A Human-Computer Collaboration Intent Recognition Method and System Based on Proxy Coupling

Abstract

This invention provides a method and system for human-computer collaborative intent recognition based on proxy coupling, comprising: constructing a virtual centroid proxy model with surface constraints; mapping user physical operations to continuous proxy motion through viscoelastic coupling dynamics to complete real-time encoding of motion intent; discretizing the interactive surface using a triangular mesh to establish a bidirectional mapping between three-dimensional space and two-dimensional parameter space; generating the desired centroid trajectory based on the coupled dynamics of the virtual and real centroids, transforming foot-ground contact behavior into dynamic nonlinear complementary constraints and solving for the optimal contact force; integrating the proxy model and constraints into model predictive control, and generating an exoskeleton adaptive motion trajectory by combining the alternating direction multiplier method and a neurodynamic solver. This invention eliminates the dependence on predefined gait templates, can capture users' time-varying motion intents in real time, improves adaptability to dynamic unstructured environments and the naturalness of human-computer collaboration, reduces computational costs, and achieves deep integration of intent recognition and motion control.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton robot control technology, specifically to an intent recognition method and system based on proxy coupling for autonomous adaptive control of exoskeletons in dynamic environments. Background Technology

[0002] Current mainstream control strategies for exoskeleton robots are based on predefined gait templates. Technicians use preset, fixed gait sequences and joint motion trajectories to plan the exoskeleton's movement patterns, thereby achieving assisted lower limb control for the wearer. To overcome the limitations of fixed gait templates, existing technologies incorporate sensing devices such as electromyography (EMG) sensors, plantar pressure sensors, and inertial measurement units (IMUs). By collecting multimodal information such as the wearer's EMG signals, plantar contact forces, and center of mass (CoM) motion data, they attempt to dynamically adjust the execution parameters of the predefined gait template. Some studies also combine algorithms such as model predictive control (MPC) and proportional-derivative (PD) control to optimize motion trajectories, attempting to improve the exoskeleton's adaptability to the environment. Furthermore, in modeling foot contact behavior, existing technologies generally simplify the mixed characteristics such as contact switching and contact force changes into fixed time series, matching the movement rhythm of the gait template through preset contact logic.

[0003] From the perspective of technical implementation effectiveness, existing exoskeleton control and intent recognition solutions still have many insurmountable limitations, failing to meet the needs of human-machine collaboration in dynamic environments. The core control method based on predefined gait templates is essentially passive control based on fixed motion patterns, unable to actively and in real-time capture the wearer's time-varying intent during movement, such as dynamic adjustments to walking speed, flexible changes in direction, and sudden movements like bending or turning. This directly leads to stiff exoskeleton performance, poor adaptability to dynamic, unstructured environments, and significantly reduces the naturalness of human-machine collaboration. While improved solutions for multimodal sensor data can collect various physical signals related to the wearer's movement, they suffer from complex data fusion logic, high computational costs, and a lack of efficient intent encoding mechanisms. They cannot directly and accurately map discrete sensor data into the wearer's continuous motion intent, easily leading to intent recognition lag and bias, making it difficult to achieve direct linkage between intent and control. The simplified modeling of foot contact behavior ignores the physical changes and mechanical constraints of contact force and step height under different motion states. This leads to a lack of physical feasibility in the motion control of exoskeletons, making them prone to problems such as motion instability and sudden changes in contact force, which affect the safety of human-machine collaboration. At the same time, in existing technologies, the intent recognition module and motion control module are mostly separate, making it impossible to directly embed user intent into the control loop. The lack of a lightweight integration framework makes it difficult to achieve real-time adaptive control of exoskeletons.

[0004] In summary, existing human-machine collaborative intent recognition and control technologies for exoskeleton robots still face four core technical challenges that urgently need to be addressed: First, control strategies rely excessively on predefined gait templates, failing to respond in real-time to the wearer's time-varying motion intentions and exhibiting poor adaptability to dynamic environments. Second, the fusion of multimodal sensor data and intent encoding are inefficient, lacking an effective mechanism to directly map user physical operations into motion intentions. Third, the modeling of foot-ground contact behavior ignores physical characteristics, lacking an effective mechanical constraint mechanism to ensure the physical feasibility of movement. Fourth, the intent recognition and motion control frameworks are separate, failing to achieve deep integration of intent and control, and thus failing to meet the real-time and naturalness requirements of human-machine collaboration.

[0005] Therefore, there is an urgent need in the market for a lightweight, highly adaptive human-computer collaboration intent recognition method that can directly embed user intent into the exoskeleton control loop to achieve template-free real-time adaptive collaborative control. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for human-computer collaboration intent recognition based on proxy coupling.

[0007] A method for human-computer collaboration intent recognition based on proxy coupling according to the present invention includes: Step S1: Construct a virtual centroid proxy model with surface constraints. Define the human-computer interaction surface based on the inertial reference coordinate system. Describe the proxy motion through viscoelastic coupled dynamics and surface normal constraint forces. Map the user's physical operation into the proxy's continuous motion to complete the continuous encoding of motion intent. Step S2: Discretize the interactive surface using a triangular mesh, and establish a bidirectional mapping between the three-dimensional space and the two-dimensional parameter space through the projection matrix and lifting matrix of each patch, so as to realize the mapping and motion control of the surrogate position in the parameter space; Step S3: Based on the coupled dynamic relationship between the virtual centroid agent and the real centroid of the exoskeleton, generate the expected centroid trajectory that encodes the user's motion intention in real time; Step S4: Transform the foot-ground contact behavior into a dynamic nonlinear complementary constraint. Construct an error vector model using nonlinear complementary functions and KKT conditions. Transform the optimization solution into a dynamic evolution process of ordinary differential equations to obtain the optimal contact force solution. Step S5: Integrate the virtual centroid surrogate model and dynamic nonlinear complementary constraints into the model predictive control (MPC), construct a multi-task weighted objective function, and solve to generate the exoskeleton adaptive motion trajectory.

[0008] Preferably, the virtual centroid surrogate dynamic equation is:

[0009] in, Here is the damping matrix. For proxy position, To control the force, it is connected to the real center of mass through viscoelastic coupling. This represents a virtual force applied from the outside. Represents the surface normal vector The constraint forces acting, the superscript v indicates the virtual force calculated in the model, and the subscript... , represents the virtual viscoelastic attraction / control force generated by the relative motion between the external environment and the agent and the real centroid of the exoskeleton, respectively; n is the normal vector of the interactive surface.

[0010] Preferably, the interactive surface is a three-dimensional freeform surface or a polyhedral surface, discretized into... vertices and A triangular facet; The projection matrix Used to map three-dimensional Cartesian space coordinates to a local two-dimensional parameter space; The lifting matrix Used to reverse the updated two-dimensional parametric space coordinates to three-dimensional space.

[0011] Preferably, the foot-to-ground contact satisfies a nonlinear complementary constraint:

[0012] in Let i be the step height of the i-th foot. Let i be the position vector of the i-th contact point. Let be the force vector of the i-th foot in contact with the ground.

[0013] Preferably, the MPC objective function in step S5 includes weighted penalty terms for centroid position tracking, centroid linear velocity tracking, foot yaw angle, fuselage attitude, control input smoothing, and leg lift height constraint. The Alternating Direction Multiplier Method (ADMM) is used to decompose the nonlinear model predictive control (MPC) optimization problem into parallel subproblems. Combined with a neurodynamic solver, the KKT conditions are transformed into ordinary differential equations to achieve optimal solution calculation.

[0014] According to the present invention, a human-computer collaborative intent recognition system based on agent coupling is provided for implementing the method, comprising a sensor module, an agent management module, a surface processing module, an intent fusion module, and a control execution module; The sensor module consists of an exoskeleton back inertial measurement unit and a plantar pressure sensor, which collects real-time data on the center of mass and plantar contact force. The agent management module constructs and maintains a virtual centroid agent model, calculates viscoelastic coupling force, and updates the agent dynamic state. The surface processing module completes the discretization of the interactive surface triangular mesh and performs bidirectional mapping calculation between the three-dimensional Cartesian space and the two-dimensional parameter space; The intent fusion module integrates agent output and sensor data, processes dynamic nonlinear complementary constraints, and generates a desired centroid trajectory containing contact information. The control execution module runs the agent model and MPC algorithm, and drives the exoskeleton joint motors through EtherCAT communication.

[0015] Preferably, the agent management module has built-in damping matrix and stiffness matrix configuration units. Based on the position and velocity deviation between the real center of mass and the virtual agent, it calculates the virtual viscoelastic control force in real time and superimposes the surface normal constraint force to update the agent's motion state.

[0016] Preferably, the surface treatment module is adapted to the interactive curved surface of the arc-shaped industrial exoskeleton and the flat rehabilitation exoskeleton, stores the projection matrix and lifting matrix of each triangular facet, and completes the real-time mapping of three-dimensional to two-dimensional coordinates and boundary trajectory adjustment.

[0017] Preferably, the intent fusion module has a built-in dynamic nonlinear complementary constraint solving unit, configured with a relaxation factor and a Fischer-Burmeister function, to transform the foot-ground contact constraint into a dynamic system and quickly solve for the optimal contact force.

[0018] Preferably, the control execution module integrates an ADMM decomposition unit and a neural dynamics solver, transforming the MPC optimization problem into a parallel quadratic programming subproblem, and achieving trajectory solution through dynamic evolution of ordinary differential equations.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs a virtual centroid proxy model, surface constraints, and viscoelastic coupling mechanism to directly and dynamically capture and encode the user's time-varying motion intentions, completely eliminating the dependence on predefined gait templates, enabling the exoskeleton to autonomously adapt to the dynamic unstructured environment, and significantly improving the naturalness of human-machine collaboration.

[0020] 2. This invention uses triangular mesh to discretize the interactive surface, establishes a bidirectional mapping between three-dimensional space and two-dimensional parameter space, realizes efficient calculation of proxy position and continuous trajectory update, simplifies the motion processing logic of complex curved surfaces, reduces computational costs, and ensures the real-time performance of intent recognition.

[0021] 3. This invention integrates the proxy model and contact complementary constraints into the MPC framework, and combines the efficient surface processing of triangular mesh discretization with the fast calculation of the ADMM solver to achieve deep integration of intent recognition and motion control, achieving millisecond-level response and adapting to real-time adaptive cooperative control in dynamic unstructured environments. Attached Figure Description

[0022] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the shared autonomous control framework described in this invention, showing the integrated structure of the Intent Recognition Module (ICC) and the MPC controller; Figure 2 This is a schematic diagram of the human-computer collaborative intent recognition system based on agent coupling according to the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0024] This invention can significantly enhance the adaptability of exoskeletons and is suitable for scenarios such as industrial handling and rehabilitation training.

[0025] Example 1 According to the present invention, a human-computer collaborative intent recognition method based on agent coupling dynamically captures user intent through a virtual CoM agent model and combines complementary constraints to ensure motion stability. Specifically, the method includes: Step S1: Construct a virtual center of mass (CoM) surrogate model. The coordinate system is defined as fixed in the inertial frame. The reference coordinate system below Interactive surfaces For a manifold embedded in a three-dimensional Cartesian space, the surrogate position and the actual center of mass position of the exoskeleton All are defined in three-dimensional space In the middle, the agent's initial position Set to be on the interactive surface (triangular mesh facet) At a point on the , the three-dimensional coordinates are mapped to initial coordinates in the two-dimensional parameter space via a projection matrix. The coupling force is determined by a three-dimensional linear damping matrix. and stiffness matrix The system dynamically captures the user's motion intentions through surface constraints and viscoelastic coupling. The agent's position is constrained by an interaction surface (such as the curved surface of an exoskeleton's back), and its motion is described by differential equations. The interaction surface refers to a free-form surface or polyhedral surface constructed in three-dimensional space, serving as the human-computer interaction manifold. In specific implementation scenarios, for example, an industrial exoskeleton can be configured as an arc-shaped surface, while a rehabilitation exoskeleton can be configured as a flat surface. The agent dynamics are expressed as:

[0026] in, Here is the damping matrix. For proxy position, To control the force, it is connected to the real center of mass through viscoelastic coupling. This represents a virtual force applied from the outside. Represents the surface normal vector The constraint forces acting, the superscript v indicates the virtual force calculated in the model, and the subscript... , These represent the virtual viscoelastic attraction / control force generated by the relative motion between the external environment and the agent's and the exoskeleton's real center of mass, respectively; surface constraints are represented by normal vectors. Ensure the agent remains on the surface at all times. This step utilizes the viscoelastic connections of surface constraints to map user actions to agent motion, enabling continuous encoding of intent.

[0027] Step S2: Discretize the interactive surface using a triangular mesh to establish a bidirectional mapping between 3D space and 2D parameter space, enabling the mapping and motion control of the proxy position in the parameter space. Specifically, the interactive surface is divided into triangular meshes, and each facet is mapped using a projection matrix. and boosting matrix This achieves the mapping between three-dimensional coordinates and two-dimensional parameter space. The projection matrix is ​​involved. Used to map three-dimensional Cartesian space coordinates to a local two-dimensional parameter space; lifting matrix This is used to reverse-reconstruct the updated two-dimensional parametric space coordinates into three-dimensional space. It discretizes the three-dimensional polyhedral surface into... vertices and Each triangular facet. Defined by three vertices, each vertex corresponds to unique 3D spatial coordinates and 2D parametric spatial coordinates, with bidirectional mapping achieved through a specific constant vector and projection / lifting matrix for each facet. Proxy motion achieves continuous trajectory updates through ray intersection and boundary handling in the parametric space.

[0028] Step S3: Based on the coupled dynamics of the agent and the real centroid, generate the expected CoM trajectory in real time and encode the user intent; Step S4: Combining dynamic nonlinear complementary conditions, the foot-ground contact behavior is expressed as complementary constraints to ensure physical feasibility. First, the nonlinear complementary problem (NCP) consisting of foot-ground contact force and step height is transformed into a dynamic nonlinear complementary system (DNCS). Then, relaxed Fischer-Burmeister functions and other nonlinear complementary problem functions are introduced to transform the complementary constraints with non-smooth points into equality and inequality equations, and KKT conditions are established. Finally, a unified error vector model is constructed based on the KKT conditions, transforming the optimization solution process into a dynamic evolution process of ordinary differential equations (i.e.,...). The optimal contact force solution is obtained through dynamic system convergence.

[0029] Step S5: Integrate the surrogate model and complementary constraints into Model Predictive Control (MPC) to generate an adaptive motion trajectory. The MPC objective function is defined as follows: The integral term includes weighted penalties for multiple underlying motion control tasks, specifically: tracking the contact point centroid position of the desired landing point, tracking the CoM linear velocity of the desired centroid velocity, the foot yaw angle task, the fuselage attitude orientation task, the force and joint regularization task for smoothing control inputs, and the swing height task for constraining leg lift height. The total prediction time domain length is set to... The value is 1.0s, and it is discretized into... It consists of 20 control steps.

[0030] Furthermore, in conjunction with the appendix Figures 1 to 2 The specific steps for dynamic motion optimization and multi-task planning of the exoskeleton of the present invention are described below: First, the proxy model is initialized by defining the interaction surface (such as the curved surface of the exoskeleton's back) and setting the initial position of the virtual CoM proxy. The surface is discretized using a triangular mesh, and each facet is associated with a projection matrix. and boosting matrix .

[0031] Then, intent capture occurs when the user expresses an intent through physical interaction, and the agent is located. Updated based on the kinetic equations:

[0032] Among them, control force Calculated from the deviation between the true centroid and the proxy:

[0033] in, Represents a three-dimensional linear damping matrix. Represents the three-dimensional linear stiffness matrix. , These represent the three-dimensional linear velocity of the actual centroid of the exoskeleton and the three-dimensional linear velocity of the virtual centroid proxy, respectively. , These represent the three-dimensional position coordinates of the actual centroid of the exoskeleton and the three-dimensional position coordinates of the virtual centroid proxy, respectively.

[0034] Next, surface motion processing is performed. The agent moves along the velocity direction in the parameter space and detects the surface boundaries by ray intersection. If the agent touches the boundary, its trajectory is adjusted along the edge vector to ensure continuity and surface constraints.

[0035] Next, complementary constraint application: foot contact force With BBK Linked through dynamic complementary conditions:

[0036]

[0037] Where h() represents the step height function of the i-th foot. This represents the position vector of the i-th contact point. Represents a nonlinear complementary operator. The transpose vector representing the normal direction of the contact plane. Let represent the ground contact force vector of the i-th foot. A relaxation factor is used. To avoid numerical singularities, the preferred value is And it is optimized using a neural network solver.

[0038] Finally, the MPC integration uses the agent output as the reference trajectory for MPC, combined with task objectives (such as CoM tracking and foot placement), to solve for the optimal motion. The computation is accelerated using ADMM and a neurodynamic solver. First, the large nonlinear model predictive control (MPC) optimization problem is reconstructed into a consistent form using the Alternating Direction Multiplier Method (ADMM), which decomposes it into multiple parallel subproblem update steps (sequentially updating variable z, auxiliary variables...). and dual variables Within each iteration of ADMM, a local quadratic programming (QP) problem is generated. A neurodynamic solver is then introduced to transform the KKT conditions of this optimization problem into a dynamic system of ordinary differential equations containing an error vector. By simulating the evolution of the dynamic system, the optimal solution for the contact force is quickly obtained, significantly reducing single-step computation latency and achieving millisecond-level response.

[0039] Furthermore, the overall control architecture and implementation process of the agent-coupled human-computer collaboration intent recognition method of the present invention are described in detail below with reference to Figure 1: This invention deeply integrates the intent recognition module with the Model Predictive Control (MPC) controller to construct a human-machine collaborative and shared autonomous control framework. This framework corresponds to steps S1-S5 in Example 1, realizing a fully closed-loop operation of real-time intent encoding, spatial mapping, trajectory generation, contact constraint solving, and adaptive motion control, as detailed below: First, the top-level intent perception and encoding layer corresponds to the human-exoskeleton collaboration module in Figure 1. This layer corresponds to steps S1 and S2 in Embodiment 1, and its core function is to capture and continuously encode the user's motion intent in real time, serving as the input source for the entire control framework.

[0040] Based on the inertial reference coordinate system, a human-computer interaction surface is defined, and a virtual centroid (CoM) proxy model with surface constraints is constructed. The proxy motion is described by viscoelastic coupled dynamics and surface normal constraint forces. Through this model, the user's physical operations are directly mapped to the proxy's continuous motion, completing the real-time encoding of motion intentions and completely eliminating the dependence on predefined gait templates.

[0041] Triangular meshes are used to discretize the interactive surface. By using the projection matrix and lifting matrix of each patch, a bidirectional mapping between three-dimensional Cartesian space and two-dimensional parameter space is established, which realizes parameter space calculation of surrogate position and continuous trajectory update, reducing the computational cost of complex surface motion processing.

[0042] The final output of this layer encodes the user's desired motion state based on their motion intent. (such as the desired center of mass position and velocity trajectory), serving as the core reference command for the downstream controller.

[0043] Secondly, the bottom-level planning and optimization layer corresponds to the dynamic nonlinear complementary conditions, walking constraints and task planning, nonlinear model predictive control, and unified nonlinear optimization solver in Figure 1. This layer corresponds to steps S3, S4, and S5 in Example 1, and its core functions include generating the desired trajectory, solving the foot-ground contact constraints, and performing MPC optimization calculations, providing optimal control input for physical execution.

[0044] Based on the coupled dynamic relationship between the virtual centroid agent and the real centroid of the exoskeleton, the expected centroid trajectory encoding the user's motion intention is generated in real time and used as the tracking target of MPC (corresponding to step S3 in embodiment 1).

[0045] The foot-ground contact behavior is transformed into a dynamic nonlinear complementary constraint. A relaxation factor and a Fischer-Burmeister function are introduced, and an error vector model is constructed in combination with KKT conditions. The optimization solution is transformed into a dynamic evolution process of ordinary differential equations to obtain the optimal foot-ground contact force solution and ensure the physical feasibility of the motion (corresponding to step S4 in Example 1).

[0046] A multi-task weighted objective function is constructed, which includes weighted penalty terms for centroid position tracking, centroid linear velocity tracking, foot yaw angle, fuselage attitude, control input smoothing, and leg lift height constraint. The virtual centroid surrogate model and dynamic nonlinear complementary constraints are integrated into the MPC framework to generate a control sequence containing the contact force state of each prediction step in the prediction time domain.

[0047] The nonlinear MPC optimization problem is decomposed into parallel quadratic programming subproblems using the alternating direction multiplier method (ADMM). The KKT conditions are transformed into ordinary differential equations by combining a neurodynamic solver. The solution is solved quickly through dynamic system convergence, and the optimal foot contact force solution for the current control cycle is output (corresponding to step S5 in Example 1).

[0048] Finally, the intermediate physical execution and state feedback layer corresponds to the PD control + Jacobian module and the centroid dynamics module in Figure 1. This layer is the physical execution unit of the control framework, receiving the upper-layer intent commands and the lower-layer optimization results to complete the exoskeleton body driving and state closed-loop feedback.

[0049] Receive the desired centroid trajectory from the top-level output and the optimal foot contact force solution from the bottom-level solver. The system calculates joint space feedback torques using a proportional-derivative (PD) control algorithm, and then transforms task space commands into exoskeleton joint driving torques using Jacobian matrix mapping. These joint torques act on the centroid dynamics module, driving the exoskeleton to complete physical motion evolution and outputting the exoskeleton's actual physical state in real time. This actual physical state is fed back to the top-level human-exoskeleton collaboration module, correcting the virtual centroid proxy motion and forming an intent recognition closed loop. Simultaneously, it is passed down to the bottom-level nonlinear model predictive control module, providing a real-time state reference for the MPC, thus completing the fully closed-loop adaptive collaboration between intent recognition and motion control.

[0050] Based on the collaborative operation of the above three-layer architecture, this invention achieves millisecond-level response of intent recognition and motion control through the S1-S5 method of Example 1, thereby improving the adaptability of the exoskeleton in dynamic unstructured environments and the naturalness of human-machine collaboration.

[0051] Example 2 The present invention also provides a human-computer collaboration intent recognition system based on proxy coupling. The human-computer collaboration intent recognition system based on proxy coupling can be implemented by executing the process steps of the human-computer collaboration intent recognition method based on proxy coupling. That is, those skilled in the art can understand the human-computer collaboration intent recognition method based on proxy coupling as a preferred embodiment of the human-computer collaboration intent recognition system based on proxy coupling.

[0052] According to the present invention, a human-computer collaborative intent recognition system based on agent coupling is provided, such as... Figure 2As shown, it includes: Sensor modules: An inertial measurement unit (IMU) is installed on the back of the exoskeleton to collect CoM data in real time; a pressure sensor is embedded in the sole of the foot to monitor contact force.

[0053] The agent management module is used to build and maintain the virtual CoM agent model, calculate the viscoelastic coupling force, and update the agent's dynamic state. This module incorporates damping and stiffness matrix configuration units. Based on the position and velocity deviation between the real center of mass and the virtual agent, it calculates the virtual viscoelastic control force in real time and superimposes the surface normal constraint force to update the agent's motion state. This module constructs a three-dimensional freeform surface as the human-computer interaction manifold, establishing a viscoelastic coupling model between the virtual center of mass agent and the actual center of mass of the exoskeleton, and capturing the wearer's movement intentions in real time.

[0054] Surface processing module: Responsible for discretizing the triangular mesh of the interactive surface and calculating the bidirectional mapping between 3D Cartesian space and 2D parameter space. This module adapts to the interactive surfaces of curved industrial exoskeletons and flat rehabilitation exoskeletons, stores the projection and lifting matrices of each triangular facet, and performs real-time mapping from 3D to 2D coordinates and boundary trajectory adjustment.

[0055] The intent fusion module integrates agent output and sensor data, processes dynamic nonlinear complementary constraints, and generates the desired CoM trajectory containing contact information. This module incorporates a dynamic nonlinear complementary constraint solving unit, configured with relaxation factors and the Fischer-Burmeister function, to transform foot-ground contact constraints into a dynamic system and quickly solve for the optimal contact force.

[0056] Control Execution Module: The embedded system runs the proxy model and MPC algorithm, driving the joint motors via EtherCAT communication. This module integrates ADMM decomposition units and a neurodynamics solver, transforming the MPC optimization problem into a parallel quadratic programming subproblem, and achieving trajectory solving through dynamic evolution of ordinary differential equations.

[0057] In testing, the exoskeleton was able to adapt to the user's bending and turning intentions in real time, reducing energy consumption by 20% and improving the naturalness of movement.

[0058] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0059] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for human-computer collaboration intent recognition based on proxy coupling, characterized in that, include: Step S1: Construct a virtual centroid proxy model with surface constraints. Define the human-computer interaction surface based on the inertial reference coordinate system. Describe the proxy motion through viscoelastic coupled dynamics and surface normal constraint forces. Map the user's physical operation into the proxy's continuous motion to complete the continuous encoding of motion intent. Step S2: Discretize the interactive surface using a triangular mesh, and establish a bidirectional mapping between the three-dimensional space and the two-dimensional parameter space through the projection matrix and lifting matrix of each patch, so as to realize the mapping and motion control of the surrogate position in the parameter space; Step S3: Based on the coupled dynamic relationship between the virtual centroid agent and the real centroid of the exoskeleton, generate the expected centroid trajectory that encodes the user's motion intention in real time; Step S4: Transform the foot-ground contact behavior into a dynamic nonlinear complementary constraint. Construct an error vector model using nonlinear complementary functions and KKT conditions. Transform the optimization solution into a dynamic evolution process of ordinary differential equations to obtain the optimal contact force solution. Step S5: Integrate the virtual centroid surrogate model and dynamic nonlinear complementary constraints into the model predictive control (MPC), construct a multi-task weighted objective function, and solve to generate the exoskeleton adaptive motion trajectory.

2. The human-computer collaboration intent recognition method based on proxy coupling according to claim 1, characterized in that, The dynamic equations of the virtual centroid proxy model are as follows: in, Here is the damping matrix. For proxy position, To control the force, it is connected to the real center of mass through viscoelastic coupling. This represents a virtual force applied from the outside. Represents the surface normal vector The constraint forces acting, the superscript v indicates the virtual force calculated in the model, and the subscript... , represents the virtual viscoelastic attraction / control force generated by the relative motion between the external environment and the agent and the real centroid of the exoskeleton, respectively; n is the normal vector of the interactive surface.

3. The human-computer collaboration intent recognition method based on proxy coupling according to claim 1, characterized in that, The interactive surface is a three-dimensional freeform surface or a polyhedral surface, discretized into... vertices and A triangular facet; The projection matrix Used to map three-dimensional Cartesian space coordinates to a local two-dimensional parameter space; The lifting matrix Used to reverse the updated two-dimensional parametric space coordinates to three-dimensional space.

4. The human-computer collaboration intent recognition method based on proxy coupling according to claim 1, characterized in that, The foot-to-ground contact satisfies a nonlinear complementary constraint: in Let i be the step height of the i-th foot. Let i be the position vector of the i-th contact point. Let be the force vector of the i-th foot in contact with the ground. Represents a nonlinear complementary operator. This represents the transpose vector of the contact plane normal direction.

5. The human-computer collaboration intent recognition method based on proxy coupling according to claim 1, characterized in that, In step S5, the virtual centroid proxy model and dynamic nonlinear complementary constraints are integrated into the model predictive control (MPC). The constructed multi-task weighted objective function includes weighted penalty terms for centroid position tracking, centroid linear velocity tracking, foot yaw angle, fuselage attitude, control input smoothing, and leg lift height constraints. The Alternating Direction Multiplier Method (ADMM) is used to decompose the nonlinear model predictive control (MPC) optimization problem into parallel subproblems. Combined with a neurodynamic solver, the KKT conditions are transformed into ordinary differential equations to achieve optimal solution calculation.

6. A human-computer collaborative intent recognition system based on proxy coupling, characterized in that, The method for implementing any one of claims 1-5 includes a sensor module, an agent management module, a surface processing module, an intent fusion module, and a control execution module; The sensor module consists of an exoskeleton back inertial measurement unit and a plantar pressure sensor, which collects real-time data on the center of mass and plantar contact force. The agent management module constructs and maintains a virtual centroid agent model, calculates viscoelastic coupling force, and updates the agent dynamic state. The surface processing module completes the discretization of the interactive surface triangular mesh and performs bidirectional mapping calculation between the three-dimensional Cartesian space and the two-dimensional parameter space; The intent fusion module integrates agent output and sensor data, processes dynamic nonlinear complementary constraints, and generates a desired centroid trajectory containing contact information. The control execution module runs the agent model and MPC algorithm, and drives the exoskeleton joint motors through EtherCAT communication.

7. The human-computer collaborative intent recognition system based on proxy coupling according to claim 6, characterized in that, The agent management module has built-in damping matrix and stiffness matrix configuration units. Based on the position and velocity deviation between the real center of mass and the virtual agent, it calculates the virtual viscoelastic control force in real time and superimposes the surface normal constraint force to update the agent's motion state.

8. The human-computer collaborative intent recognition system based on proxy coupling according to claim 6, characterized in that, The surface treatment module is adapted to the interactive curved surfaces of arc-shaped industrial exoskeletons and flat rehabilitation exoskeletons, stores the projection matrix and lifting matrix of each triangular facet, and completes real-time mapping of three-dimensional to two-dimensional coordinates and boundary trajectory adjustment.

9. The human-computer collaborative intent recognition system based on proxy coupling according to claim 6, characterized in that, The intent fusion module has a built-in dynamic nonlinear complementary constraint solving unit, which is configured with relaxation factors and Fischer-Burmeister functions to transform foot-ground contact constraints into a dynamic system and quickly solve for the optimal contact force.

10. The human-computer collaborative intent recognition system based on proxy coupling according to claim 6, characterized in that, The control execution module integrates ADMM decomposition units and a neurodynamics solver, transforming the MPC optimization problem into a parallel quadratic programming subproblem, and achieving trajectory solving through dynamic evolution of ordinary differential equations.

Images