Construction of a World Model for Energy Balance in Continuous Media and a Closed-Loop Control Method and System
By constructing a world model of energy balance in a continuous medium, the problems of volume drift and grid flipping in long-term autoregressive extrapolation of soft continuous media were solved, achieving high precision and real-time performance of non-co-located closed-loop control and ensuring the safe operation of the soft object.
Patent Information
- Application Number
- CN202611414447.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies are prone to volume drift and mesh flipping in long-term autoregressive extrapolation of soft continuous media, prediction divergence in non-co-located closed-loop control, difficulty in identifying hyperelastic material parameters, and lack of full-field mechanical safety monitoring inside the continuous medium.
A world model for energy balance in a continuous medium is constructed. By obtaining the three-dimensional discrete geometric representation of the target soft object, the deformation gradient is calculated and a volume penalty term is added. Variational energy minimization is performed, and closed-loop control is carried out by combining physical boundary conditions and full-field mechanical quantities. By using batch tensor parallel solution and physical parameter identification, volume conservation and high-precision autoregressive deduction are achieved.
It achieves volume conservation and grid stability through long-range autoregressive derivation, improves the accuracy and safety of non-co-located closed-loop control, and ensures millisecond-level real-time response and physically reliable control performance.
Smart Images

Figure CN123067162A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of embodied intelligence, robot control, continuum mechanics, and computer vision, specifically to a method and system for constructing a world model of energy balance in a continuum and for closed-loop control. It can be used in industrial robot assembly and medical surgery scenarios such as soft tissue manipulation, cable laying, and flexible component manipulation, and is particularly suitable for non-co-located soft closed-loop control scenarios where the drive and observation are decoupled in spatial configuration / degrees of freedom. Background Technology
[0002] Automated operation of near-incompressible soft objects (soft continuous media), such as soft tissues, cables, and flexible components, urgently requires a real-time and physically reliable forward model for long-term autoregressive inference and closed-loop control. High-fidelity finite element solvers offer high accuracy but insufficient computational speed, making them unsuitable for closed-loop control. Graph neural network world models can perform millisecond-level inference and have some generalization ability to unseen geometry, but their autoregressive inference (using the model's own output as the next input, progressing step by step over time) accumulates "physical illusions": volume drift and mesh flipping within tens of steps, leading to long-term prediction distortion and causing closed-loop control to collapse.
[0003] To address the aforementioned issues, existing technologies have proposed three main improvement schemes, all of which suffer from fundamental flaws. First, soft loss penalty: This involves adding physical constraint terms (such as energy or volume constraints) to the training loss function to encourage rather than enforce compliance with these constraints. However, the network cannot strictly guarantee physical realism when facing complex forces or deviating from the training distribution, and drift persists during long-term autoregressive inferences. Second, post-hoc algebraic projection: This applies rigid algebraic projection (such as Projective Dynamics) correction constraints after each prediction step. However, because rigid projection pushes the network's autoregressive input out of the training distribution, the resulting numerical perturbations are gradually accumulated and amplified. (See Table 2 and appendix for comparison). Figure 10 (c) The volume deviation increases exponentially within a few steps, accumulating to 10 in long-range simulations. 5 The magnitude of the change can induce mesh flipping. Third, global volume scaling: the prediction results are scaled proportionally to keep the total volume constant. However, the conservation of total volume does not equal local differential isochoricity. The object may be abnormally flattened in one part and non-physically stretched in another part while still maintaining the total volume. This does not hold true in the mechanics of continuum.
[0004] Furthermore, when existing neural world models are used for closed-loop control of non-co-location (i.e., the driving region where the actuator applies the load and the controlled observation region have at least decoupled or non-coincident components in terms of spatial configuration or degrees of freedom), the long-term divergence of their internal prediction models can cause the controller's rolling optimization to fall into a severely distorted state space, resulting in the control commands deviating from the physically feasible solution, thereby causing instability at the actuator end or even damage to the operated object.
[0005] Therefore, existing technologies lack an energy balance world model architecture that can embed volume conservation at the model forward solution level and stably support long-term autoregressive inference and high-precision non-co-located closed-loop control.
[0006] Further comparison with existing technologies: In the real-time optimization control of deformable objects / soft continuums, existing studies have used implicit / alternating direction methods to solve the constitutive potential energy of finite elements to obtain real-time deformation, and applied it to model predictive control (Deng Hao, Xiong Jing, Xia Zeyang, "An Implicit Solution Method for Dynamic Models for Deformable Object Manipulation", Robotics, 2024, 46(1):45-53). Its volume preservation depends on the solution framework itself rather than as an explicit design criterion, and it does not couple the learning network with the energy solution; there are also studies on real-time control of soft continuum robots using learning graph neural network dynamics combined with MPC (Eberhard et al., Graph Neural Model Predictive Control for High-Dimensional Systems, (2026), whose physical consistency is a learned property and can drift with training distribution; the deformable object neural world model reconstructed from video (PhysTwin, SoMA, DeformMaster) is an implicit Gaussian / particle representation, and does not provide explicit element-wise mechanical quantities for safety monitoring; existing closed-loop action prediction schemes based on visual world models also lack continuous medium full-field stress and volume preservation constraints. In summary, the above schemes either treat volume conservation as a violable soft constraint / post-projection, or sacrifice full-field mechanical interpretability and safety boundary monitoring with implicit representation. This invention takes "near-incompressible volume conservation as the equilibrium property of the embedded energy world model forward solution" as the design criterion, and combines explicit element-wise full-field mechanical quantities (peak volume deviation, contact reaction force, energy density) safety monitoring and restraint, taking into account both real-time performance and safety boundaries, which is different from the above-mentioned existing technologies.
[0007] The advantages of this invention compared with the closest prior art (Deng Hao, Xiong Jing, Xia Zeyang, "An Implicit Solution Method for Dynamic Models for Deformable Object Operations", Robotics, 2024, 46(1):45-53) are as follows: (1) Deformation solution framework: The prior art directly solves the constitutive potential energy of the material using the implicit / alternating direction method (ADMM), and the volume retention depends on the numerical properties of the solution framework itself; This invention uses "near-incompressible volume conservation as the balance property of the forward solution of the embedded energy world model" as the explicit design criterion, incorporates the volume penalty term into the energy functional and satisfies it with the minimization of variational energy as the solution objective, so that the volume conservation is embedded in the optimization solution process, and effectively avoids the error accumulation caused by the subsequent destruction of the network input distribution by algebraic projection in the prior art from the mechanism perspective; (2) Embedded forward solution world model pipeline and conservation: The prior art usually incorporates energy / Constitutive solution is used as an offline calculation or single-step real-time control tool, rather than as an output for subsequent autoregressive deduction of the internal world model, and its volume preservation is not an explicit design criterion; the present invention uses variational energy minimization as the forward deduction engine of the energy world model, so that volume conservation is achieved by embedding the energy term in the optimization solution process, and can be further combined with warm-start relaxation with historical equilibrium state as the initial value and batch tensor parallel solution to achieve efficient and stable convergence in long-range autoregressive deduction; (3) Boundary loading and mesh independence: the prior art does not disclose the mesh independence of the contact boundary. The present invention uses Neumann surface force / contact load to eliminate displacement clamping boundary singularity, and the measured volume deviation across the mesh is maintained at 10⁻ 5Quantity; (4) Mechanical safety and monitoring: Existing technologies do not disclose applying safety penalties or displacement constraints to control costs using full-field mechanical quantities (volume deviation, contact reaction force, energy density), nor do they output node / unit-level mechanical state fields for monitoring. This invention provides explicit element-wise full-field mechanical monitoring and safety boundary control; (5) Computational power-delay cost: When existing technologies solve constitutive potential energy using implicit / alternating direction methods, single-step convergence usually requires tens to hundreds of full Jacobian or constraint iterations, with derivation delays reaching hundreds of milliseconds to seconds. It is difficult to perform real-time prediction and control within the batch tensor high-frequency model prediction control cycle with rolling horizon H=6 and candidate sampling K=64. Running; This invention uses batch tensor parallel deduction and variational gradient calculation to complete the parallel energy minimization solution of multiple candidate control sequences within millisecond-level ticks, taking into account both physical fidelity and control real-time performance; (6) Physical parameter identification and simulation to real-world transfer: Existing technologies mostly use free parameters or implicit representations to fit soft material properties, making it difficult to bridge the physical gap between simulation and reality and difficult to transfer to unseen operating conditions; This invention is based on explicit hyperelastic parameter inversion calibration of reference configuration mapping, and combines it with open-loop consistency verification criteria under unseen excitation conditions to achieve transferable physical parameter identification, ensuring that the calibration model can be directly deployed in real physical systems. The above distinguishing features, combined with the measured data in Tables 1 to 5 of the specification (comparison of long-range zero drift, closed-loop tracking accuracy, contact reaction force and energy density with black box ×7.7 / ×16.4), are sufficient to prove that this invention has outstanding substantive features and significant progress compared to the closest existing technology. Summary of the Invention
[0008] The technical problem to be solved by this invention is to overcome the technical defects of existing soft continuous medium forward prediction models, such as easy volume drift and mesh flipping in long-term autoregressive inference, the inability of soft penalties and subsequent algebraic projection to fundamentally solve volume conservation at the forward solution level, displacement clamping at the contact edge inducing boundary stress singularities that cause volume error to worsen with mesh subdivision, prediction divergence in non-co-located closed-loop control that easily leads to controller malfunction, difficulty in identifying hyperelastic material parameters leading to a physical gap between simulation and reality, and lack of full-field mechanical safety monitoring inside the continuous medium. The invention provides a method and system for constructing a world model of energy balance for continuous media and closed-loop control.
[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for constructing and extrapolating a world model of energy balance in a continuous medium, comprising: obtaining a three-dimensional discrete geometric representation of a target soft object, wherein the three-dimensional discrete geometric representation includes at least one of a polyhedral mesh, a voxel array, or a particle set; determining a geometric reference datum for the geometric representation, and obtaining reference mapping information characterizing local deformation based on the geometric reference datum; calculating the deformation gradient for each discrete element of the geometric representation based on the reference mapping information, and calculating the near-incompressible hyperelastic strain energy accordingly, wherein the near-incompressible hyperelastic strain energy includes an isochoric energy term and a volume penalty term. The volume penalty term monotonically increases as the element volume deviates from the incompressible reference and is weighted by the modulus parameter or stiffness weight, which characterizes the incompressible properties of the material, so that volume conservation is achieved by the energy term embedded in the optimization solution process. Under the preset physical boundary conditions, variational energy minimization numerical solution is performed on the near-incompressible hyperelastic strain energy during inference and deduction to solve the equilibrium displacement of the unconstrained physical degrees of freedom, so as to obtain the equilibrium state field that maintains volume conservation and construct the energy world model. The result characterizing the equilibrium state is output for autoregressive deduction at subsequent time steps, or the energy world model is used as the internal prediction model of the upper-level controller.
[0010] Furthermore, in one possible implementation, the variational energy minimization serves as a physically consistent equilibrium layer embedded in the forward inference pipeline of the energy world model. During inference, variational energy minimization is applied using the historical equilibrium state of the previous inference cycle as a warm-start initial value, updating the solution to an equilibrium state field that maintains volume conservation. The resulting solution is used for subsequent inferences or as an internal prediction model for upper-level controllers (such as rolling horizon optimization). The technical advantage is that this equilibrium layer achieves volume conservation as an embedded equilibrium property in the forward solution, independent of the specific generation method of the candidate state field.
[0011] Furthermore, in one possible implementation, the calculation of the deformation gradient includes: right-multiplying the deformation matrix corresponding to the current configuration by the inverse of the undeformed gradient matrix of the reference configuration to obtain a deformation gradient of the form F = A·A0⁻¹; when the discrete element (such as a tetrahedral element) encounters a critical or flip risk, a lower bound constraint is applied to the element volume ratio or the corresponding element's state is reinitialized. The technical effect is that a strict reference configuration ensures physical consistency of the deformation gradient, and a lower bound constraint prevents numerical flipping into an unsolvable state.
[0012] Furthermore, in one possible implementation, the variational energy minimization employs at least one of the following: the finite-memory BFGS quasi-Newton method (i.e., L-BFGS), the nonlinear conjugate gradient method, or the Newton-Raphson method. During the solution process, the energy curvature differences in different degrees of freedom directions are normalized or preprocessed (including row scaling). The technical effect is to alleviate the ill-conditioned energy potential surface conditions caused by near-incompressible bulk modulus (curvature differences can reach 10). 6 (by multiples), ensuring stable, efficient, and convergent numerical solutions.
[0013] Furthermore, in one possible implementation, the numerical solution for variational energy minimization under preset physical boundary conditions includes: applying Neumann surface forces or contact loads (including external potential energy terms) to the contact or loading regions related to external contact, introducing the external load in a continuously distributed force manner. The technical effect is: eliminating the boundary stress singularity induced by rigid displacement clamping at the contact edge across the entire surface, ensuring that volume deviation does not increase with discrete mesh subdivision, and achieving mesh-independent near-incompressible volume conservation.
[0014] Furthermore, in one possible implementation, the step of performing variational energy minimization numerical solution during inference to obtain an equilibrium field that maintains volume conservation includes: the variational energy minimization uses historical equilibrium states prior to the current inference cycle as warm-start initial values, and inputs these historical equilibrium states into the variational energy minimization solver for optimization iteration. The technical effect is that by using the preceding physical equilibrium configuration as the spatial starting point for energy optimization, it enables rapid and stable convergence to the true physical equilibrium state with a smaller iteration step budget on the extremely ill-conditioned energy potential surface caused by the high bulk modulus of the near-incompressible hyperelasticity, avoiding numerical oscillations and deformation distortions caused by cold-start search.
[0015] Furthermore, in one possible implementation, the autoregressive inference is a long-range open-loop inference; the method further includes: using the cell volume ratio deviation not exceeding a preset threshold as a continuous criterion at each moment of the inference, and continuing the autoregressive inference when the criterion is met. Its technical effect is: achieving zero-drift and mesh-flip-free long-range autoregressive inference, providing an explicit and quantifiable conservation metric for physical state assessment and closed-loop safety monitoring.
[0016] Furthermore, in one possible implementation, the state field includes a structured latent space latent state representation extracted by a physical sensing encoder, maintaining the conservatism constraint of the continuous medium. This structured latent space latent state representation and the variational energy minimization are jointly solved by introducing the conservatism constraint into the latent space energy functional. Alternatively, the energy world model serves as a physical basis model, combined with at least one of a data-driven residual compensation network, a high-frequency kinematic extrapolation module, or a nonlinear stiffness observer, forming a composite forward prediction pipeline in a cascaded, parallel, or residual superposition manner. The technical effect is that while reducing the solution dimension of high-dimensional degrees of freedom, it forcibly maintains the conservatism constraint of the continuous medium, and effectively compensates for unmodeled high-frequency dynamics and contact disturbances using data-driven residuals or observers, improving the accuracy and environmental adaptability of the composite forward inference.
[0017] To address the aforementioned technical problems, this invention also provides a system for closed-loop control of a soft continuum using an energy world model constructed using the above method. The system includes: an actuator for applying displacement constraints, contact pressure, or tensile loads to a soft object; the actuator is at least one of a rigid transmission mechanical structure, a precision displacement guide rail, a continuum mechanism, or a flexible fluid drive unit; an observation device for acquiring physical sensing information characterizing the morphological state of the soft object; the physical sensing information includes at least one of optical point clouds, visual markers, morphological contours, internal strain fiber optic measurements, electromagnetic tracking coordinates, or tomographic image features, and for calculating the displacement and full-field deformation characterization of the controlled observation point; and a model prediction controller for, within a rolling horizon, using the energy world model as an internal prediction model, controlling multiple candidate... The system selects control sequences to extrapolate future states and calculates the cost representing task execution deviations, which include at least one of target tracking error, full-field contour matching degree, or control smoothing penalty. It then selects or aggregates the optimal candidate control sequence and applies the initial control to the actuator. A closed-loop feedback unit feeds back the measured state acquired by the observation device to the energy world model and the controller, forming a prediction-execution-observation-re-prediction closed loop. The controller is deployed on a local industrial control computer, edge gateway, or a remote cloud computing node connected via network communication. The optimal control sequence is sent to the actuator via a downlink communication link. The controlled observation area and the driving area where the actuator directly applies loads have at least decoupled or non-coincident components in spatial configuration or degrees of freedom. The technical effect is that by using an embedded physically conserved world model as a high-fidelity internal prediction model, it overcomes the problem of error divergence contaminating the controller in long-term extrapolation of existing neural world models, achieving high-precision trajectory tracking and mechanical safety control of large deformations in soft continuous media under non-co-located decoupled conditions.
[0018] Furthermore, in one possible implementation, the model predictive controller employs batch tensor quantization to infer the multiple candidate control sequences: the grid states and cell connection topologies of multiple candidate samples are aggregated into batch tensors (of one or a combination of [K,N,3] and [K,M,D,3], where K is the number of samples, N is the number of nodes, M is the number of cells, and D is the number of cells or the neighborhood dimension, where tetrahedral D=4 and hexahedral D=8). Multiple samples are solved simultaneously in a single tensor quantization energy calculation and variational gradient solution (see the batch tensor quantization inference structure). Figure 4 Its technical effects are: significantly improving the parallel utilization rate of the hardware computing unit and the real-time performance of closed-loop control, enabling the continuous medium energy world model to meet the real-time requirements of high-speed online deployment of robots.
[0019] Furthermore, in one possible implementation, the model predictive controller dynamically and adaptively selects the operating point within the solution depth of the internal predictive model (including at least one of the quasi-Newton iteration steps, the number of inner steps, the number of candidate control sequence samples, or the rolling horizon length) based on currently available computing resources or the target control frequency. The technical effect is that it achieves an adaptive dynamic trade-off between physical fidelity and control delay, enabling the control system to switch between high-fidelity accurate extrapolation and high-frequency real-time response.
[0020] Furthermore, in one possible implementation, the material parameters of the software object (including at least one of shear modulus, bulk modulus, and incompressible Poisson's ratio) are pre-calibrated, and the calibration results are written into the near-incompressible hyperelastic strain energy as the mechanical property benchmark of the internal prediction model. The technical effect is that the software physical parameters converge to a modulus basis with clear physical interpretability, possessing high-precision generalization ability even without an excitation trajectory, avoiding the absorption of model errors by fitting parameters without a physical entity, and ensuring the physical identifiability of the calibration parameters and the consistency of simulation migration to the real object.
[0021] To address the aforementioned technical problems, this invention also provides an optimized control method for operating soft continuous media, comprising: obtaining the full-field mechanical state of the controlled object at future moments from an internal physical prediction model within an optimized control solution framework; wherein the full-field mechanical state includes at least one of element volume distortion, contact load reaction force, deformation strain energy, strain tensor invariant, or stress characterization value; applying constraints, penalties, or action corrections based on the full-field mechanical state within the optimized control solution framework to ensure that the control quantity satisfies the internal physical safety boundary of the continuous medium; wherein the optimized control solution framework includes at least one of rolling time-domain optimization, generative trajectory planning based on diffusion denoising, safety filtering of control barrier function, or random sampling optimization. Its technical effect is that by explicitly introducing the internal mechanical field and volume distortion of the continuous medium, which are originally not directly measurable externally, into the upper-level trajectory planning and control optimization closed loop through an internal physical world model, it effectively avoids structural damage, material tearing, and overload instability under extreme contact operation conditions of traditional control, achieving autonomous control that balances operational objectives and the mechanical safety boundary of the object.
[0022] To address the aforementioned technical problems, this invention also provides a physically reliable soft continuous medium operation monitoring system, comprising: an internal physical prediction model for outputting at least one of the following full-field mechanical state fields at the node or unit level during rolling simulations or control: deformation gradient, volume ratio, energy, stress, or contact reaction force; and a monitoring or state extraction unit for driving a tactile force rendering device based on the full-field mechanical state field, outputting damage safety early warning signals, or serving as a safety perception state input for downstream operation planning modules. Its technical advantages are: overcoming the hardware bottleneck that prevents the direct embedding of physical sensors for measuring the deep mechanical state of soft continuous media; and achieving real-time reconstruction and visual monitoring of the internal micro-element level mechanical state using a physically consistent world model, providing reliable mechanical decision support for doctors' tactile perception, industrial overload early warning, and safety defense under high-risk extreme conditions.
[0023] Furthermore, in one possible implementation, the method further includes: acquiring the displacement and reaction force responses of the software through a preset multi-directional excitation trajectory, decoupling and identifying hyperelastic physical parameters mainly composed of shear modulus and bulk modulus; performing open-loop prediction consistency verification on a novel excitation not seen in the multi-directional excitation trajectory, and using the prediction error of the physical parameters on the novel excitation being at least one order of magnitude lower than that of the black-box neural fitting model as the verification criterion at the simulation level; configuring the verified physical parameters into the internal physical prediction model to drive the closed-loop control execution of the software. The technical effect is that: through joint identification of multi-directional dynamic excitation and force sensing, the software material response is converged to the real mechanical modulus with a definite physical entity, eliminating the absorption error of the meaningless free fitting knob from the mechanism, ensuring the global identifiability of material parameters, enabling the world model to have a highly faithful extrapolation generalization ability when facing unseen operation trajectories and complex large deformations, and laying a physical consistency foundation for the migration of simulation to real physical systems.
[0024] Furthermore, in one possible implementation, the method further includes: generating interactive samples containing the full-field mechanical safety state during rolling simulations using an internal physical prediction model; training a control strategy model with a continuous medium physical safety boundary strategy based on the interactive samples, wherein the control strategy model outputs control commands for driving the actuator to operate on the soft continuous medium. The technical effect is that it overcomes the training safety bottleneck of traditional model-less embodied strategies, which are prone to continuous medium tearing and instability during random exploration in real environments. It generates diverse interactive samples containing micro-level stress and volume distortion at low cost and high fidelity using a physical world model, effectively distilling and solidifying the explicit continuous medium physical safety boundary into the neural network control strategy, achieving efficient, robust, and physically hard-constrained end-to-end autonomous control.
[0025] Furthermore, in one possible implementation, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the aforementioned method for constructing and extrapolating the energy balance world model of a continuous medium, the optimization control method for operating a software continuous medium, or the method for identifying and transferring physical parameters of a software continuous medium. The computer-readable storage medium includes, but is not limited to, solid-state drives, hard disk drives, read-only memory, random access memory, optical discs, or flash memory. The technical advantage is that the aforementioned algorithms for constructing the physical world model, solving for variational energy minimization, monitoring the safety of the entire field's mechanics, and identifying and transferring parameters can be independently stored as standardized software modules or executable firmware and flexibly deployed and invoked in industrial control computers, edge gateways, or cloud computing nodes, possessing good engineering reusability and cross-platform portability.
[0026] Furthermore, in order to clearly explain the broad connotation and general extensibility of the various technical features of the present invention, the present invention provides the following superordinate definitions and explanations for the relevant technical terms (each definition is a superordinate concept, and the specific examples are only for illustrative purposes and do not constitute an exhaustive limitation on the scope of protection): (1) The "three-dimensional discrete geometric representation" includes, but is not limited to, at least one of tetrahedral / hexahedral / polyhedral meshes, voxel arrays, particle sets (such as the particle set of the material point method MPM, the smooth particle hydrodynamics SPH), or neural implicit spatial discrete sampling points; (2) The "reference mapping information" includes, but is not limited to, at least one of the inverse matrix of the undeformed gradient matrix of the reference configuration, the reference configuration metric tensor, or the shape kernel function array, used to associate the current local deformation with the undeformed reference; (3) The "volume penalty term" includes, but is not limited to, the square of the difference between the unit volume ratio and the reference value, At least one of the following: logarithmic volume ratio square, higher-order volume deviation polynomial, or fractional volume potential energy, which monotonically increases with the deviation of the unit volume from the incompressible reference; the "modulus parameter or stiffness weight characterizing the incompressible properties of the material" includes, but is not limited to, at least one of the following: bulk modulus, Lamé constant, incompressible penalty weight, or penalty factor; (4) the "numerical solution of variational energy minimization" includes, but is not limited to, at least one of the following: quasi-Newton method (including finite-memory BFGS, i.e., L-BFGS), nonlinear conjugate gradient method, Newton-Raphson method, first-order gradient descent method, or alternating direction multiplier method (ADMM); (5) the "warm start initial value" includes, but is not limited to, at least one of the following: the equilibrium state at the previous moment, or the predicted state obtained by extrapolation based on the equilibrium state at multiple historical moments; the present invention uses the historical equilibrium state as the criterion for the warm start initial value, and ablation experiments show (see the tables and appendices in the specification). Figure 10Using unrelaxed algebraic projections directly as initial values for solving can easily induce numerical reversals, so it is not a preferred implementation of the warm-start initial value in this application; in some alternative implementations, the variational energy minimization can also be achieved by using non-historical priors or cold-start initialization, and by increasing the number of solution steps to achieve equilibrium solution; (6) the “preset physical boundary conditions” include, but are not limited to, at least one of displacement constraints (Dirichlet boundary), external force or surface force load boundaries (Neumann boundary / contact load), elastic support (Robin boundary / elastic foundation), or contact interaction constraints; (7) the “physical sensing information” includes, but is not limited to, at least one of optical point cloud, visual marker pose, morphological contour, internal strain fiber measurement value, electromagnetic tracking coordinates, ultrasonic or tomographic image features; (8) the actuator includes, but is not limited to, rigid transmission mechanical structure (multi-degree-of-freedom robotic arm, At least one of the following: precision displacement guide rail / slide table), continuum mechanism (rope-driven continuum manipulator), flexible fluid drive unit (pneumatic artificial muscle, soft drive cavity), or smart material actuator (such as piezoelectric ceramic actuator, electrostrictive or shape memory alloy actuator); (9) the “decoupling or non-overlapping” of the controlled observation area and the drive area means that there is at least one non-overlapping component in the spatial configuration or degree of freedom, including the non-co-positional form of near-end drive-far-end observation, or the dislocational form of drive and monitoring; (10) the “task execution deviation” or “cost” includes, but is not limited to, at least one of the following: target point tracking error, full field contour matching degree (including chamfer distance or contour overlap), control quantity smoothing penalty, or contact load maintenance deviation; the “cost” may also further include physical safety penalty based on the full field mechanical state quantity (including unit volume distortion, stress concentration or strain energy).
[0027] Furthermore, the energy world model and state field of the present invention can also be presented in a variety of implementable architectures, including: (i) a structured latent space architecture: the state field can be a structured latent space latent state representation extracted by a physical sensing encoder, maintaining the conservation topological constraints of the continuous medium, and jointly solved in the latent space with energy minimization to meet conservation constraints; (ii) a residual composite forward prediction pipeline architecture: the energy world model, as a physical basis model, can be combined with a data-driven residual compensation network, a high-frequency kinematic extrapolation module, or a nonlinear stiffness observer to form a composite forward prediction pipeline in the form of cascade, parallel, or residual superposition; (iii) a generative planning architecture: the optimization control solution framework based on the whole field mechanical state includes rolling time-domain optimization and generative trajectory based on diffusion denoising. (iv) Cloud-edge collaborative deployment architecture: The controller is deployed on a local industrial control computer, edge gateway, or remote cloud computing node connected via network communication, and the optimal control sequence is sent to the actuator via the downlink communication link; (v) Strategy distillation architecture: The internal physical prediction model generates interactive samples containing the full field mechanical safety state in rolling simulation, and trains a control strategy model with a continuous medium physical safety boundary strategy accordingly; (vi) Online identification architecture: The identification of hyperelastic physical parameters can switch between offline multi-directional excitation calibration and online recursive identification (such as extended Kalman filtering, recursive least squares), and the open-loop consistency verification under unseen excitation is used as the reliability verification criterion for model migration. The above implementable architectures can be implemented independently or combined and operated collaboratively in complex industrial or medical scenarios.
[0028] Furthermore, to enable those skilled in the art to more fully implement the multi-architecture extension of the present invention, the present invention further provides the following engineering implementation guidelines and technical details (each implementation guideline and specific parameter is only illustrative and does not constitute a limitation on the scope of protection of the present invention): (1) Structured latent space conservation joint solution: The physical perception observation is encoded as the latent space hidden state z using a graph neural network or encoder-decoder architecture (such as a vision-based embodied joint prediction network or world model architecture), and the isochoric constraint is incorporated into the latent space energy function with the latent space volume ratio deviation penalty term (for example, a conservation penalty positively correlated with J(z)−1 is defined on the latent variable), and the conservation equilibrium hidden state z* is jointly solved using variational or energy minimization, and then the physical state or control quantity is output by the decoder to form a latent space conservation equilibrium world model; (2) Residual composite forward pipeline: The energy world model output p_phys is used as the physical base prediction, and the data-driven residual compensation network learns the residual Δp between the base model and the high-fidelity true value, and the final prediction output is p = p_phys + Δp; During training, the base residual is used as the supervision target, and during inference, the cascade timing or parallel addition of "physical base calculation → residual correction" is performed; (3) Energy-guided diffusion planning: The control trajectory or state sequence is generated by the denoising diffusion model. During the denoising iteration, the energy-guided gradient term (equivalent to the physical classifier-guided safety condition) is constructed using the full field mechanical quantities output by the energy world model (such as the peak value of unit volume distortion max|det(F)−1| or peak energy density). The sampling trajectory is constrained within the physical safety boundary of the continuous medium; (4) Control barrier function safety filtering: After optimizing the control solution outputs candidate actions, the quadratic programming (QP) safety filter is constructed using the control barrier function (CBF) inequality constraint. The control action is projected onto the safety set so that the Lyapunov or Li derivative of the full field mechanical quantities satisfies the barrier condition, realizing the online safety filtering and correction of the action; (5) Strategy distillation: The energy world model is used as the environmental prediction model in the rolling simulation. A large number of interactive samples containing the overall mechanical safety state are generated. The control policy network is trained using reinforcement learning algorithms such as Proximal Policy Optimization (PPO) or Soft Action Evaluation (SAC). The reward function includes a task tracking accuracy term and an overall mechanical safety penalty term (i.e., a heavy penalty is imposed if the overall mechanical quantity exceeds the threshold), so that the policy model is embedded with the physical safety boundary of the continuous medium; (6) Deformation gradient calculation of hexahedral or particle form: construct the Jacobian matrix for the hexahedral mesh with Gaussian integral points of isoparametric unit and obtain the volume ratio and deformation gradient. For the particle set or material point method (MPM), reconstruct the deformation gradient field in the neighborhood with shape function or shape kernel function (such as B spline kernel) and obtain the volume ratio; the above forms are all based on the undeformed reference configuration, and volume conservation is achieved in the optimization solution with the volume deviation penalty term. Its implementation path is consistent with the variational energy minimization framework of the present invention;(7) Online recursive identification: The extended Kalman filter (EKF) or recursive least squares (RLS) algorithm is used to update the parameters of the hyperelastic material as the state vector and the displacement and reaction force as the measurement signals. The parameters are updated frame by frame during closed-loop operation. Offline multi-directional excitation calibration and online recursive identification can be run in parallel or dynamically switched. The open-loop consistency verification under the condition of no excitation is used as the migration and generalization verification criterion before model deployment. (8) Cloud-edge collaborative timing: The controller is deployed on the edge gateway or cloud computing node. The optimal control sequence obtained by the energy world model is sent to the local actuator through the downlink communication link (such as industrial Ethernet, 5G or fieldbus) with a fixed control cycle (such as 100ms to 1s). The actuator executes according to the first step control quantity and sends back the real-time observation status through the uplink communication link to form closed-loop control.
[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) At the architecture level, the embedded volume penalty achieves the continuous medium volume conservation in the optimization solution, and the long-range open-loop autoregressive inference has no cumulative drift and no mesh flipping. Compared with the traditional pure black box neural world model, it has significantly improved the stability and physical confidence of long-range inference; (2) The initial value mechanism of state field temperature start is constructed by using the historical equilibrium configuration, which effectively avoids the numerical oscillation and divergence caused by starting from the unstable candidate initial value in the optimization solution, and ensures the stable convergence of the long-range large deformation autoregressive inference of 50 to 100 steps or more; (3) The rigid displacement clamping is replaced by the consistent Neumann surface force and contact load boundary, which eliminates the mechanical singularity of the local contact boundary, so that the element volume ratio deviation is stably maintained at 10⁻ at the cross-mesh resolution scale. 5Excellent level; (4) Using the energy world model as the closed-loop internal prediction model, the end tracking and configuration control accuracy of non-co-located soft large deformation operation is significantly improved, and the control performance monotonically converges to the theoretical performance limit (ideal physical benchmark) as the internal solution depth increases adaptively; (5) Through the batch tensor parallel solution architecture and the fidelity-delay dynamic adjustable mechanism, the bottleneck of serial evaluation of multiple candidate trajectories is eliminated, and the millisecond-level real-time response and continuous medium micro-element level solution accuracy are taken into account under the strict control cycle; (6) Using the full-field micro-element level mechanical state quantities (including element volume distortion, peak contact reaction force and local strain energy density) evaluated in real time in the rolling simulation by the physical world model, physical safety penalty or dynamic penalty is imposed on the control cost. Displacement clamping enables the closed-loop control system to have both high-precision target tracking and underlying mechanical safety boundary protection; actual measurements show that the peak contact reaction force and local energy density predicted by the traditional black box internal model are amplified by an order of magnitude (about 7.7 times and 16.4 times) compared with the real reference, while the present invention can strictly constrain them within about 10% of the error range of the real reference value; (7) through the multi-directional excitation offline calibration and online recursive identification collaboration of material hyperelastic parameters, the Sim-to-Real physical gap between software continuous medium simulation and real physical system is effectively alleviated, laying the underlying physical consistency foundation for cross-condition migration and hardware system deployment, and improving the migration and generalization potential of control strategy between unseen external excitation and different physical bodies.
[0030] Furthermore, this invention provides a qualitative description of model-free feedback control (such as conventional PID control) as a benchmark: Experimental data shows that while model-free control can achieve basic tracking accuracy (residual approximately 13.9%) in single-target point displacement tracking, it lacks forward physics deduction and prediction capabilities. It cannot pre-judge the full-field element volume distortion, peak contact reaction force, and local strain energy density, thus failing to provide any internal mechanical safety guarantees or proactive avoidance of over-limits in closed-loop control forward planning. In critical operational scenarios involving large soft deformations, near-incompressibility, and high safety requirements, the model-free benchmark can only serve as a basic reference for closed-loop tracking performance. It cannot replace the closed-loop safety control scheme of this invention, which possesses the predictive capabilities of a physical conservation world model, nor can it output the full-field mechanical state of the continuous medium for damage warning and safety monitoring. The above benchmark comparison forms a clear distinction with the safety optimization control and full-field mechanical monitoring scheme of this invention based on the full-field mechanical state, significantly demonstrating the outstanding technical effect of this invention in ensuring the physical safety of the controlled soft object. Attached Figure Description
[0031] Figure 1The diagram shows the hardware and software co-engineering architecture and hardware bus topology of this invention. The labels correspond to the following: 5 is the actuator (slide table / robotic arm driver), 6 is the depth observation camera, 8 is the software object, 9 is the main control computing unit (industrial computer / GPU computing card), 10 is the video memory, 11 is the hardware bus (PCIe / Ethernet / USB3.0 / CAN industrial bus), 12 is the contact probe, and 13 is the fixture base.
[0032] Figure 2 This is a system overview diagram of the forward and non-co-located closed-loop control of the energy balance world model of the present invention, showing the forward inference of the upper-level world model and the lower-level physical execution / observation, as well as the dual closed-loop architecture of "autoregressive inner loop + control outer loop"; the forward backbone is the historical equilibrium state, which is fed back into the variational energy minimization calculator after autoregression and outputs the equilibrium state field; where 2 is the variational energy minimization calculator, 3 is the equilibrium state field, 4 is the upper-level controller / MPC, 5 is the actuator, 8 is the software object, 6 is the observation device, and 7 is the closed-loop feedback.
[0033] Figure 3 The flowchart of variational energy balance time-series deduction and state update of the present invention shows the autoregressive state iteration process in which the previous equilibrium state field 3 is used as the initial value of warm start-up input to the variational energy minimization calculator 2, and is updated to the initial value of the next deduction cycle by the current equilibrium state field 3.
[0034] Figure 4 The diagram shows the structure of the batch tensor quantization derivation of the energy world model of this invention. It shows that the candidate displacement field tensor (with the form [K, N, 3]) and the unit topology / geometric mapping tensor (with the form [K, M, D, 3], where D=4 corresponds to a tetrahedron and D=8 corresponds to a hexahedron or a particle neighborhood) are respectively input into the batch tensor quantization variational energy minimization solver 2, and the solver outputs the data calculation flow of the batch equilibrium state field tensor 3 with the form [K, N, 3] in parallel.
[0035] Figure 5 This is a structural diagram of the non-co-located soft large deformation closed-loop control of the present invention, where each number corresponds to: 4 is the model predictive controller, 5 is the actuator (slide table / robotic arm actuator, whose end contact probe is pressed into the proximal drive area), 6 is the observation device (remote marker observation), 7 is the closed-loop feedback, and 8 is the continuous medium soft object (where 81 is the proximal drive area and 82 is the remote observation area); it shows the non-co-located control characteristics of the drive area and the observation area being decoupled in terms of spatial configuration and degrees of freedom.
[0036] Figure 6The diagram shows two boundary loading methods of the present invention, where (a) is displacement clamping (the flat plate 14 presses down on the entire top surface, and the volume ratio of the contact edge unit is divergent), (b) is Neumann surface force (the probe 12 is locally pressed in, the distributed contact surface force acts on the contact spot 121, the soft object is naturally concave and laterally raised, and the volume deviation remains independent of the mesh), 8 is the soft object, and 13 is the fixture base.
[0037] Figure 7 The flowchart for balancing fidelity and control delay in this invention shows the branch flow for determining computing power adequacy based on available computing resources and target control frequency. When computing power is sufficient, a high-precision deduction mode is adopted, and when computing power is limited, it dynamically switches to a lightweight low-latency mode to generate drive control instructions adapted to timing constraints.
[0038] Figure 8 This is a flowchart of the physical parameter calibration and simulation to real-world transfer process of the present invention. It shows that after obtaining the measured deformation data of the software object 8, the material parameters are inverted and calibrated. Through the dual judgment logic of "parameter convergence and passing the open-loop verification without excitation", if the verification fails, the process returns to re-acquiring data and iterative calibration. After the verification passes, the calibration parameters are written into the hyperelastic near-incompressible physical model to execute the calibration transfer process of the software dynamics deduction.
[0039] Figure 9 This is a schematic diagram showing the role of the full-field mechanical safety constraint in the cost of the model prediction controller in this invention. It shows that the energy world model rolling deduction module 4 outputs the end prediction and full-field mechanical quantities (including volume deviation, peak contact reaction force and local energy density) in parallel, and inputs them into the cost synthesis module containing safety penalty terms, thereby performing a closed-loop judgment process of clamping and over-limit alarm for the drive displacement that exceeds the safety limit.
[0040] Figure 10 The diagram shows a comparison of the ablation and convergence of the core features of the continuous medium world model of this invention. (a) is the ablation curve of the geometric reference datum mapping, showing that the fixed reference configuration, relative to the frame-by-frame incremental reference, reduces the resting configuration volume deviation max|detF−1| by more than an order of magnitude over the entire stroke. (b) is the ablation histogram of the warm-start mechanism, showing that under the same 80-step L-BFGS iteration budget, the advantage of the warm-start mechanism over the cold-start mechanism in the final state volume deviation under 60% to 120% rated stroke load increases significantly with the increase of load. (c) is the volume penalty embedded long-range zero drift curve, showing that under 100-step open-loop autoregressive deduction, the energy world model of this invention stably suppresses the volume deviation to an ultra-low level of less than 10⁻³. Compared with the divergence drift of the black-box graph neural network (GNN) and the numerical flip instability of the subsequent a posteriori projection method, the long-range physical fidelity of this invention is verified.
[0041] Figure 11 The chart shows a comparison of the tracking error and overall mechanical safety of the non-co-located closed-loop model predictive control of this invention. (a) is a histogram comparing the end-tracking error, showing the tracking accuracy advantage of the lightweight real-time mode (123.5%) and high-precision mode (6.2%) of this invention relative to the black-box graph neural network (2325%) and XPBD projection method (617.9%) in logarithmic coordinates, verifying the convergence boundary of this invention under control timing. (b) is a comparison of the multiples of the peak mechanical quantities in the entire field relative to the ideal physical baseline, showing that the black-box graph neural network exceeds the ideal safety baseline by 7.7 times and 16.4 times in peak contact reaction force and local energy density, respectively, while this invention strictly converges and remains near the ideal physical safety baseline (approximately 0.9 times, with a deviation within 10%), verifying the closed-loop control capability of this invention that combines high tracking accuracy with mechanical safety constraints. Detailed Implementation
[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Specifically: Embodiment 1 describes the forward construction and batch tensor quantization parallel derivation of the energy world model, and combines... Figure 1 , Figure 2 , Figure 3 , Figure 4 Example 2 combined Figure 10 Verification of long-range open-loop zero-drift extrapolation and core feature ablation; Example 3 combined with Figure 6 This illustrates the boundary loading implementation of displacement clamping and surface force indentation; Example 4 combines... Figure 5 and Figure 9 Describe the non-co-located closed-loop control topology, full-field mechanical safety constraints, and cost composition; Example 5 combines... Figure 7 and Figure 8 Verification of computing power scheduling, planner robustness, and physical parameter inversion calibration migration; Example 6 presents specific control applications in medical and industrial vertical scenarios; Example 7 presents a real-device hardware calibration system and its integration with... Figure 11 A comparison of tracking accuracy and overall mechanical safety landing.
[0043] The system of this invention is implemented using a hardware and software co-operating physical computing device. For example... Figure 1As shown, the physical entity consists of: a main control computing unit 9 (including an industrial computer or GPU computing card, and configured with resident video memory 10), an actuator 5 such as a high-rigidity ball screw slide drive module or robotic arm joint controller for driving the contact probe 12 to press in or displace and clamp, a point cloud / stereo industrial-grade depth observation camera 6 (structured light or stereo binocular camera) for acquiring deformation images, and a specimen clamp base 13 for fixing the software object 8 to be operated; the main control computing unit 9, the actuator 5 and the industrial-grade depth observation camera 6 are interconnected at millisecond level and have low-latency data interaction via a hardware bus 11 (including at least one of PCIe, Ethernet bus, USB 3.0 bus or CAN industrial bus).
[0044] Regarding data residency and closed-loop memory operation: Discrete surface point cloud data acquired by the industrial-grade depth observation camera 6 is directly written to the onboard memory 10 of the main control computing unit 9 via direct memory access (DMA) through the hardware bus 11 with a zero-copy mechanism, bypassing the CPU; The candidate state field tensor (with the form [K, N, 3], generated from historical equilibrium states) and the unit topology / geometric mapping tensor (with the form [K, M, D, 3], D=4 for tetrahedrons) reside in the tensor register in their entirety, and the parallel energy minimization relaxation solution of the GPU batch tensor variational energy minimization solver 2 is performed in situ; The optimal control sequence output by the solver is converted into hardware control pulses or speed commands via the hardware bus 11 and sent to the actuator 5 (slide table / robotic arm driver) to drive the end contact probe 12 to perform displacement pressing; The closed-loop feedback pose data obtained from the current deduction and observation is then transmitted back to the controller via the closed-loop feedback module 7 to complete the timing rolling closed loop.
[0045] Its technical effect is to realize a closed-loop software and hardware collaborative computing chain of "point cloud acquisition - in-situ variational energy relaxation of video memory - hardware control pulse generation", eliminate the bottleneck of frequent main memory / video memory data copying between CPU and GPU, and ensure ultra-low latency high-frequency real-time control and dynamic safety response at the level of milliseconds to tens of milliseconds from the physical device level.
[0046] Example 1: Forward Construction and Derivation of the Energy Balance World Model (also known as the Embedded Energy World Model, codenamed IsoLock in academic and engineering implementation). Combined with... Figure 1 , Figure 2 , Figure 3 and Figure 4A tetrahedral mesh (approximately 1287 nodes and 4800 elements) is created for a soft tissue block (geometric dimensions 0.16m × 0.12m × 0.08m, Poisson's ratio 0.4999, shear modulus μ 1.0kPa) used as an example of soft body object 8. A reference configuration is determined and the inverse of its undeformed gradient matrix is pre-calculated. Based on the inverse matrix of the reference configuration, the deformation gradient F = A·A0⁻¹ of each tetrahedral element is calculated, and a near-incompressible hyperelastic strain energy function is constructed (where the isochoric energy term is 0.5μ·(I1−3), the volume penalty term is 0.5λ·(J−1)², J = detF, and the bulk modulus λ can be taken as 10). 5 Up to 10 7 Under fixed boundary constraints, using the equilibrium state of the previous extrapolation period as the initial warm-start value, the energy minimization solution is performed using the L-BFGS quasi-Newton method (combining a strong Wolfe line search criterion and a preconditioner with row scaling) with batch tensor variational energy minimization solver 2. The equilibrium displacement of each free node is calculated to obtain the equilibrium state field that maintains volume conservation, so as to construct the energy world model. The measured volume deviation (the maximum value of the difference between the element volume ratio and 1, max|detF−1|) is compared as shown in Table 1 below:
[0047] Table 1: Relationship between volume conservation and bulk modulus (The median of the mesh deformation test results is taken under low load reference conditions; for high stroke test results, please refer to the ablation comparison below and Tables 2 and 3).
[0048]
[0049] like Figure 2 (System Overview) and Figure 3 As shown in the (warm-start timing loop), the warm-start energy minimization uses the preceding equilibrium state field 3 from the previous simulation cycle as the initial value for warm-start. Module by module, it performs inverse calculation of the hyperelastic strain energy based on the reference configuration, and under fixed boundary constraints, the variational energy minimization solver 2 performs quasi-Newton minimization. Finally, it outputs the current equilibrium state field 3, and then updates this output with the initial value for warm-start in the next simulation cycle. The above experimental results show that, under low-load reference conditions, volume conservation is achieved by the energy penalty term at the optimization level with minimal deviation (approximately 10⁻⁻⁶). 5 The magnitude of the deviation remains strictly constant and does not change significantly with the change of grid resolution (such as the element feature size in the range of 5 to 12), which means it has excellent grid independence. When the load amplitude increases to the high stroke, the volume deviation increases accordingly. The specific magnitude of the change is shown in the following ablation control experiment and Tables 2 and 3.
[0050] Furthermore, to demonstrate the indispensability of each key technical feature in the technical solution of this invention, under the same near-incompressible hyperelastic strain energy framework (shear modulus of 1.0 kPa, bulk modulus λ = 10),5 Pa, fixed boundary constraints, and uniformly using the L-BFGS solver based on the strong Wolfe line search criterion) set the following ablation comparison for five test soft tissue samples (sample 1 to sample 5, each containing a tetrahedral mesh of approximately 1287 nodes and 4800 elements): (1) Geometric reference datum mapping ablation (removing the inverse of the gradient matrix of the fixed reference configuration): If the deformation gradient F=A·A0⁻¹ is not calculated based on the undeformed reference configuration, but is approximated by the previous balanced configuration frame by frame as the local deformation degree (i.e., incremental reference), then the volume conservation is only The reference frame, which is drifted frame by frame, cannot be aligned with the true resting configuration. In actual measurements, when five test samples were monotonically pushed to 30%, 60%, and 90% of the rated stroke, the mean volume deviation max|detF−1| of the resting configuration was magnified by approximately 4.5 times, 5.1 times, and 4.9 times, respectively, compared to the fixed reference configuration (the fixed reference configuration remained at 5.2×10⁻³, 1.05×10⁻², and 1.91×10⁻², while the incremental reference deteriorated sharply to 2.35×10⁻², 5.42×10⁻², and 9.32×10⁻²). The above ablation confirms that using a fixed geometric reference is a necessary technical feature for strictly anchoring the volume to the true resting configuration and avoiding geometric reference state drift. (2) Warm-start mechanism ablation (cold-start control): If the equilibrium state of the previous moment is not used as the initial value for warm-start at each deduction moment, but the cold-start solution is always performed with the undeformed resting configuration, at a bulk modulus λ=10 5 Under the high ill-conditioned potential surface of Pa, with the same 80-step L-BFGS iterative budget, the mean values of the final-state volume deviation max|detF−1| for five samples at 60%, 90%, and 120% of the rated stroke were approximately 1.29 times, 1.59 times, and 2.15 times higher than those at warm start (2.3×10⁻², 4.7×10⁻², and 9.5×10⁻² for warm start, and 3.0×10⁻², 7.5×10⁻², and 2.0×10⁻¹ for cold start), and the convergence gap increased significantly with the increase of the load amplitude. This ablation confirms the necessary technical feature of warm start, which uses the preceding equilibrium state as the initial value, enabling faster convergence to the physical equilibrium point with a predetermined iterative budget on a near-incompressible high ill-conditioned potential surface, and avoiding deterioration of the extrapolation accuracy. Furthermore, even if the implementer does not fully adopt the specific preferred form of this invention for the aforementioned geometric reference datum or initial temperature start value, the optimized control method and physically reliable monitoring system proposed in this invention independently cover the software operation technology solution that includes full-field mechanical safety control and physical state monitoring, together with the aforementioned features, constituting a multi-layered patent protection and defense system. The two sets of ablation experiments above jointly demonstrate that geometric reference datum mapping and initial temperature start value respectively make significant technical contributions to ensuring the physical consistency and convergence stability of continuous medium solutions. For specific ablation mechanisms and failure comparisons, please refer to the appendix. Figure 10 .
[0051] Example 2: Long-range open-loop zero-drift simulation. For the 100-step open-loop autoregressive simulation process of probe monotonic slope displacement loading under fixed boundary constraints, the measured volume deviation comparison results are shown in Table 2 and appendix below. Figure 10 As shown in (c):
[0052] Table 2: Comparison of 100-step open-loop autoregressive volume drift
[0053]
[0054] As shown in Table 2, the volume deviation of the black-box neural network drifted to 1.17 × 10², while the subsequent algebraic projection method resulted in grid numerical flipping (volume deviation reaching 3.68 × 10²). 5 The present invention maintains the volume deviation at a stable 8.7 × 10⁻ 4 (The average value of the last 20 steps is 8.2 × 10⁻) 4 The measured volume deviation was strictly less than the preset safety threshold of 1×10⁻³, thus achieving excellent long-range zero drift while avoiding mesh cell flipping. Furthermore, to demonstrate the universality and robustness of long-range zero-drift inference in cross-data domain tasks, this invention performed 400 steps of open-loop autoregressive inference on the publicly available benchmark dataset (deforming_plate, derived from publicly published soft plate deformation finite element benchmark data, see reference: Pfaff et al., Learning Mesh-Based Simulation with Graph Networks, ICLR 2021) with a partitioned tetrahedral mesh. The measured volume deviation was strictly locked at approximately 1.8×10⁻³ under different random seed initializations. 5 In contrast, if only a pure black-box neural network model without energy variational constraints is used for forward inference, the cell volume ratio rapidly diverges and flips to the order of 10³³ or even overflows (+∞). This comparative result, based on independently published data, confirms that the long-range zero-drift and anti-flipping capabilities of this invention are not specifically dependent on a particular geometric mesh or proprietary working condition, significantly improving the cross-domain generalization and control robustness of the technical solution of this invention under different soft body morphologies and large deformation tasks.
[0055] Example 3: Verification of the independence of contact boundary loading method from mesh. (Combined with Appendix) Figure 6As shown, when the flat pressure plate 14 is used to apply full top-surface displacement clamping (Dirichlet boundary) loading to the soft object 8, the stress concentration in the singularity zone at the contact edge leads to an abnormal amplification of the element volume ratio, causing the volume deviation max|detF−1| to increase significantly with mesh subdivision. After using the contact probe 12 for local pressing and applying distributed surface force (Neumann boundary condition) loading to the contact patch 121, the mesh edge singularity is effectively eliminated, and the volume deviation remains stable and converges across different mesh resolutions, exhibiting excellent mesh independence. The actual measurement comparison results are shown in Table 3 and the appendix below. Figure 6 As shown:
[0056] Table 3: Comparison of Volume Deviation with Mesh Resolution under Different Contact Boundary Loading Methods
[0057]
[0058] like Figure 6 As shown, Figure 6 (a) shows the use of a flat pressure plate 14 to apply overall rigid displacement clamping and downward pressure to the top surface of the soft object 8. Figure 6 (b) shows the application of Neumann distributed surface force contact load to local contact patch 121 using contact probe 12. The measured results in Table 3 show that displacement clamping loading causes the volume deviation max|detF−1| in the singular region of the contact edge to increase sharply with mesh subdivision (from 6.3×10⁻² at 1521 nodes to 5.7×10⁻¹ at 10625 nodes), while Neumann surface force loading keeps the volume deviation stable at approximately 1.8×10⁻¹ across all mesh resolutions. 5 Up to 2.0×10⁻ 5 At extremely low levels, it completely eliminates mesh edge singularities and achieves strictly mesh-independent physical volume conservation.
[0059] Example 4: Non-co-located closed-loop model predictive control. Combined with... Figure 4 , Figure 5 , Figure 7 , Figure 9 and Figure 11 As shown: Using the aforementioned energy world model as the internal derivation core of the model predictive controller 4, the cross-entropy method (CEM) is employed to execute non-co-located model predictive control, with the rolling horizon H=6 and the candidate control trajectory sampling number K=64 (corresponding to...). Figure 4The first dimension of batch tensor quantization), the comprehensive cost is defined as the sum of the end-tracking error, the control smoothing penalty (weighted at 0.5), and the full-field mechanical safety over-limit penalty. To provide an objective benchmark, the closed-loop verification uses a fully convergent high-fidelity energy equilibrium solution as the reference physical truth (minimizing the same near-incompressible hyperelastic strain energy to equilibrium at the free node, which is physically equivalent to the steady-state equilibrium solution obtained by the implicit finite element method), and examines the closed-loop control performance when using a black-box graph neural network and the energy world model of this invention as the internal prediction core, respectively. Under 5 sets of independent random initialization control sequences, the measured end-tracking error (percentage of the target reference displacement) is compared as shown in Table 4 below. Figure 11 As shown in (a):
[0060] Table 4: Comparison of Predictive Control Performance of Non-Colocative Closed-Loop Models (Terminal error is expressed as a percentage of reference displacement under 5 independent random initialization tests)
[0061]
[0062] Furthermore, as the single-step optimization iteration budget (L-BFGS iteration count) of the variational energy minimization solver 2 increases, the end-point tracking error of the control loop of this invention exhibits a strictly monotonically decreasing convergence trend. The actual convergence comparison results are shown in Table 5 below:
[0063] Table 5: Internal optimization iteration steps and accuracy-delay trade-off (based on 5 independent randomized trials)
[0064]
[0065] like Figure 5 As shown, in the non-co-located closed-loop control topology, the model predictive controller 4 uses the energy world model as the core of internal deduction. Control commands are issued to the actuator 5 to actively drive the near-end drive area 81 of the software object 8. Meanwhile, the observation device 6 performs displacement and deformation calculations on the far-end observation area 82, which is decoupled from the near end in terms of spatial configuration and degrees of freedom. The calculation data is fed back to the model predictive controller 4 through the closed-loop feedback module 7, thus constructing a time-series rolling closed loop of "deduction-execution-observation-re-deduction". Regarding closed-loop tracking performance, the measured data in Table 4 shows that the black-box internal model suffers from a tracking error as high as approximately 2325.0% due to numerical divergence, while the relaxed geometric projection method (XPBD) has an error of 617.9%, the model-free PID baseline error is 13.9%, and the open-loop nominal replay error is 5.1%. In contrast, the lightweight real-time mode of this invention has an error of only 123.5±61.1%, and the high-precision mode error is further converged to 6.2±2.2%, significantly outperforming the black-box neural network model and effectively overcoming the physical distortion defects of the geometric projection method (see comparison of tracking accuracy and mechanical safety). Figure 11 ).like Figure 7As shown, the controller dynamically and adaptively switches between "high-precision simulation mode" and "lightweight low-latency mode" based on the available computing power. Furthermore, as the single-step iteration budget of the variational energy solver increases, the terminal error exhibits a strictly monotonically decreasing trend (see Table 5). In addition, during the rolling simulation cycle, the model prediction controller 4 uses the embedded energy world model to evaluate the full-field volume deviation max|detF−1|, peak contact reaction force, and local energy density in real time, and outputs the node / cell-level full-field mechanical state field. Figure 9 The above-mentioned full-field mechanical safety constraints are shown in the decision flow of the model predictive controller in the synthesis of comprehensive costs and the clamping of over-limit displacement, so that the non-co-position closed-loop control has both high tracking accuracy and full-field physical safety boundary.
[0066] Example 5: Planner Robustness and Material Parameter Calibration. The model predictive planner was set with a rolling horizon H=8, candidate sampling trajectory number K=128, and bulk modulus at λ=10. 4 Pa or 10 6 Under perturbation conditions such as Pa, the measured high-precision inference model's terminal error remained stable in the range of 6.2% to 10.7%, and the lightweight low-latency model remained in the range of 40.6% to 193.3%, both far superior to the divergence error of the black-box model (approximately 2325%), confirming the strong robustness of the method and system of this invention under different planning horizons and model mismatch perturbations; combined with Figure 8 The illustrated identifiable physical material calibration process involves probing the surface with an excitation probe and observing the soft surface response with a depth camera. The energy world model is used to extrapolate the state, and differentiable physical backpropagation is performed based on the observation residuals. The material parameters (including shear modulus μ, bulk modulus λ, and near-incompressible Poisson's ratio) obtained through inversion and updating are written into the near-incompressible hyperelastic strain energy function. This achieves a high degree of self-consistency between the extrapolated predictions and the measured calibration results, providing physical consistency assurance for the transition from simulation to real-world operation. The mechanical parameters identified and converged during the calibration process, primarily shear modulus and bulk modulus, possess clear physical dimensions. Furthermore, open-loop prediction consistency verification can be performed using a novel external excitation trajectory that was not involved in the calibration (at the simulation level, the extrapolation prediction error under external test conditions is at least one order of magnitude lower than that of the black-box model as a verification criterion). This effectively avoids the problem of excessive absorption of modeling errors by using empirically fitted free parameters lacking physical entity meaning, ensuring the true identifiability and task generalization of the material physical parameter calibration.
[0067] Example 6: System Deployment and Engineering Interaction Process in Minimally Invasive Medical Surgery and Flexible Industrial Assembly Scenarios. This example clarifies the specific hardware and software deployment and timing interaction control process of the continuous medium energy balance world model and non-co-located closed-loop control architecture of the present invention in practical engineering scenarios such as medical surgical assistive robots and flexible industrial manufacturing assembly.
[0068] (a) Scenario of soft tissue collaborative traction and safety control of medical minimally invasive surgery assist robot: The actuator 5 adopts the end of the minimally invasive surgical instrument 12 with force sensing function (e.g., minimally invasive needle holder or tissue forceps); the observation device 6 adopts a body cavity stereoscopic binocular endoscope; the soft object 8 is a biomechanical biomimetic soft tissue model of the surgical field or an ex vivo soft tissue sample (e.g., an ex vivo liver tissue module, a simulated stomach wall or an ex vivo blood vessel bundle). The stereoscopic binocular endoscope acquires the three-dimensional surface point cloud of the soft object 8 in real time and extracts spatial marker points. The energy world model calculates the energy state of large deformation inside the soft tissue in real time based on the preset anatomical voxel-tetrahedral mesh. When the master control end or autonomous planning system drives the end of the minimally invasive instrument 12 to act on the edge of the proximal driving area 81 of the soft object 8, the model prediction controller 4 uses the embedded hyperelastic variational energy balance layer to deduce the full-field deformation gradient and strain energy spatial transmission from the proximal driving displacement to the distal observation area 82 (such as the attachment point of deep sensitive blood vessels) in real time. When the predicted local tensile strain energy at the distal observation area 82 reaches or exceeds the preset safety damage threshold, the model prediction controller 4 triggers the displacement clamping mechanism, actively cuts off the incremental driving displacement in the dangerous direction and outputs a force tactile / audio-visual safety warning signal, thereby realizing the autonomous and collaborative constraint control of the instrument end trajectory with both physical realism deduction and critical mechanical safety boundary protection.
[0069] (b) High-precision assembly scenario of flexible sealing strips and wire harnesses in industrial automation: The actuator 5 uses a roller pressing probe 12 configured at the end of a high-rigidity industrial robot; the observation device 6 uses an industrial 3D structured light camera mounted above the production line; the soft object 8 is an EPDM rubber sealing strip for automotive doors or an industrial power wire harness with near-incompressible characteristics. When the sealing strip is pressed into the narrow metal groove of the door for assembly, the 3D structured light camera acquires the 3D contour point cloud of the unpressed section of the strip in real time; the rolling operation of the roller pressing probe 12 constitutes a typical non-co-position control (i.e., the proximal driving area 81 where the probe 12 applies force is decoupled from the far observation area 82 where the deformation quality and flatness of the far embedded groove section are located in the degree of freedom); the model predictive controller 4 will... Figure 8 The physical parameters of the EPDM rubber material calibrated offline are written into the near-incompressible hyperelastic strain energy function. The bulging deformation of the rubber strip section under large curvature bending is simulated in real time within the rolling time domain. The optimal pose control trajectory of the output probe 12 is optimized by the cross-entropy method (CEM) to effectively prevent local extrusion stacking and groove defects, and ensure that the sealing rubber strip section is geometrically uniformly embedded in the groove.
[0070] Example 7: Physical System Calibration and Full-Field Mechanical Safety Control Implementation. To seamlessly transfer the technical effects achieved in the simulation to the real physical environment, this example combines... Figure 1 , Figure 8 and Figure 9 A complete implementation method for a physical simulation system is presented, and the engineering implementation process for full-field mechanical safety constraints and real-time monitoring of physical state is specifically described in this simulation system.
[0071] (1) Hardware configuration and sensing configuration: A high-rigidity precision linear / right-angle sliding platform or industrial robotic arm is used as the actuator 5, and a contact probe 12 (e.g., a straight-head contact probe) is installed at its end. The soft object 8 (the soft object 8 is mounted on the clamping base 13) is driven by a position servo kinematic closed loop. As another preferred embodiment compared to the distributed surface force loading in Embodiment 3, this embodiment can also achieve controlled local contact by pressing the contact probe 12 in place through position servo driving. The observation device 6 uses a monocular camera or a depth camera 6, combined with multiple visual feature markers (e.g., ArUco marker blocks) attached to the outer surface of the soft object 8, to calculate the spatial displacement field between the end of the contact probe 12 and the local contact area in real time. In addition, a multi-dimensional force / torque sensor is connected in series between the contact probe 12 and the actuator 5 to collect contact reaction force data in real time.
[0072] (2) Calibration and transfer verification of physical parameters of the actual machine: combined with Figure 8 The process described above involves controlling the contact probe 12 to apply excitation along multiple trajectories, including orthogonal axial and oblique indentation (simultaneously acquiring full-field displacement and contact reaction force response data under different indentation directions and loading rates). Utilizing the differentiable backpropagation of the energy world model, the constitutive parameters of the soft object 8 are converged and optimized to have shear modulus μ and bulk modulus λ with real physical dimensions (introducing a multi-dimensional force sensor to achieve joint force-displacement inversion identification, effectively avoiding underdetermined degradation of mechanical absolute stiffness and reaction force amplitude parameters when relying solely on surface optical displacement inversion). Subsequently, open-loop autoregressive rolling prediction consistency verification is performed on a novel external test excitation trajectory that has not participated in calibration fitting. This completes open-loop consistency verification of the identified parameters at the simulation and virtual transfer levels (verifying that the prediction error is at least one order of magnitude lower than the black-box model). The aim is to reproduce the above verification process in physical real-machine transfer, providing a physical consistency basis for real-machine deployment. This lays an actionable framework for importing the identified and calibrated material physical parameters into the near-incompressible hyperelastic strain energy function and achieving faithful transfer for physical real-machine objects.
[0073] (3) Full-field mechanical safety monitoring, over-limit displacement restraint, and strategy distillation training: In the target deployment closed loop for real physical systems, the embedded energy world model can output the full-field mechanical state field, such as the tetrahedral element volume ratio, isochoric distortion energy, volume penalty energy, peak contact reaction force, and local energy density, at each prediction and deduction step; the model prediction controller 4 applies a nonlinear safety penalty term (e.g., as the full-field mechanical state quantity exceeds the safety limit) monotonically increasing as the full-field mechanical state quantity exceeds the safety limit in the comprehensive cost calculation. Figure 9 (As shown), or perform real-time displacement clamping and interception on the drive displacement sent to actuator 5 to meet the preset safety hard boundary. Simulation comparison test shows (see Figure 11 (b) If driven solely by a black-box internal model without physical constraints, the peak contact reaction force and local strain energy density are amplified by approximately one order of magnitude compared to the high-fidelity finite element physical reference true value model (by 7.7 times and 16.4 times, respectively), which can easily lead to soft body tearing or excessive compression instability. The energy world model of this invention, however, strictly maintains the peak contact reaction force and energy density within 10% of the deviation from the reference true value model. At the simulation verification and transfer protocol level for real physical systems, it verifies its closed-loop control capability, which combines high tracking accuracy with inherent mechanical safety. Furthermore, the large-scale physical interaction sample set containing the aforementioned full-field mechanical safety states, generated in parallel based on the rolling deduction of the energy world model, can be directly used for offline training of reinforcement learning or imitation learning policy networks with embedded physical safety boundaries. This distills and solidifies the implicit physical deduction and safety monitoring capabilities of the world model into high-frequency, real-time, end-to-end low-latency control commands, driving the actuator 5 to perform safe and efficient operations on the continuous medium soft object.
[0074] General Parameter Description and Implementation Extension: The specific values recorded in the above embodiments (including the number of grid nodes, number of cells, bulk modulus, rolling horizon, number of sampling trajectories, iteration step budget, safety threshold, etc.) are all exemplary preferred implementations. Those skilled in the art can flexibly adjust them according to the actual work object and available computing resources. The energy world model and closed-loop control functional modules of the present invention can be deployed on general computing devices (such as graphics processing units (GPUs), tensor processors (TPUs), neural network processors (NPUs), or central processing units (CPUs) for execution, or can be embedded in dedicated hardware chips such as field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), or system-on-a-chip (SoCs). The actuator for its closed-loop control can be a multi-degree-of-freedom robotic arm or a high-stiffness precision slide, and the observation device can be a depth camera, a structured light camera, or a binocular stereo vision camera. The above calculation programs, data structures, and control algorithms are embedded, stored, and run in a computer-readable storage medium contained in a local industrial control computer, edge controller, robot onboard computing unit, or remote cloud server, and are called and executed by one or more processors configured within it to realize the above energy deduction, parameter identification, and closed-loop safety control process.
[0075] The above description is merely a specific embodiment of the present invention and does not constitute a limitation on the scope of protection of the claims of the present invention. Any conventional modifications, equivalent substitutions, or partial engineering improvements made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be included within the scope of protection of the claims of the present invention.
Claims
1. A method for constructing and extrapolating a world model of energy balance in a continuous medium, characterized in that, include: Obtain a three-dimensional discrete geometric representation of the target software object, wherein the three-dimensional discrete geometric representation includes at least one of a polyhedral mesh, a voxel array, or a particle set; Determine the geometric reference datum for the geometric representation, and obtain reference mapping information characterizing local deformation based on the geometric reference datum; Based on the reference mapping information, the deformation gradient is calculated for each discrete element of the geometric representation, and the near-incompressible hyperelastic strain energy is calculated accordingly. The near-incompressible hyperelastic strain energy includes an isochoric energy term and a volume penalty term. The volume penalty term increases monotonically as the element volume deviates from the incompressible reference and is weighted by the modulus parameter or stiffness weight that characterizes the incompressible properties of the material, so that the volume conservation is achieved by the energy term embedded in the optimization solution process. Under the preset physical boundary conditions, during the reasoning and deduction process, the variational energy minimization numerical solution is performed on the near-incompressible hyperelastic strain energy to solve for the equilibrium displacement of the unconstrained physical degrees of freedom, thereby obtaining the equilibrium state field that maintains volume conservation and constructing the energy world model. The results characterizing the equilibrium state are output for autoregressive deduction at subsequent times, or the energy world model is used as the internal prediction model of the upper-level controller.
2. The method according to claim 1, characterized in that, The step of performing variational energy minimization numerical solution during inference includes: using the historical equilibrium state before the current inference cycle, or the predicted state obtained by extrapolation based on the equilibrium state at multiple historical moments, as the warm-start initial value, and inputting the warm-start initial value into the variational energy minimization solver for optimization and iterative solution.
3. The method according to claim 1, characterized in that, The extrapolation is a long-range open-loop extrapolation. During the extrapolation process, the deviation of the unit volume ratio does not exceed a preset threshold as a continuous criterion. When the criterion is met, the autoregressive extrapolation continues, so that the volume deviation remains below the preset threshold and no mesh flipping occurs.
4. The method according to claim 1, characterized in that, The three-dimensional discrete geometry is represented as a tetrahedral mesh, the geometric reference datum is an undeformed reference configuration, and the reference mapping information is the inverse of the undeformed gradient matrix of the undeformed reference configuration. The deformation gradient is calculated in the form F = A·A0⁻¹ (where A is the gradient matrix of the current configuration and A0 is the gradient matrix of the undeformed reference configuration). When a discrete element is at risk of criticality or flipping, a lower bound clamp is applied to the element volume ratio or the state of the corresponding element is reinitialized. The variational energy minimization adopts at least one of the following: finite memory BFGS quasi-Newton method (i.e., L-BFGS), nonlinear conjugate gradient method, or Newton-Raphson method. During the solution process, the energy curvature difference in different degrees of freedom directions is normalized or preprocessed (including row scaling). The preset physical boundary conditions include applying Neumann surface force or contact load (including external potential energy term) to the contact or loading region instead of applying displacement clamp to the complete surface, so as to eliminate the boundary singularity introduced by displacement clamp at the contact edge.
5. The method according to claim 1, characterized in that, The energy world model serves as the physical basis model and is combined with at least one of the following: a data-driven residual compensation network, a high-frequency kinematic extrapolation module, or a nonlinear stiffness observer, to form a composite forward prediction pipeline in the form of cascade, parallel, or residual superposition.
6. The method according to claim 1, characterized in that, Also includes: The displacement and reaction force response of the software are collected by pre-set multi-directional excitation trajectory, and the hyperelastic physical parameters, mainly shear modulus and bulk modulus, are decoupled and identified. An open-loop prediction consistency check is performed on a novel stimulus not seen in the multi-directional stimulus trajectory, so that the prediction error of the physical parameters on the novel stimulus is at least one order of magnitude lower than that of the black-box neural fitting model; the physical parameters that pass the check are configured into the internal physical prediction model to drive the closed-loop control execution of the software.
7. A system for closed-loop control of a soft continuum using an energy world model constructed by the method of claim 1, characterized in that, include: An actuator is used to apply displacement constraints, contact pressing, or tensile loads to a software object. The observation device is used to acquire physical sensing information characterizing the morphological state of the soft body, and to calculate the displacement and full-field deformation characterization of the controlled observation point. The model predictive controller is used to extrapolate the future state of multiple candidate control sequences within the rolling horizon, using the energy world model as the internal predictive model, and to calculate the cost representing the task execution deviation, select or aggregate the optimal candidate control sequence, and apply the first-step control to the actuator; the closed-loop feedback unit is used to feed back the measured state acquired by the observation device to the energy world model and the controller, forming a prediction-execution-observation-repreneurial closed loop; wherein, the controlled observation area and the driving area where the actuator directly applies loads have at least decoupled or non-coincident components in spatial configuration or degrees of freedom.
8. The system according to claim 7, characterized in that, The model predictive controller employs batch tensor quantization for the inference of the multiple candidate control sequences: the grid states and cell connection topologies of multiple candidate samples are aggregated into batch tensors (of the form [K,N,3] and [K,M,D,3] or a combination thereof, where K is the number of samples, N is the number of nodes, M is the number of cells, and D is the number of cell nodes or the neighborhood dimension (tetrahedral D=4, hexahedral D=8)). Multiple samples are solved simultaneously in a single tensor energy calculation and variational gradient solution. The model predictive controller dynamically and adaptively selects the operating point based on available computing power or the target control frequency, within the solution depth of the internal predictive model (including at least one of the following: quasi-Newton iteration steps, number of inner steps, number of candidate control sequence samples, or rolling horizon length), to balance fidelity and latency.
9. The system according to claim 7, characterized in that, The material parameters of the soft object (including at least one of shear modulus, bulk modulus and incompressible Poisson's ratio) are pre-calibrated, and the calibration results are written into the near-incompressible hyperelastic strain energy as the mechanical property reference of the internal prediction model.
10. An optimized control method for soft continuous media operation, characterized in that, include: In the optimization control solution framework, the full-field mechanical state of the controlled object at future moments is obtained by an internal physical prediction model. The full-field mechanical state includes at least one of element volume distortion, contact load reaction force, deformation strain energy, strain tensor invariant, or stress characterization value. In the optimization control solution framework, constraints, penalties, or action corrections are applied based on the full-field mechanical state to ensure that the control quantity satisfies the physical safety boundary inside the continuous medium. The optimization control solution framework includes at least one of rolling time-domain optimization, generative trajectory planning based on diffusion denoising, control barrier function safety filtering, or random sampling optimization.
11. A software continuous media operation monitoring system, characterized in that, include: An internal physical prediction model is used to output at least one of the following full-field mechanical state fields at the node or unit level: deformation gradient, volume ratio, energy, stress, or contact reaction force during rolling simulation or control; and a monitoring or state extraction unit is used to drive the haptic force rendering device based on the full-field mechanical state field, output damage safety early warning signals, or serve as a safety perception state input for downstream operation planning modules.
12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6 and 10.
