A pose alignment correction system for a mobile robot

Abstract

This invention discloses a positioning posture correction system for mobile robots, relating to the field of mobile robot technology. It includes a posture acquisition module for acquiring the current actual pose state of the mobile robot; a motion simulation module for inputting the current actual pose state into a digital twin model, simulating a set of candidate correction actions, and selecting a safe action set based on the simulation results; a motion output module for inputting the safe action set into a reinforcement learning policy network and outputting the final execution action; and a motion execution module for controlling the mobile robot to execute the final action. This invention fundamentally avoids collision and instability risks; achieves continuous self-evolution of the policy and model during use; and significantly enhances the adaptability and robustness to dynamic environments, system nonlinearity, and unknown disturbances, enabling mobile robots to reliably complete precise positioning operations in complex and demanding scenarios.

Description

Technical Field

[0001] This invention relates to the field of mobile robot technology, and in particular to a positioning and attitude correction system for mobile robots. Background Technology

[0002] With the rapid development of intelligent manufacturing, smart logistics, and flexible production systems, mobile robots (such as AGVs and AMRs) have become core equipment for tasks such as material handling, component assembly, and equipment docking. Precise "alignment"—that is, the mobile robot adjusting its end effector or body to a state that precisely matches the target position and posture—is the prerequisite and key to completing these advanced tasks. Traditional mobile robot alignment methods mainly rely on trajectory tracking based on fixed paths, visual servoing based on markers (such as QR codes or reflectors), or closed-loop control based on simple threshold judgments. Existing technologies face many bottlenecks in practical industrial applications: First, in dynamic or unstructured environments, preset paths may fail due to temporary obstacles or changes in ground conditions. Second, visual systems based on fixed markers may fail when markers are obscured, lighting changes drastically, or the environment is reflective. Firstly, positioning accuracy is severely reduced or even fails. Secondly, traditional control methods (such as PID control) are usually based on simplified linear models, which are difficult to handle the nonlinearity of robot dynamics, actuator delays, and complex contact forces generated when interacting with the environment. This can lead to oscillations, overshoots, or even collisions or jamming in contact tasks. Although some studies have attempted to introduce machine learning (such as reinforcement learning) to optimize the decision-making of mobile robots, direct trial and error training on physical robots is costly and risky, and the learning process lacks safety guarantees, making it difficult to meet the stringent reliability and safety requirements of industrial applications. The modules (perception, decision-making, and execution) of existing systems are often designed in isolation, lacking collaborative optimization and closed-loop adaptive capabilities. When the mechanical characteristics of the robot wear down over time or the working environment drifts, the system performance will gradually degrade.

[0003] However, current common solutions have many drawbacks, including: existing technology decision-making mechanisms rely on simple rules and lack forward-looking safety assessments, which can easily lead to collisions or instability; control methods are based on fixed models, which are difficult to adapt to nonlinear dynamics, time-varying systems and complex contacts, resulting in poor robustness; machine learning training requires direct trial and error on physical robots, which is costly, risky and lacks safety guarantees; and the system modules are isolated, lacking collaborative optimization and closed-loop calibration, so the overall performance cannot be continuously improved. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the problems existing in the current alignment and posture correction system for mobile robots, the present invention is proposed.

[0006] Therefore, the purpose of this invention is to provide a positioning and posture correction system for mobile robots, which is applicable to solving the problems of existing technology decision-making mechanisms relying on simple rules, lacking forward-looking safety assessments, and being prone to collisions or instability; control methods based on fixed models, which are difficult to adapt to nonlinear dynamics, time-varying systems and complex contacts, resulting in poor robustness; machine learning training requires direct trial and error on physical robots, which is costly, risky and lacks safety guarantees; and the system modules are isolated, lacking collaborative optimization and closed-loop calibration, making it impossible to continuously improve the overall performance.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a positioning posture correction system for a mobile robot, comprising: a posture acquisition module for acquiring the current actual posture state of the mobile robot; an action simulation module for inputting the current actual posture state into a digital twin model, simulating a set of candidate correction actions, and selecting a set of safe actions based on the simulation results; an action output module for inputting the set of safe actions into a reinforcement learning policy network and outputting the final execution action; and an action execution module for controlling the mobile robot to execute the final execution action.

[0008] As a preferred embodiment of the alignment posture correction system for mobile robots described in this invention, the reinforcement learning policy network is trained in the following manner: pre-trained in a simulation environment constructed by the digital twin model; the pre-trained network is deployed on the physical robot, and the reward function or parameters of the network are optimized online based on the feedback signal of its actual success rate in performing the alignment task.

[0009] As a preferred embodiment of the alignment and posture correction system for mobile robots described in this invention, the actual posture state is acquired through a visual sensor, a laser sensor, a radio positioning system, and an inertial sensor.

[0010] As a preferred embodiment of the alignment and posture correction system for a mobile robot according to the present invention, the specific content of the motion simulation module is as follows: Based on the current actual posture state, a set of candidate correction actions with different motion directions and speeds are preset in the digital twin model; dynamic forward simulation is performed on each candidate correction action, and the contact force / torque change curve of the end effector, the center of gravity stability margin of the robot body, and the interference with the known environment model are predicted simultaneously; according to the preset risk quantification rules, the predicted contact force / torque, center of gravity stability margin, and interference are calculated respectively to generate a comprehensive risk score corresponding to each candidate action; candidate actions with a comprehensive risk score lower than a first safety threshold and a center of gravity stability margin higher than a second safety threshold are included in the safe action set.

[0011] As a preferred embodiment of the alignment and posture correction system for a mobile robot according to the present invention, the specific content of the action output module is as follows: the first input terminal of the network receives the description vector of each action in the safe action set, the description vector including at least the basic motion parameters of the action; the second input terminal of the network receives the multi-dimensional risk feature vector corresponding to each safe action provided by the action simulation module, the vector encoding at least the predicted collision risk, stability score and task completion estimate; the network includes an attention mechanism layer for calculating the attention weight of each action in the safe action set, the weight being the result of the action description vector, the current robot state vector and the risk feature vector working together; the output layer of the network performs a weighted evaluation of the safe actions based on the attention weight, outputting the action with the highest comprehensive score as the final execution action or outputting a probability distribution of an action for sampling execution.

[0012] As a preferred embodiment of the alignment and posture correction system for a mobile robot described in this invention, the specific contents of the motion execution module are as follows: It receives the discrete instructions for the final executed action, combines them with the robot's dynamic constraints, and interpolates and smooths them into a continuous, executable body motion trajectory; a feedforward-feedback composite controller is used to drive the robot to execute the motion trajectory; the feedforward controller outputs the ideal driving force / torque required to execute the trajectory based on the prediction of the digital twin model; the feedback controller fine-tunes the trajectory based on the pose and force / torque information fed back in real time during robot execution; and an execution status monitor monitors in real time the deviation between the robot's actual motion trajectory and the desired trajectory, as well as the current status of the drive motor; when the deviation exceeds a dynamic threshold or the current is abnormal, a predefined safety response is triggered.

[0013] As a preferred embodiment of the alignment and attitude correction system for a mobile robot described in this invention, the predefined safety response is specifically a graded response strategy, as follows: Level 1 response: When the deviation or abnormal current value is in the first interval, the feedback controller enhances the correction force and dynamically relaxes the end-positioning accuracy requirements of the motion trajectory to prioritize a smooth transition; Level 2 response: When the deviation or abnormal current value enters the more severe second interval, the current trajectory tracking is immediately paused, and the robot is controlled to decelerate and stop along the tangent of the current motion direction; Level 3 response: When a sudden increase in current is detected indicating a possible physical collision, the power output is immediately cut off, and a rapid retreat action based on the reverse dynamics model is triggered, causing the robot's end effector to retreat in the opposite direction of the intrusion direction.

[0014] As a preferred embodiment of the alignment and posture correction system for a mobile robot described in this invention, the output of the feedforward controller is continuously optimized through an online learning unit; the online learning unit is configured to: collect and compare the output torque of the feedforward controller with the torque actually compensated by the feedback controller to eliminate tracking deviation; establish a regression model with the robot's motion state as input and the torque deviation as output; and use the output of the regression model to dynamically compensate the output of the feedforward controller in subsequent similar motion states.

[0015] As a preferred embodiment of the alignment and posture correction system for a mobile robot described in this invention, the actual motion data collected by the execution status monitor of the motion execution module is fed back to the motion simulation module for online parameter calibration of the digital twin model. Specifically, the actual motion data is compared with the data predicted by the digital twin model when simulating the corresponding action, and the key dynamic parameters in the digital twin model are adjusted by an optimization algorithm so that the predicted output of the model approximates the response of the actual physical system.

[0016] As a preferred embodiment of the alignment and posture correction system for a mobile robot described in this invention, the action execution module, when executing the motion trajectory, possesses online trajectory adaptation capability based on real-time environmental perception. Specifically, when the posture acquisition module perceives an unforeseen temporary obstacle on the preset motion path, the trajectory shaper, without changing the terminal pose specified by the final executed action, performs local smoothing replanning on the currently executed motion trajectory to generate a new local trajectory that bypasses the obstacle and satisfies the robot's dynamics constraints, which is then switched and executed by the feedforward-feedback composite controller.

[0017] In a second aspect, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the alignment posture correction system for a mobile robot as described in the first aspect of the present invention.

[0018] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the alignment posture correction system for a mobile robot as described in the first aspect of the present invention.

[0019] The beneficial effects of this invention are as follows: This invention uses digital twin simulation to perform forward-looking safety screening and quantitative evaluation of candidate actions, and combines reinforcement learning that integrates risk features for intelligent decision-making, thus constructing a "virtual pre-training-safe decision-making" mechanism, fundamentally avoiding collision and instability risks; through a dual closed loop of "simulation pre-training-physical online optimization" and "execution data feedback calibration model," the strategy and model achieve continuous self-evolution during use; the system significantly enhances its adaptability and robustness to dynamic environments, system nonlinearity, and unknown disturbances by leveraging multi-source perception, feedforward-feedback composite control, and online trajectory replanning capabilities; the deep collaboration among the various modules of the system integrates safety, intelligence, accuracy, and adaptability, enabling the mobile robot to reliably complete precise alignment operations in complex and demanding scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the implementation of the present invention in Example 1.

[0021] Figure 2 This is a dynamic adjustment diagram of the present invention in Example 1. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Example 1 Reference Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a positioning posture correction system for a mobile robot, comprising the following steps: S1: Attitude acquisition module, used to acquire the current actual pose state of the mobile robot.

[0026] Specifically, the actual pose is obtained through visual sensors, laser sensors, radio positioning systems, and inertial sensors.

[0027] Preferably, digital twin technology is used to perform high-fidelity, multi-physics (dynamics, contact mechanics, geometric interference) forward-looking simulation and quantitative risk assessment of candidate correction actions, constructing a "virtual safety fence" at the front end of the decision chain. This transforms the traditional post-event reactive safety mechanism into a pre-event predictive safety screening, fundamentally preventing dangerous actions that may lead to physical collisions, instability, or equipment overload from entering the execution stage. The richly structured risk feature vector it generates provides key physical prior knowledge for downstream reinforcement learning decision-making, enabling the agent to understand the potential consequences of actions, thereby significantly improving the reliability of decision-making and learning efficiency, and achieving a deep integration of model-based safety constraints and model-free learning capabilities.

[0028] For example, in a precision electronic component assembly scenario, a mobile robot needs to accurately place chips into circuit board slots. The attitude acquisition module uses vision and laser fusion positioning to determine that the robot's current posture has a millimeter-level lateral deviation. The motion simulation module then simulates various candidate correction actions such as "rapid straight rush", "slow approach", and "with angle fine adjustment" in a digital twin environment. The simulation prediction shows that although the "rapid straight rush" action is fast, the end contact force will exceed the limit of the chip pin; the "with angle fine adjustment" action may have insufficient stability margin because the center of gravity projection is close to the edge of the supporting polygon. Finally, based on the quantitative risk score, the system only includes a few actions with stable contact force and high stability, such as "slow approach", into the safe action set, thus eliminating the risk of assembly damage caused by overshoot or shaking from the source.

[0029] S2: Action simulation module, used to input the current actual pose state into the digital twin model, simulate a set of candidate correction actions, and select a set of safe actions based on the simulation results.

[0030] Preferably, the specific content of the motion simulation module is as follows: Based on the current actual pose state, a set of candidate correction actions with different motion directions and speeds are preset in the digital twin model; For each candidate correction action, a forward dynamic simulation is performed, and the contact force / torque change curve of the end effector, the center of gravity stability margin of the robot body, and the interference with the known environment model are predicted simultaneously. Based on the preset risk quantification rules, the predicted contact force / torque, center of gravity stability margin and interference are calculated respectively to generate a comprehensive risk score for each candidate action; Candidate actions with a comprehensive risk score below the first safety threshold and a center of gravity stability margin above the second safety threshold will be included in the safe action set.

[0031] Preferably, an attention mechanism decision network integrating risk priors is designed, combined with a two-stage training paradigm of "simulation pre-training - physical online optimization". The attention mechanism enables the reinforcement learning strategy to dynamically and interpretably weigh candidate solutions within the safe action set according to the current state, intelligently balancing efficiency and risk, and achieving optimal performance decision-making within the absolute safety boundary. The phased transfer training strategy cleverly solves the transfer problem of "simulation to reality" in robot reinforcement learning and the safety risks of physical training. It efficiently accumulates experience through the simulation environment, and then performs precise fine-tuning based on the successful feedback of high-order tasks in the physical environment, thereby realizing the rapid and safe deployment and adaptive optimization of the strategy in the real world.

[0032] For example, to train a robot to perform alignment operations on a swaying platform (simulating a transport vehicle), the system first constructs a simulation model in a digital twin environment that includes random vibrations of the platform, and pre-trains the policy network millions of times to teach it a preliminary anti-disturbance and correction strategy. After being deployed to a physical experimental platform, the robot begins real-world operations. Initially, due to the difference between the real vibration spectrum and the simulation, the success rate is about 70%. Subsequently, the system adjusts the reward function of the policy network online based on the binary result of each attempt: "successful insertion" or "failure to align." After dozens of iterations, the policy network gradually focuses on learning the most robust actions under specific frequency disturbances, increasing the success rate of actual tasks to over 98%, achieving an efficient and safe transfer from virtual experience to the physical world.

[0033] S3: Action Output Module, used to input the set of safety actions into the reinforcement learning policy network and output the final action to be executed.

[0034] Preferably, the reinforcement learning policy network is trained in the following ways: Pre-training is performed in a simulation environment for building the digital twin model; The pre-trained network is deployed on a physical robot, and the reward function or parameters of the network are optimized online based on the feedback signal of its actual success rate in performing alignment tasks.

[0035] Furthermore, the specific content of the action output module is as follows: The first input terminal of the network receives the description vectors of each action in the set of security actions. The description vectors include at least the basic motion parameters of the actions. The second input of the network receives a multi-dimensional risk feature vector corresponding to each safety action provided by the action simulation module. This vector encodes at least the predicted collision risk, stability score, and task completion estimate. The network contains an attention mechanism layer to calculate the attention weight of each action in the safety action set. This weight is the result of the interaction between the action description vector, the current robot state vector, and the risk feature vector. The network's output layer performs a weighted evaluation of security actions based on attention weights, outputting either the action with the highest overall score as the final action to be executed, or outputting a probability distribution of an action for sampling and execution.

[0036] Preferably, a dual adaptive closed loop of online learning at the execution layer and dynamic calibration of digital twin model parameters is established, which endows the system with the ability to continuously self-optimize and maintain long-term performance robustness. The online learning unit of the feedforward controller compensates for the time-varying characteristics of the system (such as mechanical wear and load changes) and unmodeled dynamics in real time through data-driven methods, so that the trajectory tracking accuracy improves over time. The synchronously running model calibration closed loop uses physical execution data to continuously correct the dynamic parameters of the digital twin model, making its predictive ability increasingly accurate. These two closed loops promote each other and form a positive feedback enhancement loop that makes the system more "intelligent" and more "precise" with use.

[0037] For example, a mobile robot used for stacking boxes in a warehouse, after working continuously for several months, experiences slight wear on its drive wheels, causing constant yaw drift during straight-line movement. During execution, the feedforward controller's online learning unit detects that, in order to achieve a straight trajectory, the feedback controller needs to continuously compensate with an additional torque. The unit associates this "deviation torque" with the current speed and updates the internal regression model. Subsequently, a pre-compensation torque to counteract this drift is automatically added to the feedforward output, restoring the trajectory accuracy to its original state. At the same time, the continuous yaw data recorded by the execution status monitor is fed back to the digital twin model, which automatically increases the asymmetry coefficient of its wheel-ground friction parameters, making future simulation predictions of new actions more accurate, thus forming an adaptive loop from physical wear to model correction.

[0038] S4: Action Execution Module, used to control the mobile robot to perform the final action.

[0039] Preferably, the specific content of the action execution module is as follows: Receive discrete instructions for the final action, combine them with the robot's dynamic constraints (maximum acceleration, jerk), interpolate and smooth them into a continuous, executable body motion trajectory; Feedforward-feedback composite controller: used to drive the robot to execute motion trajectories; The feedforward controller outputs the ideal driving force / torque required to execute the trajectory, as predicted by the digital twin model. The feedback controller fine-tunes the trajectory based on the pose (provided by the pose acquisition module) and force / torque information fed back in real time during the robot's execution. Execution status monitor: Real-time monitoring of the deviation between the robot's actual motion trajectory and the desired trajectory, as well as the current status of the drive motors; When the deviation exceeds the dynamic threshold or the current is abnormal, a predefined safety response is triggered.

[0040] Specifically, the predefined security response, specifically the tiered response strategy, is detailed below: Level 1 response: When the deviation or abnormal current value is in the first range, the feedback controller enhances the correction force and dynamically relaxes the end positioning accuracy requirements of the motion trajectory to prioritize a smooth transition. Level 2 response: When the deviation or abnormal current value enters the more serious second interval, immediately pause the current trajectory tracking and control the robot to decelerate and coast along the tangent of the current direction of motion; Level 3 response: When a sudden surge in current is detected, indicating a possible physical collision, the power output is immediately cut off, and a rapid retraction action based on the reverse dynamics model is triggered, causing the robot end effector to exit in the opposite direction of the intrusion.

[0041] Specifically, the output of the feedforward controller is continuously optimized through an online learning unit; The online learning unit is configured to collect and compare the output torque of the feedforward controller with the torque actually compensated by the feedback controller to eliminate tracking error; Establish a regression model with robot motion state (including velocity, acceleration, and trajectory curvature) as input and torque deviation as output; The output of this regression model is used to dynamically compensate the output of the feedforward controller under subsequent similar motion states.

[0042] Specifically, the actual motion data collected by the execution status monitor of the motion execution module is fed back to the motion simulation module for online parameter calibration of the digital twin model. The details are as follows: The actual motion data (pose, velocity, current) is compared with the data predicted by the digital twin model when simulating the corresponding action. The key dynamic parameters in the digital twin model (such as wheel-ground friction coefficient and motor response time constant) are adjusted by the optimization algorithm so that the model's predicted output approximates the response of the actual physical system.

[0043] Furthermore, the action execution module possesses online trajectory adaptation capabilities based on real-time environmental awareness when executing motion trajectories, as detailed below: When the attitude acquisition module detects an unforeseen temporary obstacle on the preset motion path; The trajectory shaper performs local smoothing replanning on the current motion trajectory without changing the terminal pose specified by the final execution action, generating a new local trajectory that bypasses obstacles and satisfies the robot's dynamic constraints, and is then switched and executed by the feedforward-feedback composite controller.

[0044] Preferably, a graded safety response mechanism and online trajectory adaptive capability are implemented during the execution process, which significantly enhances the system's resilience and flexibility in the face of uncertainty and sudden interference. The graded safety response mechanism takes differentiated measures (from adjusting control parameters to emergency rollback) according to the severity of the deviation, avoiding the interruption of production cycle caused by the traditional single emergency stop strategy. It maximizes the smoothness and continuity of operation while ensuring safety. The online trajectory adaptive capability allows the system to adjust the motion trajectory locally in real time and smoothly without changing the final alignment target when encountering unforeseen dynamic obstacles. This combines the advantages of open-loop planning with the agility of closed-loop response, greatly expanding the reliable working range of the system in unstructured dynamic environments.

[0045] For example, when the robot is performing the alignment task of feeding material to the machine tool, a small, scattered part suddenly rolls into the path ahead of it (recognized by the 3D vision of the attitude acquisition module). The motion execution module immediately initiates online adaptation: the trajectory shaper calculates a smooth local arc trajectory around the part within milliseconds without changing the final feeding target pose, and ensures that the acceleration meets the constraints. At the same time, due to the trajectory change caused by replanning, the execution status monitor detects an increase in tracking deviation and triggers a first-level response: the feedback controller gain is automatically increased, and the end effector is temporarily allowed to have a slightly larger tolerance on the horizontal plane to prioritize the smoothness of the bypassing motion. Once the obstacle is bypassed and the deviation returns to normal, the system resumes the high-precision tracking mode and finally completes the feeding accurately.

[0046] In summary, this invention utilizes digital twin simulation for forward-looking safety screening and quantitative evaluation of candidate actions, and combines reinforcement learning with risk features for intelligent decision-making, constructing a "virtual pre-training - safety decision-making" mechanism that fundamentally avoids collision and instability risks. Through a dual closed loop of "simulation pre-training - physical online optimization" and "execution data feedback calibration model," the strategy and model achieve continuous self-evolution during use. Leveraging multi-source perception, feedforward-feedback composite control, and online trajectory replanning capabilities, the system significantly enhances its adaptability and robustness to dynamic environments, system nonlinearity, and unknown disturbances. The deep collaboration among the system's modules integrates safety, intelligence, accuracy, and adaptability, enabling the mobile robot to reliably complete precise alignment operations in complex and demanding scenarios.

[0047] Example 2 is an embodiment of the present invention, which differs from the previous embodiment in that: If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0048] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0049] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0050] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0051] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A positioning and attitude correction system for a mobile robot, characterized in that: include: The attitude acquisition module is used to acquire the current actual pose state of the mobile robot; The motion simulation module is used to input the current actual pose state into the digital twin model, simulate a set of candidate correction actions, and select a set of safe actions based on the simulation results. The action output module is used to input the set of safety actions into the reinforcement learning policy network and output the final execution action. The action execution module is used to control the mobile robot to perform the final execution action.

2. The alignment and attitude correction system for a mobile robot as described in claim 1, characterized in that: The reinforcement learning policy network is trained in the following manner: Pre-training is performed in the simulation environment where the digital twin model is constructed; The pre-trained network is deployed on a physical robot, and the reward function or parameters of the network are optimized online based on the feedback signal of its actual success rate in performing alignment tasks.

3. The alignment and attitude correction system for a mobile robot as described in claim 1, characterized in that: The actual pose state is obtained through visual sensors, laser sensors, radio positioning systems, and inertial sensors.

4. The alignment and attitude correction system for a mobile robot as described in claim 1, characterized in that: The specific contents of the motion simulation module are as follows: Based on the current actual pose state, a set of candidate correction actions with different motion directions and speeds are preset in the digital twin model; For each candidate correction action, a forward dynamic simulation is performed, and the contact force / torque change curve of the end effector, the center of gravity stability margin of the robot body, and the interference with the known environment model are predicted simultaneously. Based on the preset risk quantification rules, the predicted contact force / torque, center of gravity stability margin, and interference are calculated respectively to generate a comprehensive risk score for each candidate action; Candidate actions with a comprehensive risk score lower than the first safety threshold and a center of gravity stability margin higher than the second safety threshold are included in the set of safe actions.

5. The alignment and attitude correction system for a mobile robot as described in claim 1, characterized in that: The specific contents of the action output module are as follows: The first input terminal of the network receives the description vector of each action in the security action set, and the description vector includes at least the basic motion parameters of the action; The second input terminal of the network receives a multi-dimensional risk feature vector corresponding to each safety action provided by the action simulation module. This vector encodes at least the predicted collision risk, stability score, and task completion estimate. The network contains an attention mechanism layer for calculating the attention weight of each action in the safety action set. This weight is the result of the interaction between the action description vector, the current robot state vector, and the risk feature vector. The output layer of the network performs a weighted evaluation of the safety actions based on the attention weights, and outputs the action with the highest comprehensive score as the final execution action or outputs a probability distribution of an action for sampling and execution.

6. The alignment and attitude correction system for a mobile robot as described in claim 1, characterized in that: The specific contents of the action execution module are as follows: The discrete instructions for the final action are received, and combined with the robot's dynamic constraints, they are interpolated and smoothed into a continuous, executable body motion trajectory. Feedforward-feedback composite controller: used to drive the robot to execute the motion trajectory; The feedforward controller outputs the ideal driving force / torque required to execute the trajectory, as predicted by the digital twin model. The feedback controller fine-tunes the trajectory based on the pose and force / torque information fed back in real time during the robot's execution. Execution status monitor: Real-time monitoring of the deviation between the robot's actual motion trajectory and the desired trajectory, as well as the current status of the drive motors; When the deviation exceeds the dynamic threshold or the current is abnormal, a predefined safety response is triggered.

7. The alignment and attitude correction system for a mobile robot as described in claim 6, characterized in that: The predefined security response is specifically a tiered response strategy, the details of which are as follows: Level 1 response: When the deviation or abnormal current value is in the first range, the feedback controller enhances the correction force and dynamically relaxes the end positioning accuracy requirement of the motion trajectory to prioritize a smooth transition; Level 2 response: When the deviation or abnormal current value enters the more serious second interval, immediately pause the current trajectory tracking and control the robot to decelerate and coast along the tangent of the current direction of motion; Level 3 response: When a sudden surge in current is detected, indicating a possible physical collision, the power output is immediately cut off, and a rapid retraction action based on the reverse dynamics model is triggered, causing the robot end effector to exit in the opposite direction of the intrusion.

8. The alignment and attitude correction system for a mobile robot as described in claim 6, characterized in that: The output of the feedforward controller is continuously optimized through an online learning unit; The online learning unit is configured to collect and compare the output torque of the feedforward controller with the torque actually compensated by the feedback controller to eliminate tracking deviation; Establish a regression model with the robot's motion state as input and the torque deviation as output; The output of the regression model is used to dynamically compensate the output of the feedforward controller under subsequent similar motion states.

9. The alignment and attitude correction system for a mobile robot as described in claim 6, characterized in that: The actual motion data collected by the execution status monitor of the motion execution module is fed back to the motion simulation module for online parameter calibration of the digital twin model, as detailed below: By comparing actual motion data with the data predicted by the digital twin model when simulating the corresponding action, the key dynamic parameters in the digital twin model are adjusted through optimization algorithms so that the model's predicted output approximates the response of the actual physical system.

10. The alignment and attitude correction system for a mobile robot as described in claim 6, characterized in that: When executing the motion trajectory, the action execution module has the ability to adapt the trajectory online based on real-time environmental awareness, as detailed below: When the attitude acquisition module senses an unforeseen temporary obstacle on the preset motion path; The trajectory shaper performs local smoothing replanning on the currently executing motion trajectory without changing the terminal pose specified by the final execution action, generating a new local trajectory that bypasses obstacles and satisfies robot dynamics constraints, and is then switched and executed by the feedforward-feedback composite controller.

Images