micro-probing action → update the hole pose estimation → then perform a second insertion. This is not a simple retry, but a recovery with information gain.

[0083] The hole pose estimation update is based on visual, force, tactile information and end-effector pose deviation after insertion failure. The residual between the current estimate and the actual contact state is calculated, and the hole center position and pose are corrected by filtering, optimization or multimodal fusion network.

[0084] The low-level dynamics world model is designed for real-time closed-loop operation at the edge. Its inputs include: short-term visual features; point cloud / pose / contact information; force / tactile information; joint states; device fingerprints; and skill parameters issued from the middle layer.

[0085] Its output is: future short-term state prediction. Action block A_(t:t+h-1); Risk value R_t; Execution permission flag Permit_t.

[0086] The core function of low-level dynamic world models is not long-form video generation, but rather short-term prediction, contact assessment, and action selection. A preferred approach is to employ one or more hybrid implementations of the following: state-space models; lightweight Transformers; physical prior constraint networks; differentiable physics modules; and hybrid controllers combining learning models with small-step MPC.

[0087] This invention further introduces the execution permission flag Permit_t. Specifically, before the end side actually issues an action, it must simultaneously satisfy the following: logical preconditions are met; geometric reachability conditions are met; low-level dynamic predictions have not triggered constraints such as collisions, excessive forces, excessive currents, and out-of-bounds errors; and uncertainty is below a threshold or has been taken over by the upper layer.

[0088] An actual action command is issued only when all of the above conditions are met. This transforms "able to generate an action" into "an action that can be executed after verification".

[0089] (1) The logical preconditions are satisfied This indicates that the preceding state on which the current subtask depends has been established, such as: the target has been identified, the previous subtask has been completed, the gripper is currently empty / securely gripped, the workpiece is in place, the current task mode is correct, and it is not in a faulty or disabled state.

[0090] (2) The geometrically reachable condition is satisfied. This refers to the existence of a feasible spatial path and target pose for the robot in the current environment, including: the target pose is within the workspace, the inverse kinematics exists, the end effector pose is reachable, the path has no static collisions, and the spatial clearance requirements between the fixture / workpiece / obstacle are met.

[0091] (3) The dynamic and safety constraints are satisfied This means that the short-term future actions predicted by the low-level dynamic world model will not trigger safety constraints, including: no collisions, no excessive force, no excessive flow, no joint overruns, no speed / acceleration exceeding limits, and no clamping slip instability.

[0092] (4) The uncertainty condition is satisfied This means that the uncertainty of current perception, state estimation, dynamic prediction and task judgment is below the allowable threshold; if it is above the threshold, it is necessary to confirm whether the edge or cloud-side upper-layer module has taken over the replanning or high-precision inference.

[0093] If the logical preconditions are not met, further verification will not be performed, the action will be directly prohibited, and the system will return to the high-level task world model or the mid-level skill world model to perform task rearrangement, precondition step completion, or state restoration.

[0094] If the geometric reachability condition is not met, further verification will not be performed, the action will be directly prohibited, and re-sensing, target relocation, path replanning, posture adjustment or skill change will be triggered.

[0095] If the dynamic and safety constraints are not met, the issuance of actions is prohibited, and at least one of the following actions is preferably performed: reduce speed, shorten action step size, switch to conservative control mode, modify approach direction, insert micro-probe action, or roll back to the most recently verified node.

[0096] If the uncertainty conditions are not met, the current action is prohibited from being executed directly, and at least one of the following actions is triggered: increase model accuracy, request edge / cloud side takeover, perform supplementary observations, perform micro-probing actions, or wait for manual takeover.

[0097] Actions are only allowed to be sent from the endpoint when all four conditions are met; if any condition is not met, the current action is blocked and the corresponding correction, rollback, replanning, or takeover process is initiated.

[0098] For mass production scenarios, this invention first divides the same type of robot into several equipment clusters according to mechanical structure, actuator batch, sensor batch, camera scheme, and tool head scheme.

[0099] Establish cluster-level template parameters for each device cluster: A_cluster={A_kin,A_dyn,A_vis,A_delay,A_tool}; which include: kinematic parameters; dynamic parameters; visual parameters; time delay parameters; and tool-side parameters.

[0100] For each specific device i, the present invention extracts the device fingerprint at the time of factory delivery or initial deployment and forms a single device residual: A_i equals A_cluster plus ΔA_(device,i); the ΔA_(device,i) includes, but is not limited to: encoder zero-position deviation; joint friction spectrum; hysteresis and compliance deviation; tool installation error; camera extrinsic parameter deviation; execution chain inherent delay; gripper elasticity / wear characteristics.

[0101] During deployment, an online correction variable is introduced: A_i(t) = A_cluster + ΔA_(device,i) + ΔA_online(t); where ΔA_online(t) is updated by real-time observation residuals.

[0102] The residual can come from: visual pose error; tactile event error; force curve error; error between actual trajectory and twin trajectory; network delay drift; temperature drift and wear drift.

[0103] The present invention preferably uses one or more of the following methods for updating: extended Kalman filtering; particle filtering; factor graph optimization; differentiable residual backpropagation; multi-model switching estimation.

[0104] To further address the sim2real challenge, this invention introduces a micro-probing action mechanism. When the endpoint detects significant uncertainty, but not enough to directly abort the task, the system automatically generates a low-risk micro-action, such as: a slight lift and downward view; a light touch and rebound; a small-angle trial insertion; a slight sway; or a small-range scan.

[0105] The system uses the observations of these micro-motions to update pose, friction, contact relationships, or gripping state. This mechanism allows the robot to proactively acquire information before failure, thereby significantly reducing erroneous executions caused by insufficient cognition.

[0106] This invention divides the edge-side world model and policy model into multiple precision versions, such as FP16, INT8, INT4, or mixed precision versions. The edge dynamically selects the inference precision based on the following factors: current task risk level; edge latency budget; temperature / power consumption budget; uncertainty level; and the type of skill being executed.

[0107] For example, a lower precision can be used in the coarse positioning stage before grasping; switch to high precision in the insertion, tightening, and contact-sensitive stages; and prioritize increasing precision when there is high uncertainty, rather than blindly continuing to execute with low precision.

[0108] To avoid the problem of "low-bit models appearing to run but actually exhibiting implicit distortion," this invention trains a confidence mapping head for each quantization version. That is, the edge output not only shows the action but also the "confidence of that action under the current quantization version." If the confidence level is below a threshold, the precision is automatically increased or the data is unloaded to the edge.

[0109] This invention proposes a dual-channel encoding for action blocks that is more suitable for end-side execution: the first channel is a smooth action channel, representing end-point trajectory, attitude changes, and continuous control of the gripper; the second channel is an event pulse channel, representing sparse events such as contact instant, force threshold trigger, gripper switching, braking, and retraction.

[0110] The specific approach is as follows: First, decompose the future H-step action sequence into smooth principal components and sparse event components. Then, compress and encode the smooth principal components and index and encode the event components.

[0111] The endpoint only decodes the first K steps of the current rectified horizon. This preserves the information density of long action blocks while avoiding decoding all tokens in every control cycle. Long task failures often occur not because the current frame is unclear, but because the system forgets "which step it has already reached". Therefore, this invention does not save all frames, but saves event memory tokens.

[0112] Event memory tokens can be written triggered by the following events: successful capture; contact with the target surface; disappearance or reappearance of a visually critical object; occurrence of a force peak; termination of a subtask; manual intervention; and occurrence of a safety shielding action. During replanning, the system prioritizes retrieving event memories rather than feeding the entire long video to the model. This significantly reduces context length and edge memory usage.

[0113] The long task execution process of this invention is as follows: S601: Receive target instructions and process constraints Instructions can come from natural language, MES work orders, teaching scripts, or the host control system; Process constraints include: speed limit, force limit, restricted areas, accuracy requirements, cycle time requirements, and priority.

[0114] S602: Generate a unified state packet and instantiate the device digital twin. The system collects visual, depth, joint, force, and tactile information to generate a unified state package Z_t; at the same time, it instantiates a digital twin of the current device based on the device cluster template and device fingerprint.

[0115] S603: High-level mission world model generates sub-objective trees and fallback nodes The high-level model generates a sub-target tree based on the target and the current state diagram. The sub-target tree preferably adopts the following structure: current node; prerequisites; success conditions; failure conditions; fallback node; set of alternative skills.

[0116] S604: Mid-level skill world model generates current skills and parameters. The system selects appropriate skills from the skill library based on the current sub-objective and generates parameters; for example, for the task of "retrieving a box after opening the cabinet door", the middle layer can select in sequence: "handle grabbing skill" → "door pulling skill" → "collision avoidance and reaching skill" → "box grabbing skill" → "retreat skill".

[0117] S605: Low-level dynamic world model performs short-term predictions and generates action blocks. The low-level model makes short-term predictions for the next h steps based on current observations and skill parameters, and outputs action blocks; If abnormal contact, gripping slippage, posture out of bounds, or conflict between the chassis and the robotic arm are predicted, the risk value will be set high.

[0118] S606: Performing a license check The safety shielding module checks the following: whether the logical preconditions are met; whether the geometric reachability and collision avoidance are met; whether the risk value is within the permissible range; whether the current uncertainty exceeds the threshold; whether the confidence level of the current quantization model is sufficient; and whether hardware-level safety constraints are triggered. If not, proceed to S611.

[0119] S607: Issue servo actions The action block that passes the permission check is sent to the servo controller or end effector; The control method can be joint space, operating space, impedance control, force control, or hybrid control.

[0120] S608: Execute and collect feedback The system collects feedback from visual, tactile, force, encoder, current, temperature, and network latency data during the execution process.

[0121] S609: Post-condition verification and progress submission This invention requires that "post-condition verification" be performed after each subtask is completed; When verification is successful, the system marks the node as "verified and submitted"; If the task fails, do not proceed to the next subtask; instead, roll back or repair it. This avoids the situation where "one wrong step leads to a complete mistake" in long tasks.

[0122] S610: Rollback to the most recently verified node The system saves a snapshot of the state of the most recently verified node, including: subtask number; logical state diagram; key object pose; clamping state; event memory token; and device twin parameters. In case of failure, it can roll back to the most recently trusted state instead of restarting the entire task.

[0123] The rollback rules include: A rollback is triggered when any of the following conditions occur in the current subtask g_k: the postcondition verification of the current subtask fails; during execution, the target is lost, the pose deviation exceeds the threshold, the contact is abnormal, the force / torque is abnormal, the slippage or collision risk exceeds the threshold is detected; the number of consecutive retries for the current skill exceeds the set threshold; the current state does not meet the safety permission conditions for continuing execution; or the model uncertainty or environmental uncertainty exceeds the set threshold.

[0124] Rollback does not mean returning to the previous step, but rather rolling back to the most recently verified successful task node. The most recently verified successful task node refers to the most recently submitted sub-task node that has passed its postcondition verification and been confirmed by the system.

[0125] The fallback node can be selected according to the following hierarchy: first fall back to the starting node of the current subtask g_k and re-execute the current subtask; if the current subtask fails because the preceding state is distorted, fall back to its direct predecessor subtask g_(k-1); if the predecessor node's state is also unreliable, fall back to the most recently verified successful ancestor node with a complete state; if the entire local task chain fails, fall back to the root node or stage starting point of the high-level task world model's replanning.

[0126] After rollback, instead of immediately retrying mechanically, at least one of the following operations is performed: restore the state snapshot corresponding to the rollback node; re-perceive the target, pose, contact state, and environment state; call the mid-level skill world model to reselect the skill or regenerate the skill parameters; call the low-level dynamics world model to re-predict short-term actions; perform micro-probe actions to supplement state information if necessary; if the failure is due to insufficient model confidence, switch to a higher-precision model or request edge / cloud-focused planning.

[0127] After completing the state recovery and re-verification after the rollback, the system restarts execution from the subtask corresponding to the rollback node Back(g_k). If the current subtask is repairable, it continues at the atomic task level. If the current subtask is not repairable, the remaining subtask sequence from the rollback node to the target node is regenerated by the high-level task world model, and then execution continues.

[0128] If the number of repeated failures of the same rollback node exceeds the set threshold, a higher-level rollback, strategy switch, manual takeover, or task termination will be triggered.

[0129] S611: Micro-detection / Local repair / Accuracy improvement / Replanning The system performs the following actions based on the failure type: micro-probing; local relocation; accuracy adjustment; switching to a more conservative technique; requesting edge-oriented planning; requesting cloud-side rare scenario simulation.

[0130] The S611 processing method prioritizes "local first, then global; low cost first, then high cost; edge side first, then edge cloud".

[0131] The system first classifies anomalies based on failure type, including at least: perception failures: target loss, occlusion, large pose deviation; contact failures: insertion failure, slippage, force anomalies; model failures: insufficient confidence, large quantization error, excessive uncertainty; skill failures: continuous failure of the current skill or incompatibility; planning failures: no solution for local paths, failure of preconditions, broken task chains; and rare scenario failures: exceeding the coverage of the local experience base or side model.

[0132] Process in the following order: (1) Micro-detection Suitable for mild uncertainty in perception or mild uncertainty in contact, with the lowest cost, and should be prioritized for implementation; (2) Local relocation It is applicable when the target position, small-range attitude, or contact status can be corrected through re-sensing. (3) Increased precision It is suitable for situations where the confidence of the current low-bit model is insufficient, but the task itself can still be partially continued. (4) Switch to a more conservative skill This applies to situations where the current skill fails multiple times, but the task can still be completed using a low-risk alternative skill; (5) Request edge-focused planning This is applicable when a local action layer cannot be repaired and it is necessary to regenerate sub-targets, paths, or skill parameters. (6) Request cloud-side rare scenario simulation It is suitable for situations where abnormal patterns are rare, cannot be reliably handled by the side, and require calling large models or long-term counterfactual inferences. It has the lowest priority but the strongest capabilities.

[0133] If a problem can be solved through local observation, it should not be replanned first; if a problem can be solved through edge repair, it should not be requested from the edge side first; if a problem can be solved through edge replanning, it should not be requested from the cloud side first; if the current risk is higher than the safety threshold, low-priority steps can be skipped and the process can directly proceed to conservative skills, edge side takeover, or cloud side takeover.

[0134] The overall process of offline training, mass production calibration, and online learning in this invention is as follows: collect real data, simulation data, synthetic data, and historical fault data; construct a device cluster template library and a single device fingerprint library; train the world base model and the synthetic environment generation model; jointly train high-level / mid-level / low-level world models; add physical consistency loss, cross-layer consistency loss, and computing power budget loss; distill into edge-side models and terminal-side models; quantize and generate multi-precision versions and confidence mapping heads; perform hardware-in-the-loop verification, fault injection verification, batch deployment, online residual synchronization, log backfeeding, and periodic retraining.

[0135] This invention allows the use of one or more of the following types of data: real robot teaching data; factory production logs; historical fault and rework data; publicly available embodied datasets; digital twin simulation data; counterfactual data generated from world models; and fault injection data from rare scenarios.

[0136] Among these, real-world data across the embodiment and data generated from the world-foundation model can be used to improve generalization capabilities. Publicly available information shows that the OpenX-Embodiment and RT-2 / OpenVLA routes demonstrate the importance of real-world data across the embodiment for general operational strategies; Cosmos and Genie2 illustrate that the world-foundation model can serve as a synthetic environment and a predictive foundation.

[0137] To ensure that the model is predictable, executable, and adaptable to edge devices, this invention preferably employs the following joint loss: L equals λ1 multiplied by L_pred, plus λ2 multiplied by L_phys, plus λ3 multiplied by L_cross, plus λ4 multiplied by L_task, plus λ5 multiplied by L_safe, plus λ6 multiplied by L_budget, plus λ7 multiplied by L_quant.

[0138] Wherein: L_pred: future state prediction loss; L_phys: physical consistency loss, including contact, collision, force threshold, and velocity threshold; L_cross: cross-layer consistency loss, aligning high-level logical states with low-level observations; L_task: task success-related loss; L_safe: safety constraint loss; L_budget: edge computing power budget loss; L_quant: performance preservation loss after quantization. L_budget is a significant innovation of this invention, directly incorporating "edge operability" into the training objective, rather than passively pruning it after training.

[0139] This invention preferably employs a three-stage distillation process: the cloud-side open-world model is distilled into the edge-side task / skill model; the edge-side model is distilled into the end-side short-term model; and the end-side model is then distilled into a multi-precision version.

[0140] Distillation content includes not only action output, but also: predicted state; risk value; uncertainty; termination condition; and quantified version confidence level.

[0141] The end-side execution control device of the present invention includes at least: 1. Main Control Processing Unit Used to execute low-level dynamic world models, action block decoding, security shielding logic, and real-time control logic; can be implemented by CPU, GPU, NPU, FPGA, ASIC, or a combination thereof.

[0142] 2. Safety processing unit Operating independently of the main control processing unit, it is used to monitor emergency stops, light curtains, access control, joint overruns, driver failures, and heartbeat loss. When the main control fails or the action is abnormal, the safety processing unit can hardwire the servo enable or brake output.

[0143] 3. Multimodal sensing interface unit It can access data such as vision, depth, force, touch, encoder, IMU, temperature and current.

[0144] 4. Time synchronization unit The timestamps of samples from different sensors are aligned in a unified manner in order to construct a unified state packet.

[0145] 5. Real-time control unit Generate joint commands, end-effector trajectories, impedance parameters, or force control commands based on the output of the end-side model.

[0146] 6. Communication Unit It communicates with the edge and cloud sides, and with the servo drive or chassis controller.

[0147] 7. Model Storage and Security Unit It supports storage-side models, multi-precision versions, device fingerprints, and log caching, and can verify model versions.

[0148] The key point of this invention is that the AI ​​main control link and the security hardware link are independent of each other. Even if the world model, VLA, or scheduling module malfunctions, the security link can still forcibly pull down the enable signal to prevent dangerous actions from continuing to be executed.

[0149] Within a single edge control cycle, the following steps are preferably executed sequentially: Acquisition and Synchronization Phase: Acquire visual, force, and joint states and synchronize them with time; State Fusion Phase: Generate a unified state package; Short-Term Prediction Phase: Predict the next h steps using a low-level dynamic world model; Skill Verification Phase: Verify skills based on the current subtask, skill template, and postcondition template; Safety Permission Phase: Perform checks on logical constraints, geometric constraints, dynamic constraints, and hardware safety constraints; Action Issuance Phase: Issu the first K steps of the current action block; Feedback Write-back Phase: Write the execution feedback into the event memory and online twin synchronization module.

[0150] The key to this timing is that the high-level and middle-level layers do not need to run in every millisecond cycle; the edge only executes the steps that must be completed in a short-term closed loop.

[0151] This ensures real-time performance.

[0152] Accordingly, this invention also provides an embodied intelligence platform system device based on a world model, which includes at least: a data engine module; a world base model module; a device cluster template library module; a device residual identification module; a high-level task world model module; a mid-level skill world model module; a low-level dynamics world model module; a dynamic precision scheduling module; an online digital twin synchronization module; an event memory module; a security shielding module; a real-time control module; a model distillation and quantization module; and a log feedback and closed-loop learning module. Each module can be deployed on the same device or distributed across cloud, edge, and endpoint devices.

[0153] In Example 1, which is a 3C precision assembly scenario, the task is to "grab the flexible ribbon cable and insert it into the connector, and then close the latch".

[0154] This task requires long duration, multi-stage alignment, micro-contact, and high precision.

[0155] The execution process of this invention is as follows: 1. The high-level task world model divides tasks into "Identify cable", "Grab cable", "Move to connector", "Align", "Micro-insertion", "Detection in place", "Release", and "Lock closure"; 2. The mid-level skill world model selects insertion skills for the "micro-insertion" stage, along with a micro-probing strategy; 3. Low-level dynamic world model combined with force perception and device fingerprinting predicts changes in insertion force; 4. If the insertion force is abnormal, the system will not continue to push directly, but will instead perform a slight retraction, a light touch detection, an attitude correction, and a second insertion. 5. Each insertion step must satisfy the postcondition verification before it can be submitted to the next node.

[0156] In Example 2, which is a logistics sorting and palletizing scenario, the task is to "grab packages of different shapes from the mixed bins, classify them and place them into the corresponding pallets, and complete the palletizing".

[0157] The execution process of this invention is as follows: 1. The high-level model first generates a sub-target tree consisting of "identification-grabbing-classification-placement-stack verification"; 2. For the middle layer, select either the suction skill or the claw skill based on the package's material, size, and orientation; 3. The equipment cluster template allows multiple robots of the same model to share a basic model; 4. Single-device residual compensation for different robot gripper wear and camera deviation; 5. If a package slips, the event memory module records a "failed grabbing event", and the system switches to a more conservative grabbing strategy and increases the accuracy of the local model.

[0158] In Example 3, a long-task scenario for a home service robot is described, where the task is to "open the cabinet door, take out a cup, fill it with water, place it on the table, and then close the cabinet door." This task is characterized by multiple rooms, multiple objects, multiple sub-tasks, and a long time sequence.

[0159] The execution process of this invention is as follows: 1. Construct a task tree and define multiple fallback nodes in the high-level task world model; 2. The mid-level skill world model switches between "opening the door - taking the cup - moving - getting water - placing - closing the door"; 3. End-side event memory only retains key events, such as "the door is open," "the cup is held firmly," and "the water has been filled." 4. If the target is lost during the move, the system will prioritize retrieving the event memory and performing a scan for micro-detection, rather than replanning the entire task. 5. If visual vapor or reflections increase uncertainty during the water connection stage, the system will automatically request side replenishment for replanning or improve accuracy operation.

[0160] All data collection and extraction in this invention are carried out under compliant and legal conditions.

Claims

1. An embodied intelligent control method based on a world model, characterized in that, include: Collect multimodal sensor data and device status, generate a unified status package containing device fingerprints, and instantiate a single device digital twin based on a device cluster template; The overall task is decomposed into a sequence of verifiable sub-objectives with preconditions, postconditions, and fallback nodes by a high-level task world model. The mid-level skill world model generates corresponding skill parameters, while the low-level dynamic world model makes short-term predictions and generates action blocks. The logical preconditions, geometric reachability conditions, safety constraints, and uncertainty conditions are verified sequentially. Once all conditions are met, the action is issued and executed. Collect execution feedback, verify the post-conditions of the subtask, and if they pass, advance to the next sub-objective; otherwise, fall back to the most recently verified node and correct the digital twin parameters.

2. The embodied intelligent control method based on a world model according to claim 1, characterized in that, In the process of generating a unified state package containing device fingerprints, a device cluster template for similar devices is established, standardized calibration data and operating data of single devices are collected, individual feature parameters of joint transmission visual time delay temperature drift and aging are extracted, a residual fingerprint relative to the cluster template is formed, and a three-layer world model and digital twin module are injected to complete the correction of state prediction and action generation.

3. The embodied intelligent control method based on a world model according to claim 2, characterized in that, In the instantiation of a single device digital twin based on a device cluster template, a dual mechanism is adopted: an offline mass production template closed loop and an online device synchronization closed loop. The offline closed loop updates the template by combining device cluster partitioning and cluster-level simulation training with batch data backhaul. The online closed loop continuously corrects the residual parameters of a single device by observing the residual in real time.

4. The embodied intelligent control method based on a world model according to claim 3, characterized in that, The high-level task world model decomposes the total task into different implementation methods according to the task type. For open semantic tasks, a language model is used in combination with a symbolic constraint solver. For industrial process tasks, a graph neural network is used in combination with a task state machine. For tasks with strong rules and requiring causal interpretability, an embodied task knowledge graph and a causal reasoning module are used.

5. The embodied intelligent control method based on a world model according to claim 4, characterized in that, The rollback to the most recently verified node includes prioritizing the rollback to the starting node of the current subtask; if the preceding state is distorted, rolling back to the direct predecessor node; if the local task chain fails, rolling back to the root node of the high-level replanning; after the rollback, restoring the state snapshot and re-sensing before continuing execution.

6. The embodied intelligent control method based on a world model according to claim 4 or 5, characterized in that, In the low-level dynamic world model, short-term prediction and action block generation are performed using a dual-channel encoding method. Low-frequency smooth principal components of the action sequence are extracted through frequency domain transformation or smooth curve fitting. Sparse detection is performed on the abrupt change points in the reconstructed residual to obtain event components. The end-side only decodes the preceding actions required in the current time domain.

7. The embodied intelligent control method based on a world model according to claim 6, characterized in that, The verification of the uncertainty conditions also includes a dynamic precision scheduling step, where multiple precision versions of the model are deployed on the edge, and the precision is switched according to the task risk level, latency budget and uncertainty level. The confidence level is output through the confidence mapping head jointly trained with the quantization model. If the confidence level is lower than the threshold, the precision is increased or the data is offloaded to the edge cloud.

8. The embodied intelligent control method based on a world model according to claim 4 or 7, characterized in that, In the process of reverting to the most recently verified node and correcting the digital twin parameters, when the uncertainty does not reach the shutdown threshold, a micro-probing action mechanism is triggered to automatically generate a trial action. The observation results are used to update the pose, contact, friction, or clamping state. After the state correction is completed, the verification is re-executed.

9. The embodied intelligent control method based on a world model according to claim 8, characterized in that, Once all conditions are met, actions are issued and executed. If the logical preconditions are not met, the process returns to the high-level task world model for task rearrangement or state restoration. If the geometric reachability conditions are not met, re-sensing or path replanning is triggered. If the dynamic and safety constraints are not met, the speed is reduced, the step size is shortened, or a conservative control mode is switched. If the uncertainty conditions are not met, the model accuracy is increased or a replanning request is made for the edge or cloud side.

10. The embodied intelligent control method based on a world model according to claim 9, characterized in that, The safety constraints are determined based on the total risk value obtained by weighted fusion of the collision contact pose, the task execution position, and the model risk components.

Images