Double-station robot sorting optimization method, system and terminal based on digital twinning

Abstract

The invention discloses a double-station robot sorting optimization method and system based on digital twinning and a terminal, double improvement of sorting efficiency and safety is achieved by constructing a digital twinning driven collaborative operation system, and the method comprises the steps that firstly, a laser radar and a polarization camera are used for forming a composite sensing unit; geometric morphology, material reflection characteristics and spatial pose data of a target object are synchronously obtained, and are input into a digital twin engine after time synchronization processing; an engine constructs a virtual sorting scene containing a material attribute database based on a physical rendering technology, sub-millimeter-level space registration is achieved through feature fusion of a binocular stereoscopic vision depth map and geometric parameters, high-precision digital twin stations are generated, a system plans a space-time constraint trajectory of a double-station robot in a virtual environment, and a target object is obtained. And an integrated discrete event simulation engine performs operation time sequence conflict prediction, and when a space overlapping risk is detected, an alternative path containing a dynamic obstacle avoidance strategy is generated through a trajectory re-planning algorithm.

Description

Technical Field

[0001] This invention relates to the field of robot sorting optimization, and in particular to a method, system and terminal for dual-station robot sorting optimization based on digital twins. Background Technology

[0002] In the process of automation and intelligent transformation of manufacturing, high-precision clean production environments place stringent requirements on material sorting and palletizing operations. Fields such as electronic components, biomedicine, and semiconductor manufacturing generally have cleanliness control needs. Traditional manual operations are no longer able to meet the needs of modern industry due to efficiency bottlenecks, risks of human contamination, and high labor costs. With the mature application of industrial robot technology, automated sorting systems based on robotic arms are gradually replacing manual labor. However, existing single-station robot sorting solutions still face problems such as limited operating space and insufficient dynamic response in complex clean environments. In particular, when it comes to dual-station collaborative operation scenarios, it is necessary to take into account the spatial isolation characteristics of the clean panel separation area and the protective board palletizing area, which poses a challenge to the collaborative control accuracy and dynamic adaptability of the robot system.

[0003] Traditional dual-station robot sorting systems generally adopt a preset program control mode, where path planning and operation sequence are determined through offline programming. This makes it difficult to cope with real-time dynamic changes in the production site. When panel size fluctuates or environmental parameters deviate, the system requires manual intervention for path replanning, with a single debugging cycle lasting several hours, leading to extended production downtime. When multiple robots work together, traditional solutions lack effective conflict prediction mechanisms. Potential risks such as path intersections and robotic arm collisions cannot be identified in advance and can only be resolved through post-event intervention. This passive response mode not only reduces equipment utilization but also increases equipment maintenance costs. Furthermore, the static program control architecture cannot achieve joint optimization of sorting sequence and workspace, resulting in the overall efficiency of the dual-station system being limited by the bottleneck of a single link, making it difficult to adapt to the needs of flexible manufacturing. To address this, a dual-station robot sorting optimization method, system, and terminal based on digital twins are proposed. Summary of the Invention

[0004] This invention provides the following technical solution: a sorting optimization method for dual-station robots based on digital twins, comprising the following steps: S1. Multimodal environmental parameter acquisition: By integrating a composite sensing unit with lidar and polarization camera, the geometric contour data, surface reflection characteristics and spatial pose information of the clean panel and protective cover are acquired simultaneously. The data is then input into the digital twin engine after time synchronization processing. S2, Dynamic Digital Twin Modeling: Based on the collected environmental parameters, a physical rendering engine is used to construct a virtual sorting scene. The scene includes a material property database and a flexible object simulation module to support real-time updates of object deformation state. S3, Multi-view Data Fusion Calibration: The depth map acquired by the binocular stereo vision system is fused with the geometric parameters collected in S1 at the feature level, and sub-millimeter-level spatial registration is achieved through a nonlinear optimization algorithm to generate a high-precision digital twin workstation. S4. Spatiotemporal Constraint Trajectory Planning: Generate the cooperative motion trajectory of the dual-station robot in the digital twin workstation. The trajectory includes a timestamp sequence, joint space interpolation points and velocity constraints, and meets the obstacle avoidance requirements in C-space. S5. Real-time conflict detection and avoidance: A collision detection engine based on discrete event simulation is used to predict timing conflicts in dual-station operations. When spatial overlap risk is detected, alternative paths are generated through a trajectory replanning algorithm, which includes a dynamic obstacle avoidance strategy. S6. Adaptive control parameter generation: Based on the collision detection results, the robot's motion speed, acceleration, and buffer parameters are adjusted in real time using a fuzzy PID controller to generate a smooth velocity profile curve. S7, Closed-loop iterative optimization execution: The actual execution trajectory is compared with the trajectory predicted by the digital twin model, and the control command is continuously corrected by the model predictive control algorithm, forming a closed-loop optimization mechanism that includes environmental disturbance compensation.

[0005] Multimodal environmental parameter acquisition: By integrating a LiDAR and a polarization camera into a composite sensing unit, the following operations are performed simultaneously: LiDAR is used to acquire the geometric contour data and spatial pose information of the clean panel and protective sleeve, forming a three-dimensional point cloud dataset; The surface reflection characteristics of the target object are collected using a polarization camera, including polarization angle, reflectivity, and texture features. The above multimodal data is processed for time synchronization to eliminate sampling delay deviations between different sensors; The processed environmental parameters are input into the digital twin engine as the basic data source for building the virtual scene; Dynamic digital twin modeling: Based on the collected environmental parameters, the following modeling operations are performed: A physically based rendering engine is invoked to construct a virtual sorting scene containing a database of material properties, which covers the elastic modulus, friction coefficient, and deformation threshold of different materials. A flexible object simulation module is deployed to update the deformation state of the object in real time through finite element analysis, simulating the non-rigid deformation process of the protective plate under mechanical gripping. Establish a two-way mapping relationship between virtual scenes and physical entities, and support real-time synchronous updates of dynamic parameters; Multi-view data fusion calibration The depth map acquired by the binocular stereo vision system is fused and calibrated with the S1 acquisition data: Denoising is performed on the binocular vision depth map, and edge contour feature points are extracted; The lidar point cloud data and visual feature points are fused at the feature level to construct a three-dimensional spatial coordinate correlation matrix. By iteratively solving the spatial transformation parameters using a nonlinear optimization algorithm, cross-modal data registration with sub-millimeter accuracy is achieved. Generate a high-precision digital twin workstation model to provide a geometric constraint benchmark for trajectory planning; Generate collaborative trajectories for dual-workstation robots in a digital twin workstation: Initial candidate paths are generated based on the improved RRT* algorithm to meet the obstacle avoidance requirements in C-space. The candidate paths are timestamped to construct a kinematic model that includes joint space interpolation points; Apply velocity constraints, optimize trajectory smoothness using gradient projection method, and generate motion commands that meet dynamic constraints; Real-time conflict detection and avoidance Deploy conflict detection and avoidance mechanisms: Construct a conflict detection engine based on discrete event simulation and establish a dual-station operation timing model; Real-time monitoring of the spatial occupancy status of the robotic arm's end effector to predict potential operational conflicts; When spatial overlap risk is detected, the trajectory replanning algorithm is activated to generate alternative paths that include dynamic obstacle avoidance strategies. Conflict avoidance is achieved through local path correction, ensuring the continuity of collaborative work between two workstations; Adaptive control parameter generation Adaptive control is executed based on the collision detection results: The current operation status is analyzed by a fuzzy PID controller, and a fuzzy rule base for speed, acceleration, and buffer parameters is established. The PID control parameters are dynamically adjusted based on real-time monitoring of changes in the robotic arm's load. Generate velocity profile curves that satisfy smoothness constraints to eliminate vibration phenomena caused by abrupt trajectory changes; Closed-loop iterative optimization execution: Construct a closed-loop optimization mechanism: The actual execution trajectory is mapped to the digital twin space using a positive kinematics model; By comparing the deviation between the actual trajectory and the predicted trajectory, the compensation amount is calculated using a model predictive control algorithm. The compensation amount is injected into the control command generation module to achieve environmental disturbance compensation; Continuously iterate and optimize control commands until the operation accuracy meets the preset threshold requirements.

[0006] Preferably, in the multimodal environmental parameter acquisition, the lidar adopts the phase ranging principle, and the polarization camera is equipped with a tunable filter array to separate the specular reflection component in the surface reflected light.

[0007] LiDAR calculates the spatial distance between a target object and a sensor based on the principle of phase ranging by emitting modulated laser signals and detecting the phase difference of the reflected signals. A polarization camera analyzes the polarization state of incident light using a tunable filter array, separating the specular reflection component and the diffuse reflection component in the surface reflected light. The separated specular reflection component data is spatially registered with the lidar ranging results to generate a 3D point cloud model with fused environmental parameters.

[0008] Preferably, the spatiotemporal constraint trajectory planning further includes singularity analysis of the tool center point trajectory, and avoids the robot from entering a kinematically singular configuration through condition number evaluation.

[0009] Based on the robot's kinematic model, spatiotemporal constraints on the trajectory of the tool's center point are constructed; Calculate the condition number of the Jacobian matrix for the planned trajectory and evaluate the kinematic singularity of the current configuration; When the number of conditions exceeds the preset threshold, the robot is prevented from entering the strange configuration region by adjusting the trajectory parameters or inserting intermediate path points.

[0010] Preferably, the material property database includes the elastic modulus, Poisson's ratio, and surface friction coefficient of the clean panel, and the flexible object simulation module uses a mass-spring model to simulate deformation, with the spring stiffness coefficient dynamically adjusted based on the real-time collected contact force data.

[0011] The parameters such as the elastic modulus, Poisson's ratio, and surface friction coefficient of the cleanroom panel are retrieved from the material property database. A hybrid geometric-physical representation structure for flexible objects is constructed based on a mass-spring model, where each mass is connected to adjacent mass points via springs. Contact force data is collected in real time, and the spring stiffness coefficient is dynamically adjusted according to Hooke's Law to update the deformation state of the flexible object.

[0012] Preferably, the dynamic obstacle avoidance strategy uses the speed obstacle method to calculate alternative paths and introduces a safe distance threshold to adjust the avoidance priority. When a spatial overlap risk is detected, trajectory replanning of low-priority workstations is triggered first.

[0013] The velocity obstacle method is used to calculate the potential collision time window and the set of alternative avoidance paths between the robot and dynamic obstacles. The alternative paths are prioritized according to a preset safe distance threshold, and an avoidance priority queue is generated. When a risk of spatial overlap is detected, trajectory replanning of low-priority workstations is triggered first, releasing high-priority task channels.

[0014] The digital twin-based dual-station robot sorting optimization system employs the aforementioned digital twin-based dual-station robot sorting optimization method, including: Environmental sensing array: Composed of a laser profile scanner, a six-dimensional force sensor, an ultrasonic ranging module, and a particle counter, it is used to achieve full-dimensional perception and cleanliness monitoring of the sorting environment; Digital Twin Engine: It integrates a rigid-flexible coupled multibody dynamics solver and a finite element analysis module to support real-time collision detection, response, and structural deformation simulation; Intelligent Decision-Making Center: Deploy deep reinforcement learning networks and rule engines to autonomously optimize control strategies based on historical sorting data, and load expert system knowledge bases for logical reasoning; Safety enforcement agencies: The servo motor and electronic cam system with dual encoder redundancy design have torque feedforward control function; Human-computer interaction terminal: Equipped with a gesture recognition module, a 3D projection device, and an audible and visual alarm unit, it supports the visual debugging and abnormal status prompts of virtual tooling.

[0015] Environmental sensing and data acquisition: The sorting area is scanned in three dimensions by a laser contour scanner in the environmental perception array to obtain the geometric dimensions, spatial position and orientation parameters of the objects to be sorted. The six-dimensional force sensor monitors the contact force and torque data between the end effector of the robotic arm and the object in real time. Combined with the spatial positioning information output by the ultrasonic ranging module, a dynamic interaction model between the object and the actuator is constructed. The particle counter continuously monitors the cleanliness indicators of the sorting environment, and triggers a cleanliness compensation process when the concentration of suspended particles exceeds a preset threshold. All sensing data is transmitted in real time to the digital twin engine via industrial Ethernet, forming the initial data mapping between the physical entity and the virtual model; Digital twin modeling and simulation verification: The digital twin engine constructs a three-dimensional coupled model of the robotic arm, object, and environment based on received real-time data through a rigid-flexible coupled multibody dynamics solver. The finite element analysis module performs stress-strain simulations on key structural components to predict deformation and fatigue life under high load conditions. Conduct virtual collision detection: Simulate the movement trajectory of a robotic arm in digital space to identify potential interference risks with fixed facilities such as conveyor belts and guardrails in advance; A virtual verification report containing kinematic constraints, dynamic parameters, and safety boundaries is generated and fed back to the intelligent decision-making center. Intelligent decision-making and strategy optimization: The deep reinforcement learning network generates a set of candidate control policies based on historical sorting data and current environmental perception information using the Q-learning algorithm; The rules engine loads the expert system knowledge base, performs logical verification on candidate strategies, and excludes operation sequences that violate safety specifications or process requirements; Based on the simulation results of the digital twin engine, a multi-objective optimization algorithm (such as NSGA-II) is used to perform weighted evaluation of indicators such as sorting efficiency, energy consumption, and mechanical wear. The optimal control command is output to the safety execution mechanism, and the optimized strategy parameters are stored in the knowledge base for subsequent iterations. Safety Execution and Motion Control: The servo motor system with dual encoder redundancy performs dual-channel sampling of the robotic arm joint angles and uses Kalman filtering to achieve data fusion and fault diagnosis. The electronic cam system generates a flexible motion curve based on decision commands and compensates for trajectory deviations caused by inertial loads through torque feedforward control. The predicted state of the digital twin model is compared with the feedback state of the physical entity in real time, and adaptive compensation adjustment is triggered when the deviation exceeds a certain range. Upon detecting an unexpected sudden change in contact force, the safety braking procedure is immediately initiated, the robotic arm enters zero-force suspension mode, and an audible and visual alarm is triggered. Human-machine collaboration and visual debugging: The operator uses the gesture recognition module to call the 3D projection device and drags and adjusts the movement path and grasping posture of the robotic arm in the virtual tooling interface. The digital twin engine renders the adjusted motion trajectory in real time and displays the collision risk level and structural stress distribution at key nodes; The audible and visual alarm unit adopts a graded warning mechanism: a yellow indicator light indicates that the cleanliness level exceeds the standard, and a red flashing light accompanied by a buzzer alarm indicates that a collision or overload has occurred; All debugging records and abnormal events are automatically generated into log files, which can be exported through the human-machine interface terminal for process traceability and system optimization.

[0016] Preferably, in the environmental perception array, the laser profile scanner and the six-dimensional force sensor are arranged in a spatial orthogonal manner, the ultrasonic ranging module is symmetrically installed at the four corners of the sorting platform, the particle counter is suspended above the work area through a telescopic bracket, and the data from each sensor are fused into the digital twin engine after being aligned with timestamps.

[0017] The laser profile scanner and the six-dimensional force sensor are deployed in a spatially orthogonal manner. The laser profile scanner is horizontally installed above the sorting platform, and the six-dimensional force sensor is vertically embedded in the surface of the sorting platform. The two form an orthogonal reference plane of a three-dimensional spatial coordinate system. The ultrasonic ranging modules are symmetrically fixed at the four corners of the sorting platform by rigid brackets. The transmitting surface of each module maintains a certain distance from the edge of the platform, forming a ranging network covering the entire platform area. The particle counter is suspended directly above the work area via an electrically telescopic bracket. The extension range of the bracket dynamically adjusts the vertical distance between the sampling port and the platform according to the working height. The raw data streams output by each sensor are given a unified timestamp through a hardware time synchronization module, and after preprocessing, they are transmitted to the digital twin engine for spatial coordinate system calibration and multimodal data fusion.

[0018] Preferably, the rigid-flexible coupled multibody dynamics solver and finite element analysis module of the digital twin engine achieve parallel computation through a shared memory architecture, the collision detection module adopts a layered bounding box algorithm, the response module has a built-in elastic deformation compensation table, and the structural deformation simulation accuracy is corrected by closed-loop calibration using real-time measured six-dimensional force data.

[0019] The rigid-flexible coupled multibody dynamics solver and finite element analysis module achieve parallel computation through a shared memory architecture. The solver is responsible for the kinematics calculation of the robot arm body, while the finite element module synchronously processes the contact deformation analysis of the end effector. The collision detection module uses a layered bounding box algorithm. First, it performs coarse screening using axis-aligned bounding boxes (AABB), and then performs precise intersection testing on potential collision areas using oriented bounding boxes (OBB). The response module has a built-in elastic deformation compensation table. Based on the contact force data measured in real time by the six-dimensional force sensor, the material elastic modulus correction coefficient is obtained by looking up the table. The structural deformation simulation results are compared with the physical world measurements through a closed-loop feedback loop. When the error exceeds the threshold, the finite element mesh is re-divided and the material parameters are iteratively optimized.

[0020] Preferably, the deep reinforcement learning network of the intelligent decision-making center adopts a dual-delay deep deterministic policy gradient architecture. The output layer of the actor network in the deep reinforcement learning network and the expert rule base of the rule engine are dynamically integrated through a gating mechanism. The control policy optimization process records the mapping relationship between the geometric features of objects and the grasping posture in historical sorting data, and generates interpretive decision logs for visualization by the three-dimensional projection device of the human-computer interaction mechanism.

[0021] The deep reinforcement learning network adopts a dual-delay deep deterministic policy gradient (TD3) architecture, with the actor network and the critic network deployed in independent computing units, and policy exploration and value evaluation decoupled through asynchronous gradient updates; The actor network output layer and the expert rule base of the rule engine are integrated through a gating mechanism. When a new type of object geometric feature is detected, the gating unit automatically switches to the deep learning path to generate a grasping strategy. The control strategy optimization module continuously records historical sorting data and establishes a database mapping the relationship between object geometric parameters (length, width, height, and center of gravity coordinates) and grasping posture (gripper opening degree and contact point distribution). The decision log generation unit transforms the optimization process into structured text, including the basis for strategy selection, parameter adjustment trajectory, and expected effect evaluation, which is then visualized in real time on the user interface via a 3D projection device.

[0022] A digital twin-based dual-station robot sorting optimization terminal, based on the aforementioned digital twin-based dual-station robot sorting optimization system, includes: Heterogeneous computing accelerator cards: It is equipped with a neural network processor and a vision processing unit to support FP16 precision computing and hardware-level sparsity acceleration; Safety isolation device: Unidirectional data diodes and physical isolation transformers are used to ensure complete isolation between the control network and the information network; Hot-swappable storage array: Configure a RAID6 redundant disk array and enterprise-grade SSD cache to support real-time mirror backup and fast rollback of sorting process data.

[0023] Data acquisition and preprocessing stage: By deploying an industrial-grade vision sensor array in a dual-station robot operation scenario, multi-dimensional feature data of the objects to be sorted are collected in real time, including the object's geometric contour, surface texture, mass distribution, and spatial coordinate information. The vision data is protocol stripped and transmitted unidirectionally via a unidirectional data diode in a safety isolation device, ensuring that control commands can only flow unidirectionally from the information network to the control network, preventing the reverse penetration of external networks into the production control system. Heterogeneous computing acceleration stage: The collected raw visual data is first input into the visual processing unit in the heterogeneous computing accelerator card, where noise reduction, edge enhancement, and feature point extraction are completed through a hardware-level image preprocessing pipeline. The neural network processor simultaneously loads a pre-trained lightweight convolutional neural network model and uses an FP16 precision computing architecture to perform real-time inference on the preprocessed feature data, outputting object category determination results and grasping posture parameters. The calculation process uses hardware-level sparsity acceleration technology to optimize the skipping of zero-value elements in the neural network weight matrix, thereby reducing the inference time of a single frame image. Digital twin model mapping stage: The object recognition results and posture parameters output by the heterogeneous computing accelerator card are synchronously mapped to the pre-constructed digital twin model of the dual-station robot. The twin model simulates the kinematic constraints and load dynamics of the robotic arm in real time based on the physics engine, and generates multiple sets of candidate sorting trajectories that meet the requirements of collision avoidance, cycle balance and energy consumption through iterative optimization algorithms. The optimization results are output to the dual-station robot control system after being level-converted by the physical isolation transformer in the safety isolation device. Execution feedback and data consolidation phase: During the execution of the optimized sorting trajectory by the dual-station robot, the hot-swappable storage array continuously records real-time status data such as joint angles, end effector torque, and work cycle through a RAID6 redundant disk array. When a sorting failure or abnormal shutdown event is detected, the enterprise-grade SSD cache immediately activates the data rollback mechanism to quickly restore the system state to the time point of the most recent successful sorting operation, ensuring production continuity. Under normal operating conditions, historical data is synchronously stored on at least two physically isolated hard disk media in a mirror backup manner.

[0024] In summary, compared with the prior art, the present invention provides a digital twin-based dual-station robot sorting optimization method, system, and terminal, which has the following beneficial effects: This invention systematically solves the core deficiencies of traditional sorting systems in terms of dynamic adaptation, conflict prevention, and efficiency optimization by constructing a digital twin-driven dual-station collaborative sorting system. At the environmental perception level, the composite sensing unit breaks through the limitations of traditional single sensors. By working together with lidar and polarization camera, the system synchronously captures the geometric shape, material properties and spatial pose information of the target object. Combined with time synchronization processing mechanism, it ensures that the initial state consistency between the virtual workstation and the physical scene reaches sub-millimeter level accuracy. This multi-modal data fusion capability enables the system to respond in real time to dynamic disturbances such as clean panel deformation and protective sleeve position offset, effectively overcoming the problem of operation interruption caused by environmental parameter offset in traditional systems. At the system architecture level, the environmental sensing array achieves full-dimensional monitoring of the sorting environment through the collaborative work of a six-dimensional force sensor and a particle counter. Compared with traditional cleanliness detection solutions, the recognition rate of contaminant particles is improved. The digital twin engine integrates a rigid-flexible coupling dynamic solver, enabling the structural deformation simulation accuracy to reach the micron level, providing reliable assurance for flexible manufacturing scenarios. The deep reinforcement learning network of the intelligent decision-making center is trained with historical data, which enables the control strategy to optimize autonomously and improve efficiency. Combined with the logical reasoning ability of the expert system knowledge base, it forms a hybrid decision-making mode that integrates data-driven and experience-based knowledge. The safety actuator adopts a dual encoder redundancy design and combines the torque feedforward control of the electronic cam system to improve position repeatability and significantly reduce the risk of collision during collaborative operations. The heterogeneous computing architecture of the terminal device accelerates the real-time rendering efficiency of the digital twin model through the collaborative acceleration of the neural network processor and the vision processing unit, and shortens the system response latency. The security isolation device uses physical isolation transformers and unidirectional data diodes to build a protection system that meets the security standards of industrial control systems and effectively blocks potential network attack paths. The combination of RAID6 redundancy in hot-swappable storage arrays and enterprise-grade SSD caching enables real-time mirror backup of sorting process data, shortens fault recovery time, and provides data security for the continuous and stable operation of flexible production lines. The entire solution achieves a paradigm shift in sorting operations from passive response to proactive optimization through the deep integration of digital twin technology and industrial control systems, fully adapting to the stringent requirements of flexible manufacturing for dynamic adaptability and system efficiency. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the dual-station robot sorting optimization method based on digital twins according to the present invention.

[0026] Figure 2 This is a schematic diagram of the dual-station robot sorting optimization system based on digital twins of the present invention.

[0027] Figure 3 This is a schematic diagram of the dual-station robot sorting optimization terminal based on digital twins according to the present invention. Detailed Implementation

[0028] 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.

[0029] Please see Figure 1This invention provides a technical solution: a dual-station robot sorting optimization method based on digital twins, comprising the following steps: S1. Multimodal environmental parameter acquisition: By integrating a composite sensing unit with lidar and polarization camera, the geometric contour data, surface reflection characteristics and spatial pose information of the clean panel and protective cover are acquired simultaneously. The data is then input into the digital twin engine after time synchronization processing. S2, Dynamic Digital Twin Modeling: Based on the collected environmental parameters, a physical rendering engine is used to construct a virtual sorting scene. The scene includes a material property database and a flexible object simulation module to support real-time updates of object deformation state. S3, Multi-view Data Fusion Calibration: The depth map acquired by the binocular stereo vision system is fused with the geometric parameters collected in S1 at the feature level, and sub-millimeter-level spatial registration is achieved through a nonlinear optimization algorithm to generate a high-precision digital twin workstation. S4. Spatiotemporal Constraint Trajectory Planning: Generate the cooperative motion trajectory of the dual-station robot in the digital twin workstation. The trajectory includes a timestamp sequence, joint space interpolation points and velocity constraints, and meets the obstacle avoidance requirements in C-space. S5. Real-time conflict detection and avoidance: A collision detection engine based on discrete event simulation is used to predict timing conflicts in dual-station operations. When spatial overlap risk is detected, alternative paths are generated through a trajectory replanning algorithm, which includes a dynamic obstacle avoidance strategy. S6. Adaptive control parameter generation: Based on the collision detection results, the robot's motion speed, acceleration, and buffer parameters are adjusted in real time using a fuzzy PID controller to generate a smooth velocity profile curve. S7, Closed-loop iterative optimization execution: The actual execution trajectory is compared with the trajectory predicted by the digital twin model, and the control command is continuously corrected by the model predictive control algorithm, forming a closed-loop optimization mechanism that includes environmental disturbance compensation.

[0030] Multimodal environmental parameter acquisition: By integrating a LiDAR and a polarization camera into a composite sensing unit, the following operations are performed simultaneously: LiDAR is used to acquire the geometric contour data and spatial pose information of the clean panel and protective sleeve, forming a three-dimensional point cloud dataset; The surface reflection characteristics of the target object are collected using a polarization camera, including polarization angle, reflectivity, and texture features. The above multimodal data is processed for time synchronization to eliminate sampling delay deviations between different sensors; The processed environmental parameters are input into the digital twin engine as the basic data source for building the virtual scene; Dynamic digital twin modeling: Based on the collected environmental parameters, the following modeling operations are performed: A physically based rendering engine is invoked to construct a virtual sorting scene containing a database of material properties, which covers the elastic modulus, friction coefficient, and deformation threshold of different materials. A flexible object simulation module is deployed to update the deformation state of the object in real time through finite element analysis, simulating the non-rigid deformation process of the protective plate under mechanical gripping. Establish a two-way mapping relationship between virtual scenes and physical entities, and support real-time synchronous updates of dynamic parameters; Multi-view data fusion calibration The depth map acquired by the binocular stereo vision system is fused and calibrated with the S1 acquisition data: Denoising is performed on the binocular vision depth map, and edge contour feature points are extracted; The lidar point cloud data and visual feature points are fused at the feature level to construct a three-dimensional spatial coordinate correlation matrix. By iteratively solving the spatial transformation parameters using a nonlinear optimization algorithm, cross-modal data registration with sub-millimeter accuracy is achieved. Generate a high-precision digital twin workstation model to provide a geometric constraint benchmark for trajectory planning; Generate collaborative trajectories for dual-workstation robots in a digital twin workstation: Initial candidate paths are generated based on the improved RRT* algorithm to meet the obstacle avoidance requirements in C-space. The candidate paths are timestamped to construct a kinematic model that includes joint space interpolation points; Apply velocity constraints, optimize trajectory smoothness using gradient projection method, and generate motion commands that meet dynamic constraints; Real-time conflict detection and avoidance Deploy conflict detection and avoidance mechanisms: Construct a conflict detection engine based on discrete event simulation and establish a dual-station operation timing model; Real-time monitoring of the spatial occupancy status of the robotic arm's end effector to predict potential operational conflicts; When spatial overlap risk is detected, the trajectory replanning algorithm is activated to generate alternative paths that include dynamic obstacle avoidance strategies. Conflict avoidance is achieved through local path correction, ensuring the continuity of collaborative work between two workstations; Adaptive control parameter generation Adaptive control is executed based on the collision detection results: The current operation status is analyzed by a fuzzy PID controller, and a fuzzy rule base for speed, acceleration, and buffer parameters is established. The PID control parameters are dynamically adjusted based on real-time monitoring of changes in the robotic arm's load. Generate velocity profile curves that satisfy smoothness constraints to eliminate vibration phenomena caused by abrupt trajectory changes; Closed-loop iterative optimization execution: Construct a closed-loop optimization mechanism: The actual execution trajectory is mapped to the digital twin space using a positive kinematics model; By comparing the deviation between the actual trajectory and the predicted trajectory, the compensation amount is calculated using a model predictive control algorithm. The compensation amount is injected into the control command generation module to achieve environmental disturbance compensation; Continuously iterate and optimize control commands until the operational accuracy meets the preset threshold requirements; By fusing multimodal data from lidar and polarization cameras, geometric contours and surface characteristic parameters are acquired simultaneously. Compared with single-modal perception solutions, the completeness of environmental features is improved, providing a data foundation for accurate modeling. The physically based rendering engine works in conjunction with the flexible object simulation module to reduce the latency of deformation simulation in virtual scenes, effectively supporting the real-time requirements of dual-station collaborative operations. A nonlinear optimization algorithm is used to achieve sub-millimeter-level cross-modal data registration, which keeps the spatial deviation between the digital twin workstation and the physical entity within a certain range, significantly improving the reliability of trajectory planning. The spatiotemporal constraint trajectory planning algorithm shortens the cycle time of the dual-station robot, while reducing the number of interruptions through conflict detection and avoidance mechanisms, thereby improving the overall system throughput. The adaptive fuzzy PID control strategy shortens the system's response time to load changes, improves the smoothness of the velocity profile curve, and effectively suppresses mechanical vibration and trajectory deviation. Through closed-loop iteration of digital twin models and physical entities, the system achieves autonomous environmental disturbance compensation, thereby improving the success rate of operations under complex conditions and significantly reducing the need for manual intervention.

[0031] In the acquisition of multimodal environmental parameters, the lidar adopts the phase ranging principle, and the polarization camera is equipped with a tunable filter array to separate the specular reflection component in the surface reflected light.

[0032] LiDAR calculates the spatial distance between a target object and a sensor based on the principle of phase ranging by emitting modulated laser signals and detecting the phase difference of the reflected signals. A polarization camera analyzes the polarization state of incident light using a tunable filter array, separating the specular reflection component and the diffuse reflection component in the surface reflected light. The separated specular reflection component data is spatially registered with the lidar ranging results to generate a 3D point cloud model with fused environmental parameters. High-precision distance measurement is achieved by using the phase-based ranging principle. Combined with the ability of a polarization camera to separate specular reflection components, ranging errors caused by highly reflective surfaces are effectively suppressed, improving the accuracy and robustness of multimodal environmental parameter acquisition under complex lighting conditions.

[0033] In spatiotemporal constrained trajectory planning, the singularity analysis of the tool center point trajectory is also included, and the condition number evaluation is used to avoid the robot from entering a kinematically singular configuration.

[0034] Based on the robot's kinematic model, spatiotemporal constraints on the trajectory of the tool's center point are constructed; Calculate the condition number of the Jacobian matrix for the planned trajectory and evaluate the kinematic singularity of the current configuration; When the number of conditions exceeds the preset threshold, the robot is prevented from entering the strange configuration region by adjusting the trajectory parameters or inserting intermediate path points. By avoiding kinematically singular configurations in advance through singularity analysis, the robot can prevent motion loss of control due to joint lock-up or sudden velocity changes, significantly improving the safety and execution efficiency of trajectory planning.

[0035] The material property database includes the elastic modulus, Poisson's ratio, and surface friction coefficient of the cleanroom panel. The flexible object simulation module uses a mass-spring model to simulate deformation, and the spring stiffness coefficient is dynamically adjusted based on the real-time collected contact force data.

[0036] The parameters such as the elastic modulus, Poisson's ratio, and surface friction coefficient of the cleanroom panel are retrieved from the material property database. A hybrid geometric-physical representation structure for flexible objects is constructed based on a mass-spring model, where each mass is connected to adjacent mass points via springs. Real-time acquisition of contact force data; dynamic adjustment of spring stiffness coefficient based on Hooke's Law; updating the deformation state of flexible objects. By constructing a high-fidelity deformation model based on real material parameters and combining it with a dynamic stiffness adjustment mechanism, we can accurately simulate the interaction process between flexible objects and rigid panels, providing reliable data support for precision assembly tasks.

[0037] The dynamic obstacle avoidance strategy uses the speed obstacle method to calculate alternative paths and introduces a safe distance threshold to adjust the avoidance priority. When a spatial overlap risk is detected, trajectory replanning of low-priority workstations is triggered first.

[0038] The velocity obstacle method is used to calculate the potential collision time window and the set of alternative avoidance paths between the robot and dynamic obstacles. The alternative paths are prioritized according to a preset safe distance threshold, and an avoidance priority queue is generated. When a risk of spatial overlap is detected, trajectory replanning for low-priority workstations is triggered first, while high-priority task channels are released. By quantifying avoidance priorities using safe distance thresholds, a graded response to dynamic obstacles can be achieved while ensuring the continuity of the overall task, effectively reducing the probability of trajectory conflicts and the risk of system downtime in multi-workstation collaborative scenarios.

[0039] Please see Figure 2 The dual-station robot sorting optimization system based on digital twins adopts the aforementioned dual-station robot sorting optimization method based on digital twins, including: Environmental sensing array: Composed of a laser profile scanner, a six-dimensional force sensor, an ultrasonic ranging module, and a particle counter, it is used to achieve full-dimensional perception and cleanliness monitoring of the sorting environment; Digital Twin Engine: It integrates a rigid-flexible coupled multibody dynamics solver and a finite element analysis module to support real-time collision detection, response, and structural deformation simulation; Intelligent Decision-Making Center: Deploy deep reinforcement learning networks and rule engines to autonomously optimize control strategies based on historical sorting data, and load expert system knowledge bases for logical reasoning; Safety enforcement agencies: The servo motor and electronic cam system with dual encoder redundancy design have torque feedforward control function; Human-computer interaction terminal: Equipped with a gesture recognition module, a 3D projection device, and an audible and visual alarm unit, it supports the visual debugging and abnormal status prompts of virtual tooling.

[0040] Environmental sensing and data acquisition: The sorting area is scanned in three dimensions by a laser contour scanner in the environmental perception array to obtain the geometric dimensions, spatial position and orientation parameters of the objects to be sorted. The six-dimensional force sensor monitors the contact force and torque data between the end effector of the robotic arm and the object in real time. Combined with the spatial positioning information output by the ultrasonic ranging module, a dynamic interaction model between the object and the actuator is constructed. The particle counter continuously monitors the cleanliness indicators of the sorting environment, and triggers a cleanliness compensation process when the concentration of suspended particles exceeds a preset threshold. All sensing data is transmitted in real time to the digital twin engine via industrial Ethernet, forming the initial data mapping between the physical entity and the virtual model; Digital twin modeling and simulation verification: The digital twin engine constructs a three-dimensional coupled model of the robotic arm, object, and environment based on received real-time data through a rigid-flexible coupled multibody dynamics solver. The finite element analysis module performs stress-strain simulations on key structural components to predict deformation and fatigue life under high load conditions. Conduct virtual collision detection: Simulate the movement trajectory of a robotic arm in digital space to identify potential interference risks with fixed facilities such as conveyor belts and guardrails in advance; A virtual verification report containing kinematic constraints, dynamic parameters, and safety boundaries is generated and fed back to the intelligent decision-making center. Intelligent decision-making and strategy optimization: The deep reinforcement learning network generates a set of candidate control policies based on historical sorting data and current environmental perception information using the Q-learning algorithm; The rules engine loads the expert system knowledge base, performs logical verification on candidate strategies, and excludes operation sequences that violate safety specifications or process requirements; Based on the simulation results of the digital twin engine, a multi-objective optimization algorithm (such as NSGA-II) is used to perform weighted evaluation of indicators such as sorting efficiency, energy consumption, and mechanical wear. The optimal control command is output to the safety execution mechanism, and the optimized strategy parameters are stored in the knowledge base for subsequent iterations. Safety Execution and Motion Control: The servo motor system with dual encoder redundancy performs dual-channel sampling of the robotic arm joint angles and uses Kalman filtering to achieve data fusion and fault diagnosis. The electronic cam system generates a flexible motion curve based on decision commands and compensates for trajectory deviations caused by inertial loads through torque feedforward control. The predicted state of the digital twin model is compared with the feedback state of the physical entity in real time, and adaptive compensation adjustment is triggered when the deviation exceeds a certain range. Upon detecting an unexpected sudden change in contact force, the safety braking procedure is immediately initiated, the robotic arm enters zero-force suspension mode, and an audible and visual alarm is triggered. Human-machine collaboration and visual debugging: The operator uses the gesture recognition module to call the 3D projection device and drags and adjusts the movement path and grasping posture of the robotic arm in the virtual tooling interface. The digital twin engine renders the adjusted motion trajectory in real time and displays the collision risk level and structural stress distribution at key nodes; The audible and visual alarm unit adopts a graded warning mechanism: a yellow indicator light indicates that the cleanliness level exceeds the standard, and a red flashing light accompanied by a buzzer alarm indicates that a collision or overload has occurred; All debugging records and abnormal events are automatically generated into log files, which can be exported through the human-machine interface terminal for process traceability and system optimization. By leveraging the synergy between the environmental perception array and the digital twin engine, dynamic mapping of all elements in the sorting process is achieved, thereby improving the system's response speed to complex working conditions. The deep integration of rigid-flexible coupling dynamics model and finite element analysis enables the positioning accuracy of the robotic arm end effector to reach a certain level, significantly reducing the sorting failure rate caused by structural deformation. The hybrid decision-making architecture of deep reinforcement learning and rule engine improves sorting efficiency compared to traditional methods while ensuring process compliance. The combination of dual encoder redundancy design and torque feedforward control improves the repeatability of joint position control and effectively extends the service life of the servo system. The integration of a 3D visualization debugging interface with gesture interaction technology shortens system debugging time and reduces reliance on the skills of professional operators. The integration of particle counters and cleanliness compensation processes ensures that the sorting environment continuously meets the ISO 14644-1 Class 5 standard, making it particularly suitable for cleanliness-sensitive scenarios such as electronic components.

[0041] In the environmental perception array, the laser profile scanner and the six-dimensional force sensor are arranged in a spatial orthogonal manner. The ultrasonic ranging module is symmetrically installed at the four corners of the sorting platform. The particle counter is suspended above the work area through a telescopic bracket. The data from each sensor are fused into the digital twin engine after being aligned with timestamps.

[0042] The laser profile scanner and the six-dimensional force sensor are deployed in a spatially orthogonal manner. The laser profile scanner is horizontally installed above the sorting platform, and the six-dimensional force sensor is vertically embedded in the surface of the sorting platform. The two form an orthogonal reference plane of a three-dimensional spatial coordinate system. The ultrasonic ranging modules are symmetrically fixed at the four corners of the sorting platform by rigid brackets. The transmitting surface of each module maintains a certain distance from the edge of the platform, forming a ranging network covering the entire platform area. The particle counter is suspended directly above the work area via an electrically telescopic bracket. The extension range of the bracket dynamically adjusts the vertical distance between the sampling port and the platform according to the working height. The raw data streams output by each sensor are given a unified timestamp through a hardware time synchronization module, and after preprocessing, they are transmitted to the digital twin engine for spatial coordinate system calibration and multimodal data fusion. Orthogonal arrangement spatially decouples laser scanning data from force data, enabling independent compensation for measurement deviations caused by mechanical vibration and improving the accuracy of 3D contour reconstruction. Symmetrical ultrasonic ranging networks eliminate blind spots in platform edge detection, achieving a certain range of planar positioning accuracy and effectively avoiding the risk of edge collisions during the sorting process; The telescopic bracket allows the particle counter to adapt to different working heights, and can capture particles of a certain size range at standard heights, thus improving the coverage of air cleanliness monitoring. The timestamp alignment mechanism ensures the spatiotemporal consistency of multi-sensor data in the digital twin engine, reducing the latency of dynamic scene reconstruction.

[0043] The rigid-flexible coupled multibody dynamics solver and finite element analysis module of the digital twin engine achieve parallel computation through a shared memory architecture. The collision detection module adopts a hierarchical bounding box algorithm, the response module has a built-in elastic deformation compensation table, and the accuracy of structural deformation simulation is corrected by closed-loop calibration using real-time measured six-dimensional force data.

[0044] The rigid-flexible coupled multibody dynamics solver and finite element analysis module achieve parallel computation through a shared memory architecture. The solver is responsible for the kinematics calculation of the robot arm body, while the finite element module synchronously processes the contact deformation analysis of the end effector. The collision detection module uses a layered bounding box algorithm. First, it performs coarse screening using axis-aligned bounding boxes (AABB), and then performs precise intersection testing on potential collision areas using oriented bounding boxes (OBB). The response module has a built-in elastic deformation compensation table. Based on the contact force data measured in real time by the six-dimensional force sensor, the material elastic modulus correction coefficient is obtained by looking up the table. The structural deformation simulation results are compared with physical world measurements through a closed-loop feedback loop. When the error exceeds the threshold, finite element mesh re-division and iterative optimization of material parameters are triggered. The shared memory architecture improves the efficiency of multiphysics coupling calculations, achieving a real-time simulation refresh rate of 1kHz with a 4-core CPU configuration; The layered bounding box algorithm reduces the computational cost of collision detection, thus shortening the detection response time in complex scenes; The elastic deformation compensation table covers 20 common engineering materials, keeping the contact force prediction error within a certain range and significantly improving the sorting success rate of brittle items; The closed-loop correction mechanism reduces the state synchronization error between the digital twin model and the physical entity, ensuring the physical feasibility of the virtual debugging results.

[0045] The deep reinforcement learning network of the intelligent decision-making center adopts a dual-delay deep deterministic policy gradient architecture. The output layer of the actor network in the deep reinforcement learning network and the expert rule base of the rule engine are dynamically integrated through a gating mechanism. The policy optimization process records the mapping relationship between the geometric features of objects and the grasping posture in historical sorting data, and generates interpretive decision logs for visualization by the three-dimensional projection device of the human-computer interaction mechanism.

[0046] The deep reinforcement learning network adopts a dual-delay deep deterministic policy gradient (TD3) architecture, with the actor network and the critic network deployed in independent computing units, and policy exploration and value evaluation decoupled through asynchronous gradient updates; The actor network output layer and the expert rule base of the rule engine are integrated through a gating mechanism. When a new type of object geometric feature is detected, the gating unit automatically switches to the deep learning path to generate a grasping strategy. The control strategy optimization module continuously records historical sorting data and establishes a database mapping the relationship between object geometric parameters (length, width, height, and center of gravity coordinates) and grasping posture (gripper opening degree and contact point distribution). The decision log generation unit transforms the optimization process into structured text, including the basis for strategy selection, parameter adjustment trajectory, and expected effect evaluation, which is visualized in real time on the operation interface through a 3D projection device. The TD3 architecture improves the policy convergence speed and achieves a certain sorting success rate after 100,000 training iterations, which is significantly better than the traditional Q-learning method. The gating fusion mechanism retains the interpretability advantage of the rule engine, improves the credibility of decisions under abnormal conditions, and reduces the frequency of manual intervention. The geometry-attitude mapping database covers more than 300 standard industrial parts, which shortens the adaptation time of new objects to within a certain time and improves the changeover efficiency of the production line. Visualized decision logs enable operators to trace the system's decision-making path, reducing average fault diagnosis time and significantly enhancing human-machine collaboration efficiency.

[0047] Please see Figure 3 A digital twin-based dual-station robot sorting optimization terminal, based on the aforementioned digital twin-based dual-station robot sorting optimization system, includes: Heterogeneous computing accelerator cards: It is equipped with a neural network processor and a vision processing unit to support FP16 precision computing and hardware-level sparsity acceleration; Safety isolation device: Unidirectional data diodes and physical isolation transformers are used to ensure complete isolation between the control network and the information network; Hot-swappable storage array: Configure a RAID6 redundant disk array and enterprise-grade SSD cache to support real-time mirror backup and fast rollback of sorting process data.

[0048] Data acquisition and preprocessing stage: By deploying an industrial-grade vision sensor array in a dual-station robot operation scenario, multi-dimensional feature data of the objects to be sorted are collected in real time, including the object's geometric contour, surface texture, mass distribution, and spatial coordinate information. The vision data is protocol stripped and transmitted unidirectionally via a unidirectional data diode in a safety isolation device, ensuring that control commands can only flow unidirectionally from the information network to the control network, preventing the reverse penetration of external networks into the production control system. Heterogeneous computing acceleration stage: The collected raw visual data is first input into the visual processing unit in the heterogeneous computing accelerator card, where noise reduction, edge enhancement, and feature point extraction are completed through a hardware-level image preprocessing pipeline. The neural network processor simultaneously loads a pre-trained lightweight convolutional neural network model and uses an FP16 precision computing architecture to perform real-time inference on the preprocessed feature data, outputting object category determination results and grasping posture parameters. The calculation process uses hardware-level sparsity acceleration technology to optimize the skipping of zero-value elements in the neural network weight matrix, thereby reducing the inference time of a single frame image. Digital twin model mapping stage: The object recognition results and posture parameters output by the heterogeneous computing accelerator card are synchronously mapped to the pre-constructed digital twin model of the dual-station robot. The twin model simulates the kinematic constraints and load dynamics of the robotic arm in real time based on the physics engine, and generates multiple sets of candidate sorting trajectories that meet the requirements of collision avoidance, cycle balance and energy consumption through iterative optimization algorithms. The optimization results are output to the dual-station robot control system after being level-converted by the physical isolation transformer in the safety isolation device. Execution feedback and data consolidation phase: During the execution of the optimized sorting trajectory by the dual-station robot, the hot-swappable storage array continuously records real-time status data such as joint angles, end effector torque, and work cycle through a RAID6 redundant disk array. When a sorting failure or abnormal shutdown event is detected, the enterprise-grade SSD cache immediately activates the data rollback mechanism to quickly restore the system state to the time point of the most recent successful sorting operation, ensuring production continuity. Under normal operating conditions, historical data is synchronously stored on at least two physically isolated hard disk media in a mirror backup manner. By leveraging the collaborative work of the neural network processor and vision processing unit in the heterogeneous computing accelerator card, the processing latency for object recognition and pose estimation in complex scenarios is reduced to industrial-grade real-time requirements. Compared to traditional CPU solutions, this improves computational efficiency and significantly enhances the dynamic response capability of the dual-station robot system. The security isolation device adopts a combination design of unidirectional data diodes and physical isolation transformers to completely block the bidirectional data path between the control network and the information network. It has been tested by a third-party organization and can resist 23 types of typical industrial network attacks, including TCP reset attacks and man-in-the-middle attacks. The system network security level reaches IEC 62443-3-3 standard SL3 level. The RAID6 redundancy architecture of the hot-swappable storage array can tolerate the simultaneous failure of any two hard drives without data loss. The IOPS performance provided by the enterprise-grade SSD cache is improved compared to traditional mechanical hard drives, which shortens the time for real-time mirror backup and failure rollback of sorting operation data and reduces the system mean time to failure (MTTR). The bidirectional mapping mechanism between the digital twin model and the physical system enables the sorting trajectory optimization process to be independent of manual teaching. Actual production verification has shown that it can improve the efficiency of collaborative operation of dual-station robots, while reducing energy consumption by dynamically adjusting the motion parameters of the robotic arm.

[0049] This solution systematically addresses the core shortcomings of traditional sorting systems in terms of dynamic adaptation, conflict prevention, and efficiency optimization by constructing a digital twin-driven dual-station collaborative sorting system. At the environmental perception level, the composite sensing unit overcomes the limitations of traditional single sensors. By working in tandem with a lidar and a polarization camera, the system synchronously captures the geometric shape, material properties, and spatial pose information of the target object. Combined with a time synchronization processing mechanism, it ensures that the initial state consistency between the virtual workstation and the physical scene reaches sub-millimeter level accuracy. This multi-modal data fusion capability enables the system to respond in real time to dynamic disturbances such as cleanroom panel deformation and protective sleeve position offset, effectively overcoming the problem of operation interruption caused by environmental parameter offset in traditional systems.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing dual-station robotic sorting based on digital twinning, characterized in that, Comprise the following steps: S1, multi-modal environment parameter acquisition: Through the composite perception unit integrating laser radar and polarization camera, the geometric profile data, surface reflection characteristics and spatial pose information of the clean panel and the protective panel set are synchronously acquired, and the data are input into the digital twin engine after time synchronization processing; S2, dynamic digital twin modeling: Based on the collected environment parameters, a virtual sorting scene is constructed by using a physics-based rendering engine, the scene includes a material attribute database and a flexible object simulation module to support real-time updating of object deformation state; S3, multi-view data fusion calibration: The depth map obtained by the binocular stereo vision system is fused with the geometric parameters collected in S1 at the feature level, and sub-millimeter level spatial registration is realized through a nonlinear optimization algorithm to generate a high-precision digital twin station; S4, space-time constraint trajectory planning: The collaborative motion trajectory of the double-station robot is generated in the digital twin station, the trajectory includes a timestamp sequence, joint space interpolation points and speed constraint conditions, and meets the C-space obstacle avoidance requirement; S5, real-time conflict detection and avoidance: A conflict detection engine based on discrete event simulation is used to predict the timing conflict of the double-station operation, and when a spatial overlap risk is detected, an alternative path is generated through a trajectory re-planning algorithm, which includes a dynamic obstacle avoidance strategy; S6, adaptive control parameter generation: According to the conflict detection result, the robot motion speed, acceleration and buffer interval parameters are adjusted in real time through a fuzzy PID controller to generate a smooth speed profile curve; S7, closed-loop iterative optimization execution: The actual execution trajectory is compared with the predicted trajectory of the digital twin model, and the control command is continuously corrected using a model predictive control algorithm to form a closed-loop optimization mechanism that includes environmental disturbance compensation.

2. The digital-twin-based dual-station robotic sorting optimization method of claim 1, wherein: In the multi-modal environment parameter acquisition, the laser radar uses a phase distance measurement principle, and the polarization camera is equipped with a tunable optical filter array to separate the specular reflection component in the surface reflected light.

3. The digital-twin-based dual-station robotic sorting optimization method of claim 1, wherein: In the space-time constraint trajectory planning, the space-time constraint trajectory planning further includes singularity analysis of the tool center point trajectory to avoid the robot entering a kinematic singularity configuration through condition number evaluation.

4. The digital-twin-based dual-station robotic sorting optimization method of claim 1, wherein: The material attribute database includes the elastic modulus, Poisson's ratio and surface friction coefficient of the clean panel, and the flexible object simulation module uses a particle-spring model for deformation simulation, and the spring stiffness coefficient is dynamically adjusted according to the real-time collected contact force data.

5. The digital-twin-based dual-station robotic sorting optimization method of claim 1, wherein: The dynamic obstacle avoidance strategy uses a velocity obstacle method to calculate an alternative path, and introduces a safety distance threshold to adjust the avoidance priority, and when a spatial overlap risk is detected, the trajectory re-planning of the low-priority station is triggered preferentially.

6. A dual-station robotic sorting optimization system based on digital twinning, employing the dual-station robotic sorting optimization method based on digital twinning according to any one of claims 1-5, characterized in that, Comprise: Environment perception array: Composed of laser profile scanner, six-axis force sensor, ultrasonic ranging module and particle counter to realize full-dimensional perception and cleanliness monitoring of the sorting environment; Digital twin engine: Integrates rigid-flexible coupled multi-body dynamics solver and finite element analysis module to support real-time collision detection, response and structure deformation simulation; Intelligent decision-making center: Deploy deep reinforcement learning network and rule engine to autonomously optimize control strategy according to historical sorting data, and load expert system knowledge base for logical reasoning; Safety enforcement agencies: The servo motor and electronic cam system with dual encoder redundancy design have torque feedforward control function; Human-computer interaction terminal: Equipped with a gesture recognition module, a 3D projection device, and an audible and visual alarm unit, it supports the visual debugging and abnormal status prompts of virtual tooling.

7. The digital twin based dual position robotic sortation optimization system of claim 6, wherein: In the environmental perception array, the laser profile scanner and the six-dimensional force sensor are arranged in a spatial orthogonal manner, the ultrasonic ranging module is symmetrically installed at the four corners of the sorting platform, and the particle counter is suspended above the work area through a telescopic bracket. The data from each sensor are fused into the digital twin engine after being aligned with timestamps.

8. The digital twin based dual position robotic sortation optimization system of claim 6, wherein: The rigid-flexible coupled multibody dynamics solver and finite element analysis module of the digital twin engine achieve parallel computation through a shared memory architecture. The collision detection module adopts a layered bounding box algorithm, the response module has a built-in elastic deformation compensation table, and the structural deformation simulation accuracy is corrected by closed-loop calibration using real-time measured six-dimensional force data.

9. The digital twin based dual position robotic sortation optimization system of claim 6, wherein: The deep reinforcement learning network of the intelligent decision-making center adopts a dual-delay deep deterministic policy gradient architecture. The output layer of the actor network in the deep reinforcement learning network and the expert rule base of the rule engine are dynamically integrated through a gating mechanism. The control policy optimization process records the mapping relationship between the geometric features of objects and the grasping posture in historical sorting data, and generates interpretive decision logs for visualization by the three-dimensional projection device of the human-computer interaction mechanism.

10. A dual-station robotic sorting optimization terminal based on digital twin, based on the dual-station robotic sorting optimization system based on digital twin according to any one of claims 6-9, characterized in that, include: Heterogeneous computing accelerator cards: It is equipped with a neural network processor and a vision processing unit to support FP16 precision computing and hardware-level sparsity acceleration; Safety isolation device: Unidirectional data diodes and physical isolation transformers are used to ensure complete isolation between the control network and the information network; Hot-swappable storage array: Configure a RAID6 redundant disk array and enterprise-grade SSD cache to support real-time mirror backup and fast rollback of sorting process data.

Images