An ai-based simulation human cinema line creative workflow method

Abstract

This invention discloses an AI-based simulated human creative workflow method for cinemas. It includes: semantic parsing of multimodal intent data to generate a target functional and a set of physical boundary parameters; solving a system of multi-rigid-body nonlinear differential equations to obtain a sequence of motion trajectory states; performing forward kinematic mapping and rasterization rendering to output a scene atlas tensor; calculating visual loss and calculating the original visual gradient using inverse ordinary differential equations; obtaining Lyapunov energy functional indices to construct energy dissipation boundary manifold constraints; performing orthogonal projection operations on the original visual gradient to eliminate gradient components that cause the total system energy to exhibit non-physical divergence, outputting a safe physical gradient; and updating the physical boundary parameter set based on the safe physical gradient in a closed loop. This invention prevents parameter iteration from deviating from the dynamically stable state and avoids system divergence and collapse.

Description

Technical Field

[0001] This invention relates to the field of computer graphics technology, specifically to an AI-based method for simulating human creative workflows in cinemas. Background Technology

[0002] In the film and television production process, the creation of humanoid characters primarily involves two methods: motion capture-based computer graphics and neural network-based generative artificial intelligence. Motion capture-based computer graphics captures performance data from live actors and maps it to a humanoid skeletal model, supplemented by post-production human correction and offline rendering to maintain the physical accuracy of the visual effects. Generative artificial intelligence methods, on the other hand, parse the script using a multimodal large model, directly outputting end-to-end humanoid performance videos or 3D mesh sequences to improve content generation efficiency.

[0003] Existing motion capture-based production methods are unidirectional processes, lacking a mathematical connection between natural language intent and the final physical state representation. When semantic-level fine-tuning of simulated human movements is required, the system cannot perform automated inverse state solving, leading to cumbersome modification procedures. Existing generative AI methods do not integrate multi-rigid-body dynamics differential equations and environmental collision constraints at the underlying level, resulting in motion sequences that do not follow Lagrange's principles, leading to issues such as center of gravity shift, momentum non-conservation, and geometric penetration in the generated results. A common approach to addressing these issues is to backpropagate visual semantic loss to the physics engine through a differentiable rendering pipeline. However, in actual closed-loop iterations, directly updating physical boundary parameters with purely visual-guided mathematical gradients disrupts the mechanical energy conservation of the dynamic system, causing parameter evolution to deviate from physical boundaries and ultimately leading to divergence in the physics engine's state-space equations. Existing technologies cannot simultaneously parse multimodal semantic instructions and execute automated inverse control while ensuring that the closed-loop update of underlying physical parameters strictly adheres to system energy constraints. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based simulated human cinema creative workflow method to solve the technical problem in the prior art that directly using visual gradients to update physical parameters easily disrupts the stable energy state of the system, leading to divergent and crashing physics engine solutions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention provides an AI-based simulated human creative workflow method for the theatrical release, comprising the following steps:

[0007] Step S100: Receive externally input multimodal intent data, perform semantic parsing on the multimodal intent data, and generate the objective functional and initial physical boundary parameter set for solving the human dynamics simulation.

[0008] Step S200: Under the condition of satisfying the preset multi-rigid-body nonlinear differential equation system constraints, the optimal motion state trajectory sequence and driving torque sequence are calculated by solving based on the target functional and the physical boundary parameter set.

[0009] Step S300: Read the motion state trajectory sequence, perform forward kinematic mapping and rasterization rendering, and output the scene map tensor;

[0010] Step S400: Calculate the visual loss between the scene map prediction tensor and the target multimodal intent reference tensor, and calculate the original visual gradient of the visual loss with respect to the physical boundary parameter set by solving the inverse ordinary differential equation;

[0011] Step S500: Obtain the Lyapunov energy functional index under the current state, construct the energy dissipation boundary manifold constraint, perform orthogonal projection operation on the original visual gradient, eliminate the gradient components in the original visual gradient that cause the total energy of the system to diverge non-physically, and output the corrected safe physical gradient.

[0012] Step S600: Based on the corrected security physical gradient, calculate and update the numerical variables in the physical boundary parameter set, and rewrite the updated physical boundary parameter set.

[0013] Step S700: Based on the updated physical boundary parameter set, perform iterative solving of the multi-rigid-body nonlinear differential equation system until the visual loss value drops to within the set tolerance threshold range, and output the final simulated human motion state trajectory sequence.

[0014] Further, the step of semantically parsing the multimodal intent data to generate the target functional and initial physical boundary parameter set for solving the human dynamics simulation includes: performing word segmentation and part-of-speech tagging on the natural language instruction sequence to separate entity word features, and extracting the three-dimensional keypoint coordinates and depth tensor information from the synchronously input storyboard image frame sequence through a visual feature extraction network; calling a pre-trained physical feature embedding network to encode the entity word features as physical intent latent variables, projecting the physical intent latent variables onto a specific physical parameter space through a multilayer perceptron decoder, and decoding to generate the contact surface Coulomb friction coefficient matrix, the environmental viscous damping coefficient matrix, and the joint maximum driving torque vector; combining the three-dimensional keypoint coordinates and the depth tensor information to perform inverse kinematics calculation and calculate the environmental space collision volume, and instantiating a physical boundary parameter set containing the contact surface Coulomb friction coefficient matrix, the environmental viscous damping coefficient matrix, the joint maximum driving torque vector, the system initial state vector, and the environmental geometric boundary constraint set.

[0015] Further, the step of solving the optimal motion state trajectory sequence and driving torque sequence based on the target functional and the physical boundary parameter set, under the premise of satisfying the preset multi-rigid-body nonlinear differential equation set, includes: the target functional is composed of two parts: the end-state algebraic cost function and the process integral cost function; constructing a simulated human basic nonlinear differential equation set unconstrained by the environment based on the principles of Lagrange mechanics; extracting the contact surface friction coefficient matrix from the physical boundary parameter set, orthogonally decomposing the contact force using the friction cone model to construct the contact force constraint conditions of the inequality; extracting the joint maximum driving torque vector and the environmental geometric boundary constraint set from the physical boundary parameter set, and constructing the control input saturation constraint conditions and spatial obstacle avoidance constraint conditions using the symbolic distance field algorithm; The fundamental nonlinear differential equations, contact force constraints, control input saturation constraints, and spatial obstacle avoidance constraints are tensor-concatenated to transform them into nonlinear state-space evolution equations and global inequality constraint functions. Using the direct collocation algorithm, adjacent time nodes in the discrete time node sequence are selected, and the nonlinear state-space evolution equations are transformed into system dynamic collocation equality constraint equations according to Simpson's integral rule. Lagrange multiplier vectors are introduced to construct a Lagrange function containing the discretized objective functional, the system dynamic collocation equality constraint equations, and the global inequality constraint functions. The first derivative is obtained to generate the Karosh-Kuhn-Tucker optimality condition system. Newton iteration optimization is performed in the state space, and the Hessian matrix approximation of the Lagrange function with respect to the decision variable matrix is ​​calculated using the quasi-Newton method. The linearized Karosh-Kuhn-Tucker optimality condition system is solved, and the optimal motion state trajectory sequence and driving torque sequence are output when the residual value is lower than the convergence tolerance threshold.

[0016] Furthermore, the steps of obtaining the Lyapunov energy functional index under the current state and constructing the energy dissipation boundary manifold constraints include: substituting the system state vector containing instantaneous generalized coordinates and instantaneous generalized velocities into the energy functional calculation formula to calculate the Lyapunov energy functional of the simulated person at the current physical evolution moment; establishing boundary inequality equations characterizing the passivity of the system, i.e., the time derivative of the system energy functional is less than or equal to the external control input power (defined as the inner product of the control input and the generalized velocity), forming a set of physical constraint manifold surfaces in the physical boundary parameter space; obtaining the partial derivative of the energy functional index with respect to the physical boundary parameter set, traversing the parameter components and splicing them to generate a complete normal gradient column vector, which is used to indicate the parameter space evolution direction that leads to non-physical abrupt changes in system energy.

[0017] Further, the step of performing an orthogonal projection operation on the original visual gradient to output the corrected safety physical gradient includes: calculating the projection component of the original visual gradient in the direction of the normal vector gradient column vector; subtracting the projection component from the original visual gradient to generate the corrected safety physical gradient.

[0018] Furthermore, in the step of updating the physical boundary parameter set, the absolute deviation value between the system energy index of the current iteration step and the preset reference benchmark energy scalar is read, and the adaptive update step size parameter is calculated based on the absolute deviation value; the inverse closed-loop update operation of the physical boundary parameter set is performed using the safe physical gradient and the adaptive update step size parameter, and the updated physical boundary parameter set matrix is ​​overwritten to the input memory region of the nonlinear dynamics solver to trigger the numerical solution closed loop of the next nonlinear state space evolution equation.

[0019] The present invention, by adopting the above technical solution, can bring the following beneficial effects:

[0020] 1. This invention extracts features from natural language instructions and storyboard images through an intent parsing processor, and translates unstructured multimodal data into a structured set of physical boundary parameters and objective functionals using a semantic-physical association mapping table and tensor transformation interface. This achieves the effect of converting subjective multimodal instructions into dynamic boundary conditions that can be directly called by the physics engine, and solves the technical problem of mathematical coupling between semantic intent and underlying physical dynamics in traditional workflows. This enables the nonlinear dynamics solver to have a strict optimization iteration benchmark and a unified boundary value basis.

[0021] 2. This invention uses a cluster of nonlinear dynamics solvers to concatenate the basic differential equations, contact force constraints, and obstacle avoidance constraints into state-space evolution equations. It then calls a sequence quadratic programming unit to execute collocation algorithms and Newton's iterative optimization, achieving the effect of outputting the optimal motion trajectory and driving torque sequence of multiple rigid bodies while adhering to strict laws of mechanics. This avoids physical distortion and clipping phenomena caused by direct generation algorithms, ensuring that the simulated human motion performance has realistic mass distribution and dynamic logic support.

[0022] 3. This invention constructs Lyapunov dissipative boundary manifold constraints for the system by using manifold projection feedback computing nodes, and uses orthogonal projection operator actuators to project the returned original visual gradient vector into the allowed energy dissipation tangent space to calculate the safe physical gradient. This achieves the effect of eliminating mathematical components that cause abnormal energy divergence when updating physical boundary parameters in a closed loop, preventing the parameter iterative evolution from deviating from the dynamic stable state, and ensuring that each parameter update closely follows the passive dissipation law of the system. Thus, while reducing visual bias, it avoids divergence and collapse in the solution of non-conservative controlled systems. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall method of AI-based simulated human cinema creative workflow of the present invention;

[0024] Figure 2 This is a schematic diagram of the simulated human workflow system of the present invention;

[0025] Figure 3 This is a flowchart of the nonlinear multi-rigid-body dynamics solution and optimal trajectory calculation of the present invention;

[0026] Figure 4 This is a flowchart illustrating the Lyapunov energy dissipation manifold construction and gradient orthogonal projection process of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1

[0029] See attached document Figure 1-2 This invention provides an AI-based simulated human creative workflow method for cinemas, comprising the following steps:

[0030] In step S100, the data processing unit 91 nested within the central controller 90 calls the intent parsing processor 10 to receive multimodal intent data input from the external terminal. It performs semantic parsing on the multimodal intent data to generate the objective functional and initial physical boundary parameter set for solving the human dynamics simulation. Specifically, the intent parsing processor 10 extracts action feature words from natural language commands and storyboard images. The system has a pre-set association mapping table. Based on this table, the intent parsing processor 10 maps action feature words to underlying physical quantity values. For example, it maps environmental descriptions in commands to friction coefficient ranges and action amplitude descriptions to joint damping values, constructing a physical boundary parameter set containing multiple dimensional constants. Target functional The construction formula is as follows:

[0031]

[0032] In the formula, Represents the system state vector. Represents the control input vector. Represents the end of the movement. The terminal state algebraic cost function, The integral cost function represents the process.

[0033] Step S200, the nonlinear dynamics solver cluster 20 receives the target functional With physical boundary parameter set Under the premise of satisfying the preset constraints of the multi-rigid-body nonlinear differential equation system, the optimal motion state trajectory sequence and driving torque sequence are calculated. The nonlinear dynamics solver cluster 20 uses a sequential quadratic programming algorithm to iteratively search in the state space to find the target functional The state vector that takes the minimum value, and the motion state trajectory sequence output by step S200 in the current physical boundary parameter set. Under constraints, it conforms to the laws of mechanics;

[0034] In step S300, the differentiable rendering workstation cluster 30 reads the motion state trajectory sequence output by the nonlinear dynamics solver cluster 20, performs forward kinematic mapping and rasterization rendering, and outputs a scene graph tensor. The forward kinematic mapper uses a homogeneous coordinate transformation matrix to convert the underlying joint angle data into vertex coordinates of the simulated human surface topological mesh. The differentiable rasterization renderer processes the topological mesh and ambient lighting parameters to generate a scene graph tensor containing color, depth and normal features.

[0035] In step S400, the differentiable rendering workstation cluster 30 calculates the visual loss between the scene map prediction tensor and the target multimodal intent reference tensor, and calls the subsequent adjoint state gradient calculation logic circuit to calculate the visual loss with respect to the physical boundary parameter set by solving the inverse ordinary differential equation. The original visual gradient;

[0036] In step S500, the manifold projection feedback computing node 40 obtains the Lyapunov energy functional index of the nonlinear dynamics solver cluster 20 in the current state, constructs the energy dissipation boundary manifold constraint, performs an orthogonal projection operation on the original visual gradient, and outputs the corrected safe physical gradient. The manifold projection feedback computing node 40 eliminates the gradient components in the original visual gradient that cause the total energy of the system to diverge non-physically through projection matrix operations. The projection operation preserves the parameter evolution direction located in the energy stable dissipation tangent space of the physical system.

[0037] In step S600, the manifold projection feedback computing node 40 calculates and updates the physical boundary parameter set based on the corrected safety physics gradient. The numerical variables in the data will be updated, and the updated physical boundary parameter set will be rewritten into the input cache of the nonlinear dynamics solver cluster 20. The update formula for the physical boundary parameter set is as follows:

[0038]

[0039] In the formula, Representing the The physical boundary parameter set matrix after the next iteration. Representing the The current physical boundary parameter set matrix in the next iteration Representing the The step size parameter is updated adaptively in the next iteration. This represents the corrected security physical gradient;

[0040] In step S700, the nonlinear dynamics solver cluster 20 re-executes the solution process of the multi-rigid-body nonlinear differential equation system based on the updated physical boundary parameter set. The system iterates between steps S200 and S600 until the visual loss value calculated in step S400 drops to the set tolerance threshold range, at which point it is determined that the iteration has converged, the workflow ends, and the final simulated human motion state trajectory sequence is output.

[0041] Furthermore, the above method is implemented based on an AI-based simulated human cinema creative workflow system. The system is built around a central controller 90 at the hardware topology level. The central controller 90 internally integrates a data processing unit 91, which includes an intent parsing processor 10, a nonlinear dynamics solver cluster 20, a differentiable rendering workstation cluster 30, and a manifold projection feedback computing node 40. Communication connections are established between the intent parsing processor 10, the nonlinear dynamics solver cluster 20, the differentiable rendering workstation cluster 30, and the manifold projection feedback computing node 40. Low-level data interaction is carried out by relying on the peripheral component interconnection expansion bus to ensure low-latency transmission of physical tensors and rendering maps in the closed-loop iteration process.

[0042] For the allocation of parallel computing resources and the network communication mechanism between nodes in a multi-computing node environment, those skilled in the art can use a unified computing device architecture or a message passing interface to build it. The data communication protocol and routing configuration of the unified computing device architecture and the message passing interface are well known technologies in the field and will not be described in detail here.

[0043] The intent parsing processor 10 is internally equipped with random access memory and semantic parsing logic circuit. The semantic parsing logic circuit receives multimodal data sent by an external input terminal. The multimodal data includes natural language text and image segmentation sequences.

[0044] The semantic parsing logic circuit translates unstructured multimodal data into a structured set of physical boundary parameters and an objective functional. The underlying driver of the intent parsing processor 10 is equipped with a tensor conversion interface. The tensor conversion interface extracts semantic constraints from natural language and converts them into specific physical quantities in floating-point data format, such as the lower limit of the friction coefficient and the threshold of the joint damping coefficient, according to a fixed data type protocol. Then, the specific physical quantities are written into the designated physical engine memory address of the data processing unit 91. The nonlinear dynamics solver cluster 20 deploys a multi-rigid-body dynamics calculation logic circuit and a nonlinear programming solver. The nonlinear programming solver includes a sequential quadratic programming operation unit and an interior point method operation unit.

[0045] Under the constraints of specified physical boundary parameters, the nonlinear dynamics solver cluster 20 calls the microprocessor to execute the iteration of the nonlinear dynamics differential equations of the system. The central processing unit in the nonlinear dynamics solver cluster 20 is responsible for handling the logical branches and determining the system constraints. The tensor processing unit calls the hardware multiplier array to handle the Hessian matrix and Jacobian matrix operations in the evolution of the differential equations. The differentiable rendering workstation cluster 30 is equipped with a forward kinematics mapper and a differentiable rasterization renderer. The forward kinematics mapper includes a homogeneous coordinate transformation matrix multiplier and a linear hybrid skin weight calculation unit.

[0046] The forward kinematics mapper maps the generalized coordinate matrix of the bottom joints output by the nonlinear dynamics solver cluster 20 to the vertex coordinate matrix of the upper-layer simulated human 3D topological mesh through the skeleton skin weight algorithm. The differential rasterizer is built based on a differentiable graphics pipeline and has the ability to perform chain-like differentiation of computational graphs.

[0047] When the differentiable rasterizer performs forward rasterization calculations and outputs the scene map tensor, it synchronously maintains the corresponding gradient backpropagation calculation graph in the video memory. The manifold projection feedback computing node 40 is configured in the data feedback path, connecting the video memory output end of the differentiable rendering workstation cluster 30 and the memory input end of the nonlinear dynamics solver cluster 20. The manifold projection feedback computing node 40 embeds a Hamiltonian energy calculation unit and an orthogonal projection operator executor.

[0048] The Lyapunov energy calculation unit reads the mass distribution matrix and instantaneous velocity vector of each moving part of the simulated human, and calculates the total mechanical energy scalar (i.e., Hamiltonian) of the system according to mechanical formulas. Its Lyapunov energy functional... The calculation formula is:

[0049]

[0050] In the formula, Represents the instantaneous generalized coordinate vector. Represents the instantaneous generalized velocity vector. The generalized mass matrix representing the simulated human. Represents the system's potential energy. Represents the set of physical boundary parameters;

[0051] The orthogonal projection operator executor includes vector projection calculation hardware to correct the raw visual gradient data returned by the 30 differentiable rendering workstation clusters, resulting in a corrected safe physical gradient. Calculated using the following formula:

[0052]

[0053] In the formula, This represents a minimal constant that prevents mathematical singularities caused by a denominator of zero. Represents the original visual gradient vector. Represents the energy functional with respect to the physical boundary parameter set The normal vector and gradient column vector;

[0054] The manifold projection feedback computing node 40 converts the corrected safe physical gradient data into parameter update instructions. The manifold projection feedback computing node 40 has direct memory access permissions. According to the set step size scaling rules, it overwrites the physical boundary parameter set in the input cache of the nonlinear dynamics solver cluster 20 through the parameter update instructions, triggering the second round of state space closed-loop solution in the internal hardware layer of the data processing unit 91.

[0055] Example 2

[0056] See attached document Figure 3-4 Based on Example 1, step S100 specifically includes the following steps for the logic control of multimodal semantic parsing and physical boundary extraction:

[0057] In step S101, the communication interface 11 inside the intent parsing processor 10 receives multimodal intent data transmitted from an external terminal. The multimodal intent data is formed by splicing time-aligned natural language instruction sequences and storyboard image frame sequences. The lexicalization logic unit 12 inside the intent parsing processor 10 performs word segmentation and part-of-speech tagging on the natural language instruction sequences, separating entity word features that represent the moving subject, environmental state, and the degree of force applied to the action. The storyboard image frame sequence is synchronously input into the visual feature extraction network 13. The visual feature extraction network 13 extracts the three-dimensional key point coordinates representing the initial physical posture of the character and the depth tensor information representing the geometric occlusion relationship of the environment through forward inference calculation.

[0058] In step S102, the intent parsing processor 10 calls the pre-trained physical feature embedding network to perform high-dimensional space mapping based on the separated entity word features. The physical feature embedding network encodes the entity word features into continuous physical intent latent variables through a self-attention mechanism. Then, through the internal multilayer perceptron decoder, the physical intent latent variables are projected onto a specific physical parameter space. When the entity word features indicate the environmental surface state, the intent parsing processor 10 decodes the surface state features and outputs the contact surface Coulomb friction coefficient matrix and the environmental viscous damping coefficient matrix. When the entity word features indicate the force exertion of the character's action, the intent parsing processor 10 maps and generates the physical saturation upper limit value of the maximum driving torque of each degree of freedom joint in the simulated human skeleton model based on the activation weight of the latent variables.

[0059] In step S103, the intent parsing processor 10 performs inverse kinematics calculations based on the coordinates of the three-dimensional key points to generate the initial state vector of the simulated human system. The intent parsing processor 10 calculates the initial environmental space collision volume based on the depth tensor information. The intent parsing processor 10 summarizes the calculated physical parameters and instantiates the physical boundary parameter set in the data processing unit 91. Specifically, the physical boundary parameter set. Includes a matrix of friction coefficients between the simulated human foot and the ground. Environmental fluid viscosity damping coefficient matrix Maximum driving torque vector of the joint System initial state vector and the set of environmental geometric boundary constraints containing the coordinates of spatial obstacles. ;

[0060] Step S104, intending to parse the tensor conversion interface 15 at the underlying level of the processor 10 to obtain the contact surface friction coefficient matrix. Environmental fluid viscosity damping coefficient matrix Maximum driving torque vector of the joint System initial state vector and the set of environmental geometric boundary constraints According to the preset memory alignment rules, it is encapsulated into a high-dimensional floating-point parameter tensor. The intent parsing processor 10 writes the high-dimensional floating-point parameter tensor into the physical engine shared memory area of ​​the data processing unit 91 through the peripheral component interconnection expansion bus according to the direct memory access protocol, so that it can be called as a boundary condition when the nonlinear dynamics solver cluster 20 initializes the multi-rigid-body differential equation system.

[0061] Furthermore, in step S100, the steps for constructing the functional model of the moving target include:

[0062] In step S105, the functional construction logic module 16 inside the data processing unit 91 receives the semantic features and the initial state vector of the system output by the intent parsing processor 10. The functional construction logic module 16 starts the mathematical instantiation program of the Lagrange-Borza form objective functional. The objective functional is composed of two parts: the terminal state algebraic cost function and the process integral cost function, which are used to provide the convergence direction and gradient evaluation benchmark in the subsequent nonlinear programming solution.

[0063] Step S106: Based on the end-effector posture requirements in the multimodal intent data, the functional construction logic module 16 calculates and constructs the end-effector state algebraic cost function. This end-effector state algebraic cost function is used to quantitatively evaluate the Euclidean distance deviation between the actual physical state vector of the simulated human at the motion terminal time point and the preset target state vector. This embodiment discloses the end-effector state algebraic cost function. The specific mathematical expansion structure:

[0064]

[0065] In the formula, Represents the system at the terminal time The actual state vector at that time, This represents the target state vector generated by the feature mapping of the storyboard images. Represents the end-bias penalty weight matrix. Represents the set of physical boundary parameters, superscript Represents the matrix transpose operation;

[0066] Step S107: Functional construction logic module 16 extracts the physical boundary parameter set. The boundary constraint variables are configured with the terminal bias penalty weight matrix. End-point deviation penalty weight matrix The functional construction logic module 16, based on the kinematic degrees of freedom priority of each key joint of the character, constructs a weight matrix for end-effector deviation penalty. The main diagonal values ​​are injected with different values. The differential settings of the main diagonal values ​​are used to adjust the proportion of optimization computing power allocated to each joint degree of freedom when the nonlinear programming solver is approximating the end attitude.

[0067] Step S108: The functional construction logic module 16 reads the physical parameters representing action style and environmental resistance from the multimodal intention data, and constructs a process integral cost function. The process integral cost function is used to measure the total energy consumption of the simulated human along the entire time domain axis and the smoothness of the motion trajectory. This embodiment discloses the process integral cost function. The specific mathematical implementation structure:

[0068]

[0069] In the formula, The control input vector at the current moment represents the joint driving torque matrix. Represents the instantaneous generalized velocity vector; Represents the control torque penalty weight matrix; Represents the speed penalty weight matrix;

[0070] In step S109, the functional construction logic module 16 performs a dynamic weight allocation operation and reads the physical boundary parameter set. The threshold of force exertion in Scalar eigenvalues ​​of environmental viscous damping coefficient Functional construction logic module 16 is based on the force intensity threshold. Through formula Calculate and update the control torque penalty weight matrix main diagonal element , in the formula, This is the first preset constant. To prevent extremely small constants with a denominator of zero, the above inverse proportional calculation method reduces the weight matrix of the control torque penalty. The element values ​​are used to expand the system's tolerance to instantaneous driving torque. The functional construction logic module 16 is based on the scalar eigenvalues ​​of the environmental viscous damping coefficient. Through formula Calculate the update rate penalty weight matrix main diagonal element , in the formula, It is a direct proportionality coefficient. Based on the basic smoothing bias constant, the above proportional calculation method increases the velocity penalty weight matrix. The element values ​​enable the system to generate low-speed motion trajectories to suppress oscillations;

[0071] In step S110, the functional construction logic module 16 encapsulates the instantiated terminal state algebraic cost function and process integral cost function to generate a numerical target functional memory block. The functional construction logic module 16 then passes the address pointer of the numerical target functional memory block to the constraint evaluation queue of the nonlinear dynamics solver cluster 20 through the system's internal memory mapping mechanism.

[0072] Furthermore, step S200 specifically includes the following steps for defining the multi-rigid-body dynamic equations and constraints:

[0073] In step S201, the nonlinear dynamics solver cluster 20 nested within the data processing unit 91 calls the internal multi-rigid-body dynamics calculation logic circuit 21, which reads the physical boundary parameter set passed in by the intent parsing processor 10. With the system initial state vector System state vector In the underlying mathematical space, it is defined as containing the generalized coordinates of each node of the simulated human skeleton topology system. With generalized speed Column vectors;

[0074] Step S202: The multi-rigid-body dynamics calculation logic circuit 21 constructs a set of basic nonlinear differential equations for the simulated human without environmental constraints, based on the principles of Lagrange mechanics. This set of basic nonlinear differential equations describes the physical evolution of each rigid body link of the simulated human under the action of internal and external forces in the time domain. This embodiment discloses the specific mathematical structure of the set of basic nonlinear differential equations:

[0075]

[0076] In the formula, Represents the generalized acceleration column vector; Represents the generalized quality matrix of the system; Represents the matrix of Coriolis force and centrifugal force; This represents the viscous damping matrix, which contains a set of physical boundary parameters. The environmental fluid viscosity damping coefficient matrix ; Represents the gravitational potential energy gradient matrix; Represents the driving mapping matrix; Represents the column vector of joint control inputs; Jacobian matrix representing the contact point; Representing the column vector of external contact forces, the dimensions of each physical term in the fundamental nonlinear differential equation system are strictly unified as torque dimensions to ensure consistency of physical dimensions.

[0077] Step S203, Multi-rigid-body dynamics calculation logic circuit 21 extracts physical boundary parameter set The contact surface friction coefficient matrix Establish contact force constraints characterizing the interaction with the physical environment. The multi-rigid-body dynamics calculation logic circuit 21 will convert the external contact force column vector. Orthogonally decomposed into normal contact forces With tangential friction The multi-rigid-body dynamics calculation logic circuit 21 uses a friction cone model to construct inequality constraint equations:

[0078] Step S204, Multi-rigid-body dynamics calculation logic circuit 21 extracts physical boundary parameter set Maximum driving torque vector of joints and the set of environmental geometric boundary constraints Construct system path constraints and spatial obstacle avoidance constraints, and construct control input saturation constraint equations using the multi-rigid-body dynamics calculation logic circuit 21. Multi-rigid-body dynamics calculation logic circuit 21 utilizes forward kinematic mapping function generalized coordinates Mapped to the first digit of the simulated human topology mesh The multi-rigid-body dynamics calculation logic circuit 21 uses the symbolic distance field algorithm to calculate the shortest spatial Euclidean distance from the three-dimensional Cartesian space coordinate column vector of each key point to the environmental geometric boundary, and constructs a non-negative inequality anti-penetration constraint equation: In the formula, Representing the The positive kinematic mapping function of each key point Represents the symbolic distance field operator;

[0079] In step S205, the multi-rigid-body dynamics calculation logic circuit 21 performs tensor concatenation of the basic nonlinear differential equations, contact force constraints, control input saturation constraints, and spatial obstacle avoidance constraints, transforming them into standard nonlinear state-space evolution equations. With global inequality constraint functions The multi-rigid-body dynamics calculation logic circuit 21 writes the transformed nonlinear state-space evolution equation and global inequality constraint function into the shared memory stack 22 of the nonlinear dynamics solver cluster 20.

[0080] Furthermore, in step S200, the optimal trajectory calculation process based on nonlinear programming includes the following steps:

[0081] In step S206, the data processing unit 91 nested inside the central controller 90 calls the sequence quadratic programming operation unit 23 of the nonlinear dynamics solver cluster 20. The sequence quadratic programming operation unit 23 reads the numerical objective functional memory block, the nonlinear state space evolution equation and the global inequality constraint function from the shared video memory stack 22. The sequence quadratic programming operation unit 23 initializes the discretized time grid in the hardware register. The time interval of the motion terminal is divided into a finite number of discrete time steps to form a discrete time node sequence.

[0082] In step S207, the sequential quadratic programming operation unit 23 uses the direct collocation algorithm to transform the optimal control functional in the continuous time domain into a finite-dimensional nonlinear programming algebraic system. The sequential quadratic programming operation unit 23 selects adjacent time nodes in the discrete time node sequence and uses polynomial interpolation logic to construct discrete representation matrices of the state vector and control input vector. The sequential quadratic programming operation unit 23 then generates the system dynamic collocation equality constraint equations based on Simpson's integral rule.

[0083]

[0084] In the formula, Representing the The state vector at each time point Representing the The state vector at each time point A scalar representing the discrete time step size between adjacent time points. Representing the The evaluation vector of the nonlinear state-space evolution function at each time node. The vector representing the evaluation of the evolution function at the midpoint of time, where the state vector at the midpoint of time is... With control vector It participates in the optimization solution as an independent decision variable;

[0085] In step S208, the sequential quadratic programming operation unit 23 parses the numerical objective functional memory block into a discrete scalar cost function containing state decision variables and control decision variables. The sequential quadratic programming operation unit 23 introduces Lagrange multiplier vectors to construct the Lagrange function:

[0086]

[0087] In the formula, This represents the column vector of global decision variables after assembly. Represents the Lagrange multiplier vectors constrained by equality. Represents the Lagrange multiplier vectors constrained by inequality. This represents the set of all dynamic collocation equality constraint equations. Representing the set of all global inequality constraint functions, the sequence quadratic programming operation unit 23 pairs of Lagrangian functions. Taking the first derivative, we can generate the Karosh-Kuhn-Tucker optimality condition system;

[0088] In step S209, the sequential quadratic programming operation unit 23 performs Newton iteration optimization in the state space. The sequential quadratic programming operation unit 23 calculates the approximate value of the Lagrangian function with respect to the Hessian matrix of the decision variable matrix using a quasi-Newton method. The sequential quadratic programming operation unit 23 then uses the BFGS algorithm to perform the Hessian matrix approximation. The iterative update, specifically the update formula is:

[0089]

[0090] In the formula, Representing the Approximate value of Hessian matrix in the next iteration Representing the The step size vector of decision variables in the next iteration Representing the The gradient difference vector of the Lagrange function in the next iteration, the sequential quadratic programming operation unit 23 solves the linearized Karosh-Kun-Tucker optimality condition system, calculates the search direction matrix of the decision variables in the current iteration step, the sequential quadratic programming operation unit 23 executes the line search algorithm to determine the descent step size scalar, and writes the updated global decision variable column vector into memory;

[0091] In step S210, the sequential quadratic programming operation unit 23 evaluates the Karlosh-Kun-Tucker optimality condition residual value under the current global decision variable column vector. When the optimality condition residual value is lower than the preset convergence tolerance threshold, the sequential quadratic programming operation unit 23 terminates the Newton iterative optimization process. The sequential quadratic programming operation unit 23 parses the converged global decision variable column vector and outputs the optimal motion state trajectory sequence and the optimal driving torque sequence in the discrete state.

[0092] Furthermore, in step S300, the forward kinematics and scene graph rendering pipeline specifically includes the following steps:

[0093] In step S301, the data processing unit 91 nested inside the central controller 90 calls the differentiable rendering workstation cluster 30. The forward kinematics mapper 31 in the differentiable rendering workstation cluster 30 reads the optimal motion state trajectory sequence output by the nonlinear dynamics solver cluster 20 from the shared video memory stack 22. The optimal motion state trajectory sequence contains the generalized coordinate matrix of the simulated human joint at each discrete time node.

[0094] Step S302, the forward kinematics mapper 31 performs geometric position mapping from bones to the skin mesh. The forward kinematics mapper 31 uses a linear hybrid skinning algorithm to convert the underlying joint generalized coordinate matrix into vertex coordinates of the three-dimensional topological mesh. This embodiment discloses the mathematical structure of the forward mapping of vertex coordinates:

[0095]

[0096] In the formula, Represents the first deformed mesh model The three-dimensional spatial position vector of each vertex in the world coordinate system. Representing the The initial position vectors of each vertex under the bound pose, both using homogeneous coordinate representation format; Representative influence The total number of skeletal joints at each vertex. Representing the The first skeletal joint pair The skinning weight scalar of each vertex, Represented by joint generalized coordinates The calculated first The fourth-order homogeneous coordinate transformation matrix of each skeletal joint;

[0097] In step S303, the differentiable rasterizer 32, which is called internally by the differentiable rendering workstation cluster 30, receives the deformable vertex coordinates output by the forward kinematics mapper 31, and constructs a three-dimensional scene map by combining the preset camera intrinsic matrix, camera extrinsic matrix and ambient lighting parameter tensor. The differentiable rasterizer 32 performs vertex shading and primitive assembly operations to project the topological mesh in the three-dimensional space onto the two-dimensional screen coordinate system.

[0098] In step S304, the differentiable rasterizer 32 performs fragment shading calculations and outputs a scene map tensor containing multi-channel features. To support subsequent closed-loop gradient backpropagation calculations, the differentiable rasterizer 32 calculates the outgoing radiance tensor of screen pixels using continuously differentiable rendering equations. :

[0099]

[0100] In the formula, Represents the coordinates of the intersection points of the three-dimensional surface grid corresponding to a pixel in screen space. Represents the observation direction vector. This represents the emissivity at the coordinates of the intersection of the surface grid. Represents the hemispherical illumination integral domain. Represents the bidirectional reflection distribution function. Represents the incident light radiance. This represents the incident light direction vector. The scene graph tensor represents the normal vector at the intersection of the surface mesh. In terms of data structure, it is a multi-dimensional spatial matrix, which includes the color channel attribute, depth channel attribute, and surface normal channel attribute of the rendering plane. The color channel attribute records the three primary color values ​​of each pixel. The depth channel attribute records the orthogonal spatial distance from the mesh surface to the camera projection plane. The surface normal channel attribute records the spatial orientation of the normal vector of each facet of the mesh surface.

[0101] In step S305, when the differential rasterizer 32 performs forward rendering calculations, it simultaneously constructs a directed acyclic graph (DAG) for forward calculations in the system video memory. The nodes of the DAG record the tensor data states of each rendering stage, and the edges of the DAG record the matrix multiplication and rasterization interpolation operators in the graphics pipeline. The DAG provides a backpropagation data structure carrier for the subsequent calculation of the partial derivatives of the visual loss function with respect to the physical boundary parameter set.

[0102] Furthermore, in step S400, the visual loss and gradient calculation process includes the following steps:

[0103] In step S401, the visual loss assessment logic circuit 33 inside the differentiable rendering workstation cluster 30 receives the scene map tensor output by the differentiable rasterizer renderer 32, and synchronously reads the target reference video frame tensor generated by the video generation diffusion model called by the intent parsing processor 10 based on multimodal intent data. The visual loss assessment logic circuit 33 uses the target reference video frame tensor as a visual comparison benchmark, quantifies the visual deviation between the predicted rendering result and the reference target, and constructs an integral visual loss function in the time domain. This embodiment discloses the specific mathematical expansion of the visual loss function:

[0104]

[0105] In the formula, Represents a scalar value for visual loss. Represents the end time of the movement. Representing time nodes The scene map prediction tensor output by the subdivisible rasterizer 32 Representing time nodes The target multimodal intent tensor below, The square norm operation represents the Euclidean distance between tensors;

[0106] In step S402, the differentiable rendering workstation cluster 30 calls the adjoint state gradient calculation logic circuit 34. The adjoint state gradient calculation logic circuit 34 starts the backpropagation computation graph and calculates the partial derivative of the visual loss scalar value with respect to the system state vector according to the chain rule. The adjoint state gradient calculation logic circuit 34 performs tensor differential multiplication.

[0107]

[0108] In the formula, The spatial gradient matrix representing the visual loss with respect to the system state vector. The error matrix representing the visual loss with respect to the scene map prediction tensor. The Jacobian matrix representing the pixel channel in a differentiable rendering pipeline with respect to the surface topology vertex coordinates. The Jacobian matrix representing the mesh vertices with respect to the underlying state vector in the positive kinematic mapping;

[0109] In step S403, the adjoint state gradient calculation logic circuit 34 initializes the adjoint state vector variable in the shared memory stack 22. The adjoint state vector variable is used to accumulate gradient information in reverse without storing the complete history of the forward evolution state. The adjoint state gradient calculation logic circuit 34 constructs the adjoint ordinary differential equation in the continuous time domain:

[0110]

[0111] In the formula, represent The adjoint state column vector at time t, This represents the first derivative vector of the adjoint state column vector with respect to time. The equation represents the nonlinear state-space evolution function. The matrix representing the partial derivatives of the evolution function equation with respect to the state vector, and the terminal boundary conditions of the adjoint ordinary differential equation are set as follows: This ensures the effective backpropagation of the algebraic cost gradient of the terminal state;

[0112] Step S404: The adjoint state gradient calculation logic circuit 34 applies the Runge-Kutta numerical integration algorithm to solve the adjoint ordinary differential equation in reverse along the time axis. The adjoint state gradient calculation logic circuit 34 starts from the motion terminal time. From the start of the calculation to the initial time At the end of the time interval, calculate and output the sequence of evolution trajectories of the adjoint state vectors over the entire time domain. ;

[0113] Step S405, the adjoint state gradient calculation logic circuit 34 calculates the trajectory sequence of the adjoint state vector obtained from the solution. The partial derivatives of the visual loss scalar value with respect to the physical boundary parameter set are calculated. The adjoint state gradient calculation logic circuit 34 performs definite integral operations in the time domain and outputs the original visual gradient vector.

[0114]

[0115] In the formula, Represents the original visual gradient vector. The equation representing the nonlinear state-space evolution function with respect to the physical boundary parameter set The partial derivative matrix;

[0116] In step S406, the adjoint state gradient calculation logic circuit 34 calculates and outputs the original visual gradient vector. The data feedback communication bus 35 is written to the data feedback communication bus 35, which transmits the original visual gradient vector to the manifold projection feedback computing node 40.

[0117] Furthermore, in step S500, the Hamiltonian energy manifold constraint construction process includes the following steps:

[0118] Step S501: The manifold projection feedback computing node 40 reads the system state vector and physical boundary parameter set of the current iteration step of the nonlinear dynamics solver cluster 20. The system state vector includes the instantaneous generalized coordinates and instantaneous generalized velocity of the simulated human.

[0119] In step S502, the Lyapunov energy calculation unit 41 inside the manifold projection feedback computing node 40 calculates the total mechanical energy scalar of the system. The Lyapunov energy calculation unit 41 substitutes the system state vector into the energy functional calculation formula to calculate the Lyapunov energy functional of the simulated human at the current physical evolution moment. The specific calculation formula is consistent with the formula for calculating the energy functional of the system in the aforementioned data processing unit 91;

[0120] In step S503, the manifold projection feedback computing node 40 constructs a Lyapunov energy dissipation boundary manifold in the physical boundary parameter space. The manifold projection feedback computing node 40 establishes boundary inequality equations based on the system passivity theory, i.e., the time derivative of the system energy functional. Less than or equal to external control input power (defined as control input vector) With generalized velocity vector The inner product, i.e. ),satisfy The physical essence of this inequality lies in the fact that the increase in the total mechanical energy within a controlled nonlinear system is strictly limited to the upper limit of the work done by the external active force (natural damping dissipation further reduces the derivative), ensuring that the update of parameters by the visual inverse gradient does not violate the laws of mechanics and inject energy into the system out of thin air, thus ensuring that the system operates within the N-dimensional physical boundary parameter set. A set of physically constrained manifold surfaces is formed in space. ;

[0121] Step S504: The manifold projection feedback computing node 40 calculates the normal vector gradient of the energy dissipation boundary manifold at the current physical boundary parameter point. The manifold projection feedback computing node 40 obtains the partial derivative of the energy functional index with respect to the physical boundary parameter set. The normal vector gradient indicates the parameter space evolution direction that leads to a non-physical abrupt change in the system energy. This embodiment discloses the partial derivative matrix expansion structure of the normal vector gradient for the physical boundary parameter set. The first in Each parameter component The gradient of the energy functional normal vector The formula for calculating a scalar element is:

[0122]

[0123] In the formula, The first one representing the gradient of the normal vector A scalar element, manifold projection feedback computing power node 40 traversals of all Each parameter component is concatenated to generate a complete normal vector and gradient column vector. ;

[0124] In step S505, the manifold projection feedback computing node 40 constructs the local tangent space of the physically constrained manifold surface based on the normal vector gradient. The local tangent space consists of the set of all vectors orthogonal to the normal vector gradient. The manifold projection feedback computing node 40 instantiates the tangent space basis vector matrix in memory, which limits the parameter correction amount driven by visual bias in subsequent steps to be strictly distributed within the local tangent space, ensuring that the parameter iteration process follows the system's stable energy dissipation law.

[0125] Furthermore, in step S500, the process for constructing the Lyapunov energy dissipation boundary manifold constraint includes the following steps:

[0126] In step S506, the orthogonal projection operator executor 42 inside the manifold projection feedback computing node 40 reads the original visual gradient vector transmitted from the data feedback communication bus 35. The orthogonal projection operator executor 42 synchronously reads the normal gradient column vector output by the Lyapunov energy calculation unit 41. The original visual gradient vector indicates the pure mathematical optimization direction to reduce visual bias, and the normal gradient column vector indicates the physical over-boundary direction to deviate from the stable energy dissipation state of the system.

[0127] In step S507, the orthogonal projection operator executor 42 calls its internal hardware multiply-accumulator array to perform orthogonal projection operation on the high-dimensional space matrix of the gradient vector. The orthogonal projection operator executor 42 calculates the projection component of the original visual gradient vector in the direction of the normal gradient column vector. The orthogonal projection operator executor 42 subtracts the projection component from the original visual gradient vector and outputs the corrected safe physical gradient column vector. The specific mathematical structure of the orthogonal projection operation is the same as the correction formula used in the aforementioned system architecture section. The safe physical gradient column vector is distributed in the spatial geometry inside the tangent space of the Lyapunov energy dissipation boundary manifold.

[0128] Furthermore, in step S600, the specific process for updating the physical boundary parameter set includes the following steps:

[0129] In step S601, the step size control logic circuit 43 embedded in the manifold projection feedback computing node 40 calculates the adaptive update step size parameters. The step size control logic circuit 43 reads the absolute deviation value between the system Lyapunov energy functional index of the current iteration step and the preset reference energy scalar. The step size control logic circuit 43 calculates the adaptive update step size parameters based on the absolute deviation value. The specific step size reduction calculation formula is as follows:

[0130]

[0131] In the formula, Representing the The step size parameter is updated adaptively in the next iteration. This represents the preset base step size constant. Represents the damping attenuation coefficient. Representing the The Lyapunov energy functional index of the system under the next iteration. The step size reduction formula is used to prevent parameters from moving too far along the tangent plane during updates, which would lead to higher-order truncation errors and ensure that the parameter points are closely attached to the physical constraint manifold surface.

[0132] In step S602, the manifold projection feedback computing node 40 uses the calculated safe physical gradient column vector and the adaptive update step size parameter to perform the reverse closed-loop update operation of the physical boundary parameter set. The specific algebraic structure of the boundary parameter closed-loop update is the same as the formula recorded in step S600 above.

[0133] In step S603, the manifold projection feedback computing node 40 overwrites the updated physical boundary parameter set matrix to the input memory region of the nonlinear dynamics solver cluster 20 through a preset direct memory access instruction. Based on the overwritten physical boundary parameter set matrix, the data processing unit 91 triggers the underlying logic circuit to execute the numerical solution closed loop of the next nonlinear state space evolution equation. The system continues to run steps S200 to S600 until the visual loss value output by the differentiable rendering workstation cluster 30 decreases to within the preset convergence tolerance range.

Claims

1. An AI-based simulated human creative workflow method for film and television production, characterized in that, Includes the following steps: Step S100: Receive externally input multimodal intent data, perform semantic parsing on the multimodal intent data, and generate the objective functional and initial physical boundary parameter set for solving the human dynamics simulation. Step S200: Under the condition of satisfying the preset multi-rigid-body nonlinear differential equation system constraints, the optimal motion state trajectory sequence and driving torque sequence are calculated by solving based on the target functional and the physical boundary parameter set. Step S300: Read the motion state trajectory sequence, perform forward kinematic mapping and rasterization rendering, and output the scene map tensor; Step S400: Calculate the visual loss between the scene map prediction tensor and the target multimodal intent reference tensor, and calculate the original visual gradient of the visual loss with respect to the physical boundary parameter set by solving the inverse ordinary differential equation; Step S500: Obtain the Lyapunov energy functional index under the current state, construct the energy dissipation boundary manifold constraint, perform orthogonal projection operation on the original visual gradient, eliminate the gradient components in the original visual gradient that cause the total energy of the system to diverge non-physically, and output the corrected safe physical gradient. Step S600: Based on the corrected security physical gradient, calculate and update the numerical variables in the physical boundary parameter set, and rewrite the updated physical boundary parameter set. Step S700: Based on the updated physical boundary parameter set, perform iterative solving of the multi-rigid-body nonlinear differential equation system until the visual loss value drops to within the set tolerance threshold range, and output the final simulated human motion state trajectory sequence.

2. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, Step S100 includes: The natural language instruction sequence is segmented and part-of-speech tagged to separate entity word features. The synchronously input storyboard image frame sequence is processed by a visual feature extraction network to extract the 3D key point coordinates and depth tensor information. The pre-trained physical feature embedding network is invoked to encode the entity word features as physical intent latent variables. The physical intent latent variables are then projected onto a specific physical parameter space through a multilayer perceptron decoder to decode and generate the contact surface Coulomb friction coefficient matrix, the environmental viscous damping coefficient matrix, and the joint maximum driving torque vector. Combining the coordinates of the three-dimensional key points with the depth tensor information, inverse kinematics calculation is performed to calculate the collision volume in the environment space. A physical boundary parameter set is instantiated, which includes the Coulomb friction coefficient matrix of the contact surface, the environmental viscous damping coefficient matrix, the maximum driving torque vector of the joint, the initial state vector of the system, and the set of environmental geometric boundary constraints.

3. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, The objective functional is composed of two parts: an end-state algebraic cost function and a process integral cost function. The construction formula is: In the formula, Represents the system state vector. Represents the control input vector. Represents the end of the movement. The terminal state algebraic cost function, The integral cost function represents the process. Represents the set of physical boundary parameters.

4. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, Step S200 includes: Based on the principles of Lagrange mechanics, a set of fundamental nonlinear differential equations for a simulated human being unconstrained by the environment is constructed. Extract the contact surface friction coefficient matrix from the physical boundary parameter set, and use the friction cone model to orthogonally decompose the contact force to construct the contact force constraint conditions of the inequality. Extract the joint maximum driving torque vector and the environmental geometric boundary constraint set from the physical boundary parameter set, and use the symbolic distance field algorithm to construct control input saturation constraint conditions and spatial obstacle avoidance constraint conditions; The fundamental nonlinear differential equations, the contact force constraints, the control input saturation constraints, and the spatial obstacle avoidance constraints are tensor-concatenated to transform them into nonlinear state-space evolution equations and global inequality constraint functions.

5. The AI-based simulated human cinema creative workflow method according to claim 4, characterized in that, Also includes: Using the direct collocation algorithm, adjacent time nodes in the discrete time node sequence are selected, and the nonlinear state-space evolution equation is transformed into a system dynamic collocation equality constraint equation according to Simpson's integral rule. By introducing Lagrange multiplier vectors, a Lagrange function is constructed that includes the discretized objective functional, the system dynamic collocation equality constraint equation, and the global inequality constraint function. The first derivative is then obtained to generate the Karoš-Kuhn-Tucker optimality condition system. Newton iteration optimization is performed in the state space, and the approximate value of the Lagrangian function with respect to the Hessian matrix of the decision variable matrix is ​​calculated by the quasi-Newton method to solve the linearized Karosh-Kuhn-Tucker optimality condition system. When the residual value is lower than the convergence tolerance threshold, the optimal motion state trajectory sequence and driving torque sequence are output.

6. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, Step S300 includes: Using a linear hybrid skinning algorithm, the generalized coordinate matrix of the bottom joints in the optimal motion state trajectory sequence is converted into the vertex coordinates of a three-dimensional topological mesh; By combining the preset camera intrinsic matrix, camera extrinsic matrix and ambient lighting parameter tensor, the outgoing radiance tensor of screen pixels is calculated through continuously differentiable rendering equations, and a scene atlas tensor containing color channel attributes, depth channel attributes and surface normal channel attributes is generated. During forward rendering computation, a directed acyclic graph for forward computation is simultaneously constructed in video memory to record the tensor data state and operators during the rendering phase.

7. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, The process of calculating the original visual gradient of the visual loss with respect to the physical boundary parameter set by solving the inverse ordinary differential equation in step S400 includes: Tensor differential multiplication is performed according to the chain rule to calculate the partial derivative of the visual loss scalar value with respect to the system state vector; Construct the adjoint ordinary differential equation in the continuous time domain; The Runge-Kutta numerical integration algorithm is used to solve the adjoint ordinary differential equation in reverse along the time axis, and the sequence of the evolution trajectory of the adjoint state vector in the complete time domain is output. Based on the evolution trajectory sequence of the adjoint state vector, a definite integral operation is performed in the time domain to output the original visual gradient of the visual loss scalar value with respect to the physical boundary parameter set.

8. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, The process of obtaining the Lyapunov energy functional index under the current state and constructing the energy dissipation boundary manifold constraint in step S500 includes: Substituting the system state vector containing instantaneous generalized coordinates and instantaneous generalized velocity into the energy functional calculation formula, the Lyapunov energy functional of the simulated human at the current physical evolution moment is calculated. Establish boundary inequality equations characterizing the passivity of the system, i.e., the time derivative of the system energy functional is less than or equal to the external control input power, defined as the inner product of the control input and the generalized velocity, forming a set of physically constrained manifold surfaces in the physical boundary parameter space; The partial derivatives of the energy functional index with respect to the physical boundary parameter set are obtained. The parameter components are then traversed and concatenated to generate a complete normal vector gradient column vector, which is used to indicate the direction of parameter space evolution that leads to non-physical abrupt changes in the system's energy.

9. The AI-based simulated human creative workflow method for cinemas according to claim 8, characterized in that, The process of performing orthogonal projection on the original visual gradient and outputting the corrected safety physical gradient in step S500 includes: Calculate the projection component of the original visual gradient onto the column vector of the normal gradient; subtract the projection component from the original visual gradient, and calculate the corrected security physical gradient using the following formula. : In the formula, This represents a minimal constant that prevents mathematical singularities caused by a denominator of zero. Represents the original visual gradient vector. Represents the energy functional with respect to the physical boundary parameter set The normal vector and gradient column vector.

10. The AI-based simulated human creative workflow method for cinemas according to claim 1, characterized in that, Step S600 includes: Read the absolute deviation value between the system energy index of the current iteration step and the preset reference benchmark energy scalar, and calculate the adaptive update step size parameter based on the absolute deviation value; Using the calculated safety physical gradient and the adaptive update step size parameter, perform the inverse closed-loop update operation of the physical boundary parameter set; The updated physical boundary parameter set matrix is ​​overwritten to the input memory region of the nonlinear dynamics solver cluster to trigger the underlying logic circuit to execute the next numerical solution loop of the nonlinear state-space evolution equation.

Images