A mobile robot control method and system

Abstract

This invention discloses a mobile robot control method and system, relating to the field of intelligent motion control technology for mobile robots. The method includes: acquiring environmental feature data and motion state data; performing preprocessing and feature fusion to generate standardized state data; constructing a reinforcement learning training environment in an offline simulation environment to obtain a converged reinforcement learning policy model; inputting the current standardized state data of the robot into the reinforcement learning policy model to obtain and output preliminary motion control commands and reference trajectories; establishing a predictive control model (MPC), inputting the current standardized state data of the robot, performing short-term motion prediction and optimization, and solving for the final control commands that satisfy the constraints; and distributing the commands to four Mecanum wheel drivers to adjust the speed and direction of each wheel. This invention achieves high-precision trajectory tracking, adaptive obstacle avoidance, and energy efficiency optimization for mobile robots in complex dynamic environments, improving system stability and intelligence.

Description

Technical Field

[0001] This invention relates to the field of intelligent motion control technology for mobile robots, specifically a mobile robot control method and system. Background Technology

[0002] With the widespread application of mobile robots in smart warehousing, logistics distribution, and industrial manufacturing, omnidirectional mobile robots based on Mecanum wheels have become a hot topic in academia and industry due to their flexible movement capabilities and spatial adaptability. In recent years, with the improvement of sensor accuracy and the rapid development of artificial intelligence technology, the autonomous perception, path planning, and motion control methods of mobile robots have been continuously advancing. Traditional motion control methods mainly rely on PID, fuzzy control, and classical kinematic / dynamic modeling, while some cutting-edge research has begun to introduce intelligent algorithms such as deep learning and reinforcement learning to improve the robot's adaptability to complex environments and its decision-making intelligence. Especially in real-world scenarios with dynamic environmental changes, dense obstacles, and precision operations, higher demands are placed on the autonomy, robustness, and accuracy of robot control strategies.

[0003] However, existing mobile robot control technologies generally suffer from several limitations. On the one hand, traditional motion control methods rely on precise model assumptions and manual parameter tuning, making it difficult to cope with dynamic disturbances and high-dimensional inputs in complex unstructured environments, resulting in insufficient trajectory tracking accuracy and global planning capabilities. On the other hand, while end-to-end control methods based on single reinforcement learning possess a certain degree of adaptability, they face practical problems such as a lack of training samples, slow convergence speed, and inability to guarantee safety under extreme conditions, making it difficult to meet the high stability and real-time requirements of industrial-grade deployments. Furthermore, some existing fusion control strategies have significant shortcomings in the effective fusion of kinematic constraints and multi-source environmental perception data, as well as the coordination between global planning and local optimization, failing to achieve efficient, accurate, and safe control of omnidirectional mobile robots under complex tasks.

[0004] Therefore, there is an urgent need for an innovative mobile robot control method that can integrate multimodal perception data, combine global intelligent decision-making with local safety optimization, and possess high robustness and real-time performance, in order to break through the bottlenecks of existing technologies and achieve high-precision autonomous control of the robot's omnidirectional motion in complex scenarios. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed, entitled "A Mobile Robot Control Method and System".

[0006] Therefore, the technical problem solved by this invention is that existing mobile robot motion control methods have difficulties in integrating multi-source perception data to achieve global and local collaborative optimization, lack the ability to adapt to complex dynamic environments, and have insufficient motion control accuracy and safety. The invention also addresses the problem of how to achieve high-precision, robust, and adaptive motion control for omnidirectional mobile robots based on Mecanum wheels in complex environments.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a mobile robot control method, including collecting feature data of the environment in which the positioning object is located and motion state data of the robot body, to obtain environmental feature data and motion state data respectively;

[0009] Environmental feature data and motion state data are preprocessed and feature fused to generate standardized state data;

[0010] Based on standardized state data as input, a reinforcement learning training environment is constructed in an offline simulation environment, and a converged reinforcement learning policy model is obtained through iterative training.

[0011] Real-time acquisition and preprocessing of the current robot's standardized state data, inputting it into the reinforcement learning policy model to obtain and output preliminary motion control commands and reference trajectories;

[0012] A predictive control model (MPC) based on robot kinematics and dynamics constraints is established. The MPC takes the initial motion control command and reference trajectory as reference inputs, and simultaneously inputs the current standardized state data of the robot to perform short-term motion prediction and optimization, and solves to obtain the final control command that satisfies the constraints.

[0013] The final control commands are distributed to the four Mecanum wheel drives to adjust the speed and direction of each wheel, thereby enabling the robot to move in all directions with three degrees of freedom.

[0014] As a preferred embodiment of the mobile robot control method of the present invention, the method of collecting feature data of the environment in which the positioning object is located and motion state data of the robot body includes collecting information on obstacle distribution, ground type, spatial geometry, ambient light intensity and dynamic changes in the environment to obtain environmental feature data.

[0015] The robot's position, posture, speed, and the rotational speed of its four drive wheels are collected to obtain motion state data.

[0016] As a preferred embodiment of the mobile robot control method of the present invention, the preprocessing and feature fusion of environmental feature data and motion state data includes performing noise reduction, filtering, outlier removal, normalization, and time synchronization on the environmental feature data and motion state data respectively, and then performing feature splicing and dimension unification on the processed data according to a preset format to fuse them into standardized state data.

[0017] As a preferred embodiment of the mobile robot control method of the present invention, the step of obtaining a converged reinforcement learning policy model through iterative training includes setting up a reinforcement learning training system in a simulation environment that includes robot state, action space and reward function, taking standardized state data as input, iteratively training the robot motion control policy through a reinforcement learning algorithm, continuously optimizing the policy parameters until the cumulative reward value reaches the preset convergence condition, and finally obtaining a reinforcement learning policy model that can be used for robot motion control.

[0018] As a preferred embodiment of the mobile robot control method of the present invention, the following steps are included: real-time acquisition and preprocessing of the current robot standardized state data; during the actual operation of the robot, periodic acquisition of current environmental feature data and motion state data; generation of standardized state data for the current period according to a preset preprocessing process; input of the standardized state data for the current period into a reinforcement learning strategy model; and output of the corresponding preliminary motion control commands and reference trajectory information by the reinforcement learning strategy model.

[0019] As a preferred embodiment of the mobile robot control method described in this invention, the establishment of the predictive control model MPC based on robot kinematics and dynamic constraints includes establishing kinematic models describing the relationship between robot displacement, velocity, acceleration and the rotational speed of each wheel, and dynamic models describing the robot's dynamic behavior, based on the structural parameters, motion mode and force relationship of the mobile robot body.

[0020] Based on the aforementioned kinematic and dynamic models, the robot's position, attitude, velocity, and acceleration are selected as state variables to construct a predictive control optimization problem. Position constraints, attitude constraints, velocity constraints, acceleration constraints, wheel speed constraints, and driving force constraints are set in the optimization problem.

[0021] As a preferred embodiment of the mobile robot control method of the present invention, the step of obtaining the final control command that satisfies the constraints includes, in each control cycle, taking the preliminary motion control command and reference trajectory as target inputs, and taking the standardized state data of the current cycle as state inputs, using the predictive control model MPC to predict the motion state and optimize the control sequence in the future time period based on the set physical constraints, using a numerical optimization algorithm to obtain the optimal control command sequence, and selecting the first control command from the optimal control command sequence as the final control command output of the current cycle, which is used to drive the robot to perform actual motion execution.

[0022] As a preferred embodiment of the mobile robot control method described in this invention, the step of distributing the final control command to the four Mecanum wheel drivers includes: decomposing the final control command into velocity components along the X-axis, along the Y-axis, and rotating about the vertical axis according to the kinematics of the robot chassis; calculating the target rotation speed and rotation direction of each drive wheel using an inverse kinematics algorithm based on the installation parameters and rolling direction of the four Mecanum wheels; and sending the target rotation speed and rotation direction of each drive wheel to the corresponding driver to control the four Mecanum wheels to work together to achieve three-degree-of-freedom omnidirectional motion of the robot in the plane.

[0023] Secondly, embodiments of the present invention provide a mobile robot control system, including:

[0024] Data acquisition module: Collects feature data of the environment in which the positioning object is located and motion state data of the robot body, and obtains environmental feature data and motion state data respectively;

[0025] Data preprocessing and feature fusion module: preprocesses and fuses environmental feature data and motion state data to generate standardized state data;

[0026] Reinforcement learning policy model training module: Based on standardized state data as input, a reinforcement learning training environment is built in an offline simulation environment, and a converged reinforcement learning policy model is obtained through iterative training;

[0027] Reinforcement learning reasoning and instruction generation module: collects and preprocesses the current robot standardized state data in real time, inputs it into the reinforcement learning policy model, and obtains and outputs preliminary motion control instructions and reference trajectories;

[0028] Predictive control model optimization module: Establish a predictive control model MPC based on robot kinematics and dynamic constraints. The predictive control model MPC takes the initial motion control command and reference trajectory as reference inputs, and simultaneously inputs the current standardized state data of the robot to perform short-term motion prediction and optimization, and solves to obtain the final control command that satisfies the constraint conditions.

[0029] Chassis drive distribution and execution module: Distributes the final control commands to the four Mecanum wheel drives, adjusts the speed and direction of each wheel, and realizes the robot's three-degree-of-freedom omnidirectional motion.

[0030] The beneficial effects of this invention are as follows: By fusing multi-source environmental feature data and robot motion state data to construct standardized state data input, this invention achieves accurate perception of the robot and its environment. Employing a control strategy combining reinforcement learning and predictive control (MPC), it possesses both global autonomous decision-making capabilities and local safety optimization performance. This not only improves the robot's motion control accuracy and path tracking robustness in complex and dynamic environments but also effectively ensures the system's real-time performance and operational safety. Compared to existing technologies, this invention enables efficient autonomous control of a Mecanum wheel-based omnidirectional mobile robot in multi-dimensional, highly dynamic scenarios, significantly expanding the robot's adaptability and engineering application value in demanding applications such as intelligent warehousing and automated handling. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the 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:

[0032] Figure 1 The above is an overall flowchart of a mobile robot control method provided in the first embodiment of the present invention. Detailed Implementation

[0033] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0034] Example 1, referring to Figure 1 As one embodiment of the present invention, a mobile robot control method is provided, comprising:

[0035] S1: Collect feature data of the environment in which the positioning object is located and motion state data of the robot body to obtain environmental feature data and motion state data respectively.

[0036] The system collects information on obstacle distribution, ground type, spatial geometry, ambient light intensity, and dynamic environmental changes to obtain environmental characteristic data; it also collects the robot's position, posture, speed, and the rotational speed of its four drive wheels to obtain motion state data.

[0037] Step S1 of this invention clearly defines the specific content of collecting environmental feature data and robot motion state data, significantly enhancing the foundation and engineering feasibility of subsequent control methods. Specifically, environmental feature data includes obstacle distribution (i.e., spatial coordinates and boundaries of obstacles in the environment detected by LiDAR or vision sensors), ground type (e.g., flat, ramps, steps, distinguished by force sensors or visual analysis), spatial geometry (e.g., shape information of walls, corners, openings, etc. in space), ambient light intensity (which can be monitored in real time by a light sensor), and dynamic environmental change information (e.g., the appearance / disappearance of moving obstacles, personnel flow, etc., achieved through dynamic visual tracking). The robot motion state data includes position information (usually obtained by an encoder or fused SLAM to obtain the robot's global / local coordinates), attitude information (obtained by an IMU or gyroscope to obtain the robot's current rotation angle or attitude quaternion), velocity information (calculated by the encoder and IMU to determine the current linear velocity and angular velocity), and rotational speed information of the four drive wheels (independently collected by wheel speed sensors or encoders to obtain the current rotational speed of each wheel).

[0038] The core of this step lies in the full acquisition and structured management of multimodal sensing data, providing a high-dimensional and rich input foundation for subsequent feature fusion and intelligent control algorithms. Unlike existing technologies that rely on a single source (such as odometry or monocular vision for motion state estimation), this invention effectively improves the accuracy and robustness of state observation through the joint acquisition of multi-source heterogeneous data. For example, the joint acquisition of obstacle distribution and spatial structure significantly improves the feasibility assessment of obstacle avoidance and path planning; the synchronous acquisition of information such as ground type and light intensity provides crucial sensing input for subsequent adaptive motion control (such as slip detection and slip compensation). The acquisition of independent drive wheel rotation speeds not only directly monitors the actual execution of each wheel but also allows for real-time detection of wheel differences and mechanical faults, improving system health management and motion consistency capabilities.

[0039] S2: Preprocess and fuse environmental feature data and motion state data to generate standardized state data.

[0040] The environmental feature data and motion state data are subjected to denoising, filtering, outlier removal, normalization, and time synchronization. The processed data are then combined with features and dimensions according to a preset format to form standardized state data.

[0041] In this embodiment of the invention, environmental feature data and motion state data are explicitly subjected to denoising, filtering, outlier removal, normalization, and temporal synchronization processing, respectively. Then, standardized state data is formed through feature concatenation and dimensional unification. This process addresses the heterogeneity of multi-source sensing data and the uncertainties inherent in the actual data acquisition process, ensuring the validity and comparability of the input data. Specifically, denoising (such as median filtering and Kalman filtering) can effectively eliminate transient interference, sudden pulses, and high-frequency noise in the acquired signal; filtering smooths the signal for specific noise patterns such as low-pass and high-pass filters, improving signal quality; outlier removal automatically removes outliers that exceed the normal distribution range by setting statistical thresholds, sliding windows, or machine learning discrimination methods, preventing extreme errors from affecting the overall judgment; normalization unifies data with different dimensions and value ranges to [0,1] or zero mean unit variance intervals through linear normalization, Z-score standardization, etc., which helps to eliminate dimensional interference between various features, making subsequent feature fusion and model input more efficient and robust; time-series synchronization ensures effective alignment of multi-source data at the same time point and sampling frequency through interpolation alignment or distribution alignment algorithms, preventing information lag or misalignment.

[0042] Feature concatenation and dimensionality unification further concatenate all the preprocessed features in an ordered manner according to a set format (such as one-dimensional vectors or feature matrices), ensuring that each standardized state data is complete, without loss, and adaptable to the input requirements of subsequent algorithms. For high-dimensional perceptual features, methods such as principal component analysis (PCA) and linear discriminant analysis (LDA) can be used for dimensionality reduction and unification, effectively compressing redundancy and highlighting key features.

[0043] S3: Based on standardized state data as input, a reinforcement learning training environment is built in an offline simulation environment, and a converged reinforcement learning policy model is obtained through iterative training.

[0044] In a simulation environment, a reinforcement learning training system is set up that includes robot state, motion space, and reward function. Standardized state data is used as input, and the robot motion control strategy is iteratively trained through reinforcement learning algorithms to continuously optimize the strategy parameters until the cumulative reward value reaches the preset convergence condition. Finally, a reinforcement learning strategy model that can be used for robot motion control is obtained.

[0045] In this embodiment of the invention, step S3 essentially involves using standardized state data, which has undergone rigorous preprocessing and feature fusion, as input to build a controllable and highly accurate reinforcement learning training platform in a simulation environment. This training environment first sets the robot's state (such as position, velocity, posture, and environmental features), motion space (selectable control variables for the robot, such as velocity, steering, and wheel speed), and reward function (a function used to measure the merits of each behavioral strategy), enabling the training system to realistically reflect the actual motion process and environmental interaction. Using standardized state data as input not only ensures the standardization and consistency of the training samples but also makes the reinforcement learning model highly adaptable to various environments and working conditions.

[0046] Reinforcement learning algorithms typically include architectures such as Deep Q-Networks (DQNs), Policy Gradient Methods, and Actor-Critic. Through repeated interaction with the simulation environment, the agent selects actions based on its current state, the environment returns a new state and reward value, and the model continuously adjusts its policy parameters based on the accumulated reward. The iterative training process dynamically optimizes the decision network structure by continuously collecting experience and updating the policy, achieving adaptive policy adjustment. Model convergence criteria typically include the accumulated reward value stabilizing at a preset threshold, the policy network output no longer changing significantly, or the simulation task completion rate reaching a specified target.

[0047] Offline simulation training platforms can significantly reduce the time and cost of actual robot physics experiments, avoid equipment damage caused by trial and error under extreme conditions, and improve training efficiency and sample diversity. Simultaneously, by fully utilizing the high-dimensional features of standardized state data, the trained policy model can be compatible with various complex scenarios (such as dense obstacles and dynamic environmental changes), improving the generalization and robustness of decision-making capabilities. Compared with existing data-driven methods based on static policies or relying on a large number of real experiments, the S3 process of this invention not only improves the engineering deployability of the control policy model but also lays the theoretical and data foundation for subsequent online deployment and adaptive adjustment, significantly expanding the technical boundaries and application breadth of intelligent mobile robot motion control.

[0048] S4: Collect and preprocess the current robot's standardized state data in real time, input it into the reinforcement learning policy model, and obtain and output preliminary motion control commands and reference trajectories.

[0049] During the actual operation of the robot, current environmental feature data and motion state data are periodically collected, and standardized state data for the current period is generated according to a preset preprocessing process. The standardized state data for the current period is then input into the reinforcement learning strategy model, which outputs the corresponding preliminary motion control commands and reference trajectory information.

[0050] In this embodiment of the invention, step S4 emphasizes that during the actual deployment and operation phase of the robot, the system can periodically (e.g., every 10ms, 20ms, or 50ms sampling period) collect the latest environmental feature data and robot motion state data, and generate standardized state data for the current period through a preprocessing procedure consistent with the offline training phase (including noise reduction, normalization, and timing synchronization). Here, "standardized state data for the current period" includes not only external features such as obstacle distribution, ground type, spatial geometry, light intensity, and dynamic environmental changes collected in real time, but also internal motion states such as the robot's position, posture, speed, and wheel rotation speeds. This comprehensive, high-frequency data input provides an information foundation for the system to achieve adaptive decision-making in complex and dynamic environments.

[0051] The "standardized state data for the current cycle" is passed as input to the offline-trained reinforcement learning policy model in this step. The reinforcement learning policy model is essentially a function mapper trained by a deep network. It can quickly infer the optimal initial motion control commands (such as target velocity, steering angle, trajectory point sequence, etc.) and reference trajectory based on the current input state. The initial motion control commands are typically global commands for the robot chassis, such as specifying the three-degree-of-freedom velocity or spatial pose transformation for the next cycle, while the reference trajectory is the sequence of spatial path points the robot expects to traverse in the short term. The model inference process does not require complex online searches or model retraining, thus ensuring high real-time performance and deployment efficiency.

[0052] This step achieves efficient and seamless integration between deep reinforcement learning strategies and the robot's actual motion, enabling dynamic and real-time responses to changes in the robot's environment and its own capabilities. Compared to traditional control schemes that rely solely on manual rules or static models, this invention, through real-time input of high-dimensional, standardized multi-source data, ensures that the strategy model can generate adaptive initial control commands under various operating conditions, significantly improving trajectory tracking accuracy, environmental adaptability, and motion robustness. Furthermore, this real-time inference architecture can be deeply integrated with downstream predictive control modules, providing accurate and real-time reference input for subsequent physical constraint optimization and execution control. This mechanism demonstrates high engineering practical value and promising technological prospects in practical applications such as dynamic obstacle avoidance, cooperative motion, and energy consumption optimization, fully proving the feasibility and advancement of this invention.

[0053] S5: Establish a predictive control model MPC based on robot kinematics and dynamic constraints. The predictive control model MPC takes the initial motion control command and reference trajectory as reference inputs, and simultaneously inputs the current standardized state data of the robot to perform short-term motion prediction and optimization, and solves to obtain the final control command that satisfies the constraint conditions.

[0054] Based on the structural parameters, motion mode, and force relationships of the mobile robot body, kinematic models describing the relationship between robot displacement, velocity, acceleration and the rotational speed of each wheel, and dynamic models describing the robot's dynamic behavior are established respectively.

[0055] Based on the aforementioned kinematic and dynamic models, the robot's position, attitude, velocity, and acceleration are selected as state variables to construct a predictive control optimization problem. Position constraints, attitude constraints, velocity constraints, acceleration constraints, wheel speed constraints, and driving force constraints are set in the optimization problem.

[0056] In this embodiment of the invention, the kinematic model specifically refers to a mathematical expression established based on the robot's chassis structure, wheel arrangement, and installation angles, describing the mapping relationship between the robot's global coordinate position, local velocity, acceleration, and the independent rotational speeds of each Mecanum wheel. For example, three-degree-of-freedom omnidirectional motion can be bidirectionally converted between wheel speed and chassis speed (linear velocity and angular velocity) using a coordinate transformation matrix. The dynamic model, based on the robot's mass, moment of inertia, and wheel-ground friction parameters, describes the dynamic response process of each state variable of the robot under input driving force. This model not only considers the robot's inertia but also encompasses non-ideal influences in actual operation such as friction and load changes.

[0057] This step simultaneously introduces both kinematic and dynamic models to achieve a high-precision, multi-level description of the robot's actual motion process, significantly outperforming traditional control methods that only use simplified kinematic modeling. Using the robot's position, attitude, velocity, and acceleration as core state variables, and setting directly controllable physical quantities such as wheel speed and driving force as optimization constraints, the control model can strictly limit the robot's operation within physical limits and safe zones. By comprehensively introducing multi-dimensional constraints such as position, attitude, velocity, and acceleration, high-precision trajectory tracking is ensured, and theoretical guarantees are provided for improving robustness and safety under complex real-world working conditions.

[0058] Furthermore, within each control cycle, the initial motion control command and reference trajectory are used as target inputs, and the standardized state data of the current cycle are used as state inputs. The predictive control model MPC is used to predict the motion state and optimize the control sequence for the future time period based on the set physical constraints. The optimal control command sequence is obtained by solving the numerical optimization algorithm, and the first control command is selected from the optimal control command sequence as the final control command output for this cycle, which is used to drive the robot's actual motion execution.

[0059] It should also be noted that within each control cycle, the system receives preliminary motion control commands and reference trajectories (as the desired target) output by the reinforcement learning module in real time, and inputs the standardized state data of the current cycle, which has undergone fusion preprocessing, into the MPC optimizer. The MPC uses kinematic and dynamic models, combined with the aforementioned physical constraints, to predict the motion state in a future time domain (e.g., within 1-2 seconds), forming a set of optimal control command sequences covering all control moments (e.g., one decision point every 100ms, covering the next 10 steps).

[0060] During the optimization process, numerical optimization algorithms (such as quadratic programming, sequential quadratic programming, interior point method, etc.) are used to solve the constrained control problem, ensuring that the predicted trajectory does not exceed the physical and environmental boundaries at any time. Finally, only the first instruction in the optimal control sequence is selected as the final execution instruction issued to the drive system in the current control cycle. This "rolling optimization-rolling execution" approach not only significantly enhances the system's adaptability to dynamic environmental changes, but also corrects trajectory deviations caused by modeling errors or external disturbances in real time.

[0061] This section focuses on the collaborative use of MPC and reinforcement learning modules, leveraging a physical constraint closed-loop optimization mechanism to organically combine adaptive global decision-making with ontological safety control. This approach balances global planning capabilities with local physical feasibility, significantly improving the robot's control accuracy and motion safety in dynamic and complex environments. Compared to existing solutions that rely solely on RL or simple MPC, this invention better balances global optimization with real-time response, motion flexibility with safety boundaries, thus enhancing the feasibility and stability of engineering implementation.

[0062] S6: Distribute the final control commands to the four Mecanum wheel drives, adjust the speed and direction of each wheel, and realize the robot's three-degree-of-freedom omnidirectional motion.

[0063] The final control command is decomposed into velocity components along the X-axis, along the Y-axis, and around the vertical axis according to the kinematics of the robot chassis. Based on the installation parameters and rolling direction of the four Mecanum wheels, the inverse kinematics algorithm is used to calculate the target speed and rotation direction of each drive wheel. The target speed and rotation direction of each drive wheel are then sent to the corresponding driver to control the four Mecanum wheels to work together and realize the robot's three-degree-of-freedom omnidirectional motion in the plane.

[0064] The core of step S6 lies in transforming the final control commands obtained from MPC optimization into independent motion parameters for each Mecanum wheel that can actually be driven, achieving omnidirectional motion capability with three degrees of freedom (X-axis translation, Y-axis translation, and rotation around the vertical axis). Specifically, the final control commands typically include the robot's linear velocity in the global coordinate system (…). , ) and angular velocity ( These global commands need to be mapped to the target rotation speed and direction of rotation for each wheel through inverse kinematics of the robot chassis.

[0065] The arrangement of the four Mecanum wheels determines their contribution to the decomposition of different motion components. This is achieved by establishing the inverse kinematic transformation matrix (e.g.:

[0066] ;

[0067] in Here is the kinematic coefficient matrix. For the first The rotational speed of each wheel (∈{1,2,3,4}), which can efficiently map global speed commands to the target speed and direction of rotation of each drive wheel. Subsequently, the system sends the target speed and direction of each wheel to the motor driver through a real-time signal transmission interface, enabling the four Mecanum wheels to move in tandem.

[0068] This step employs a command allocation mechanism based on global inverse kinematics, ensuring that the driving motion of each wheel is strictly consistent with the robot's overall desired motion, significantly improving the accuracy of trajectory tracking and attitude control. Compared to traditional direct drive or simple speed allocation methods, this method fully considers wheel arrangement, chassis parameters, and mechanical effects, effectively suppressing practical problems such as steering deviation, sideslip, or single-wheel failure, thus improving system robustness. This distributed and refined drive command allocation also provides a basic interface for subsequent advanced control such as fault detection, active slippage compensation, and energy consumption optimization, greatly expanding the practical application boundaries and technological depth of omnidirectional mobile robots.

[0069] Example 2 is the second embodiment of the present invention, which differs from the previous embodiment in that:

[0070] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they 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 the part of the current technical solution, can be embodied in the form of a software product. This current 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 described in 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.

[0071] 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.

[0072] 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.

[0073] Example 3, an embodiment of the present invention, provides a mobile robot control system, including a data acquisition module, a data preprocessing and feature fusion module, a reinforcement learning strategy model training module, a reinforcement learning inference and instruction generation module, a predictive control model optimization module, and a chassis drive allocation and execution module.

[0074] Data acquisition module: Collects feature data of the environment in which the positioning object is located and motion state data of the robot body, and obtains environmental feature data and motion state data respectively;

[0075] Data preprocessing and feature fusion module: preprocesses and fuses environmental feature data and motion state data to generate standardized state data;

[0076] Reinforcement learning policy model training module: Based on standardized state data as input, a reinforcement learning training environment is built in an offline simulation environment, and a converged reinforcement learning policy model is obtained through iterative training;

[0077] Reinforcement learning reasoning and instruction generation module: collects and preprocesses the current robot standardized state data in real time, inputs it into the reinforcement learning policy model, and obtains and outputs preliminary motion control instructions and reference trajectories;

[0078] Predictive control model optimization module: Establish a predictive control model MPC based on robot kinematics and dynamic constraints. The predictive control model MPC takes the initial motion control command and reference trajectory as reference inputs, and simultaneously inputs the current standardized state data of the robot to perform short-term motion prediction and optimization, and solves to obtain the final control command that satisfies the constraint conditions.

[0079] Chassis drive distribution and execution module: Distributes the final control commands to the four Mecanum wheel drives, adjusts the speed and direction of each wheel, and realizes the robot's three-degree-of-freedom omnidirectional motion.

[0080] Example 4 is an embodiment of the present invention, which provides a mobile robot control method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0081] A comparative experiment on the path tracking and dynamic obstacle avoidance performance of an omnidirectional mobile robot was conducted in a smart warehouse experimental area. The experimental subject was an autonomous mobile robot equipped with four Mecanum wheels. Its sensor configuration included LiDAR (obstacle distribution), RGBD vision module (ground type and dynamic environment), inertial measurement unit (IMU), and encoder (speed, wheel speed, and position information). The experimental area was pre-set with obstacles and dynamically moving obstacles to simulate complex routes between shelves and temporary obstacle interference.

[0082] The experiment was divided into two groups: the control group used classic PID control with piecewise programming, while the experimental group used the "standardized state data + reinforcement learning strategy + MPC collaborative optimization" control method proposed in this invention. All robots first underwent static calibration and initialization, and then completed multiple rounds of tasks under a unified start and end point and multiple segments and inflection points. In each round, environmental features and the robot's motion state were collected in real time, and standardized inputs were generated through data preprocessing and feature fusion. The experimental group had previously completed offline training of the reinforcement learning strategy model using the simulation environment of this scenario, deployed it on the robot, and performed real-time inference during online runtime, combining it with MPC to form dynamically optimal motion control commands. Finally, key parameters such as robot trajectory deviation, average tracking error, maximum obstacle avoidance reaction time, number of local path adjustments, and energy consumption were fully recorded for each round of experiments.

[0083] Among them, the comparison group-1 aims for a balance between accuracy and speed, with the following specific parameters:

[0084] Proportional / integral / derivative gain: Kp 1.30, Ki 0.05, Kd 0.17;

[0085] Maximum linear velocity: 1.05 m / s; Maximum linear acceleration: 1.4 m / s²;

[0086] Rescheduling cycle: 140 ms;

[0087] Obstacle expansion radius: 0.19 m;

[0088] Pure tracking forward look distance: 0.55 m;

[0089] Wheel speed limit: 230 rpm;

[0090] Compared to the control group -2, which focuses on conservative obstacle avoidance and emphasizes stability (corresponding to slower speed and larger error in Table 1), the specific parameters are as follows:

[0091] Kp 1.10, Ki 0.06, Kd 0.15;

[0092] Maximum linear velocity: 0.92 m / s; Maximum linear acceleration: 1.1 m / s²;

[0093] Replanning cycle: 180 ms (slower, resulting in delayed dynamic response);

[0094] Obstacle expansion radius: 0.23 m (larger, more conservative detour);

[0095] Forward sight distance: 0.45 m;

[0096] Wheel speed limit: 220 rpm;

[0097] Compared to Group-3, which exhibits more aggressive tracking (corresponding to slightly faster speed and slightly lower error in Table 1), the specific parameters are as follows:

[0098] Kp 1.42, Ki 0.04, Kd 0.19;

[0099] Maximum linear velocity: 1.10 m / s; Maximum linear acceleration: 1.6 m / s²;

[0100] Rescheduling cycle: 120 ms;

[0101] Obstacle expansion radius: 0.17 m;

[0102] Forward sight distance: 0.60 m;

[0103] Wheel speed limit: 240 rpm.

[0104] The conservative configuration of Comparison Group-2 (larger expansion radius and slower replanning cycle) resulted in the lowest average speed, increased number of detours, and amplified maximum error; the higher speed and shorter replanning cycle of Comparison Group-3 improved the average error, but the energy consumption was slightly higher; Comparison Group-1 was in between.

[0105] Experimental group-1 represents a trade-off in tracking priority (corresponding to the lower error in Table 1), with the following specific parameters:

[0106] RL policy checkpoint: ckpt-178k (freezes inference, no additional exploration);

[0107] MPC prediction time domain: N = 10 steps;

[0108] Cost weights: Tracking error weight 1.00, control increment weight 0.28, smoothing weight 0.22;

[0109] Physical constraints: Maximum wheel speed 240 rpm, maximum linear acceleration 1.5 m / s²;

[0110] Safety boundary (MPC internal expansion): 0.20 m;

[0111] Experimental group-2 prioritizes energy efficiency and smoothness (corresponding to the lowest energy consumption and highest speed in Table 1, but with a slightly increased error). Specific parameters are as follows:

[0112] RL strategy checkpoint: ckpt-195k (temperature inference T = 0.05, to improve perturbation robustness);

[0113] MPC prediction time domain: N = 12 steps (longer look-ahead, which is beneficial for maintaining speed);

[0114] Cost weights: tracking error weight 0.90, control increment weight 0.35, smoothing weight 0.30;

[0115] Physical constraints: Maximum wheel speed 235 rpm, maximum linear acceleration 1.4 m / s²;

[0116] Safety boundary: 0.18 m;

[0117] Experimental group-3 prioritizes safety boundaries (corresponding to the highest completion rate, slightly higher error, and slightly higher energy consumption in Table 1), with the following specific parameters:

[0118] RL policy checkpoint: ckpt-210k (freeze inference);

[0119] MPC prediction time domain: N = 10 steps;

[0120] Cost weights: Tracking error weight 1.20, control increment weight 0.30, smoothing weight 0.24;

[0121] Physical constraints: Maximum wheel speed 240 rpm, maximum linear acceleration 1.5 m / s²;

[0122] Safety boundary: 0.22 m (larger, more conservative for narrow passages).

[0123] Experimental group-1 prioritizes tracking accuracy in its weighting, resulting in the lowest average error. Experimental group-2 increases the weighting of smoothing and control costs and extends the prediction time domain, which reduces energy consumption while maintaining speed, but slightly increases the average error. Experimental group-3 improves the safety boundary and adopts a more conservative path, achieving the highest completion rate, but with a slight increase in energy consumption and error.

[0124] The experimental data are summarized in Table 1. The test subjects include the control group (traditional method) and the experimental group (the present invention) and their specific performance under the same test task. The units of each parameter are indicated in the table header. The experimental results demonstrate the technical advantages of the robot in high-precision path tracking and high robustness in highly dynamic environments.

[0125] Table 1 Data Reference Table

[0126]

[0127] As shown in Table 1, the mobile robot using the proposed solution (experimental group) significantly outperforms the traditional control group in several key performance indicators, including path tracking accuracy, obstacle avoidance speed, energy consumption, and task completion rate, in intelligent warehousing scenarios. The average path tracking error of the experimental group is only 10.5 mm, a reduction of more than 50% compared to the control group (21.4 mm). The maximum deviation also decreased from an average of 57.8 mm in the control group to 27.2 mm, significantly improving high-precision motion control capabilities. The experimental group's average dynamic obstacle avoidance reaction time is 208 ms, an improvement of more than 45% compared to the control group's 378 ms, indicating that the reinforcement learning + MPC mechanism can adapt to environmental changes more quickly and correct motion strategies in a timely manner. The average number of local path adjustments is reduced by half, indicating that the robot does not need to frequently make large-scale trajectory changes in complex environments, effectively improving motion continuity and safety.

[0128] Furthermore, the experimental group's average energy consumption was 18.6 Wh, approximately 18% lower than the control group's 22.8 Wh, thanks to energy efficiency synergistic control under MPC optimization and improved strategies to reduce ineffective actions. Regarding task completion rate, the experimental group achieved or exceeded 99% almost throughout the entire process, while the control group exhibited significant task failures and deviations. These results fully demonstrate the high robustness and adaptability of the proposed solution in highly dynamic and complex environments.

[0129] This significant improvement in technical performance stems from the decision optimization capabilities of this invention, which integrates perception, feature fusion, deep reinforcement learning, and model predictive control. Multi-source data fusion and standardized processing provide intelligent algorithms with high-dimensional and accurate environmental and ontological perception inputs, effectively reducing the drawbacks of traditional control's excessive reliance on model accuracy and parameters. The introduction of reinforcement learning significantly enhances the adaptability of motion strategies in changing scenarios, while the MPC (Multi-Process Control) stage ensures optimal physical constraints and energy consumption during motion execution. In contrast, the comparison group, relying solely on empirical parameter adjustments, optimization without physical constraints, and weak environmental adaptation, could not maintain a balance between trajectory accuracy and energy consumption in complex environments and was easily affected by dynamic disturbances, impacting task stability. Therefore, the table data directly reflects the technical breakthroughs of this invention in terms of accuracy, response speed, robustness, and energy efficiency, fully demonstrating its creativity and engineering practical value.

[0130] 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 mobile robot control method, characterized in that, include: Collect feature data of the environment in which the positioning object is located and motion state data of the robot body to obtain environmental feature data and motion state data respectively; Environmental feature data and motion state data are preprocessed and feature fused to generate standardized state data; Based on standardized state data as input, a reinforcement learning training environment is constructed in an offline simulation environment, and a converged reinforcement learning policy model is obtained through iterative training. Real-time acquisition and preprocessing of the current robot's standardized state data, inputting it into the reinforcement learning policy model to obtain and output preliminary motion control commands and reference trajectories; A predictive control model (MPC) based on robot kinematics and dynamics constraints is established. The MPC takes the initial motion control command and reference trajectory as reference inputs, and simultaneously inputs the current standardized state data of the robot to perform short-term motion prediction and optimization, and solves to obtain the final control command that satisfies the constraints. The final control commands are distributed to four Mecanum wheel drives to adjust the speed and direction of each wheel, thereby enabling the robot to move in all directions with three degrees of freedom. The collection of characteristic data of the environment in which the positioning object is located and the motion state data of the robot body include collecting information on obstacle distribution, ground type, spatial geometry, ambient light intensity and dynamic changes in the environment to obtain environmental characteristic data; Collect the robot's position, posture, speed, and rotational speed of its four drive wheels to obtain motion state data; The establishment of the predictive control model MPC based on robot kinematics and dynamic constraints includes establishing kinematic models describing the relationship between robot displacement, velocity, acceleration and the rotational speed of each wheel, and dynamic models describing the robot's dynamic behavior, based on the structural parameters, motion mode and force relationship of the mobile robot body. Based on the aforementioned kinematic and dynamic models, the robot's position, attitude, velocity, and acceleration are selected as state variables to construct a predictive control optimization problem. Position constraints, attitude constraints, velocity constraints, acceleration constraints, wheel speed constraints, and driving force constraints are set in the optimization problem.

2. The mobile robot control method as described in claim 1, characterized in that, The preprocessing and feature fusion of environmental feature data and motion state data includes performing noise reduction, filtering, outlier removal, normalization, and time synchronization on the environmental feature data and motion state data respectively, and then splicing and unifying the dimensions of the processed data according to a preset format to fuse them into standardized state data.

3. The mobile robot control method as described in claim 2, characterized in that, The method of obtaining a converged reinforcement learning policy model through iterative training includes setting up a reinforcement learning training system in a simulation environment that includes robot state, action space and reward function, taking standardized state data as input, iteratively training the robot motion control policy through a reinforcement learning algorithm, continuously optimizing the policy parameters until the cumulative reward value reaches the preset convergence condition, and finally obtaining a reinforcement learning policy model that can be used for robot motion control.

4. The mobile robot control method as described in claim 3, characterized in that, The process of obtaining and outputting preliminary motion control commands and reference trajectories includes: periodically collecting current environmental feature data and motion state data during the actual operation of the robot; generating standardized state data for the current period according to a preset preprocessing procedure; inputting the standardized state data for the current period into a reinforcement learning strategy model; and having the reinforcement learning strategy model output the corresponding preliminary motion control commands and reference trajectory information.

5. The mobile robot control method as described in claim 4, characterized in that, The process of obtaining the final control command that satisfies the constraints includes, within each control cycle, taking the preliminary motion control command and reference trajectory as target inputs, and the standardized state data of the current cycle as state inputs, using the predictive control model MPC to predict the motion state and optimize the control sequence for future time periods based on the set physical constraints, using a numerical optimization algorithm to obtain the optimal control command sequence, and selecting the first control command from the optimal control command sequence as the final control command output for this cycle, which is used to drive the robot's actual motion execution.

6. The mobile robot control method as described in claim 5, characterized in that, The process of distributing the final control command to the four Mecanum wheel drivers includes decomposing the final control command into velocity components along the X-axis, along the Y-axis, and around the vertical axis according to the kinematics of the robot chassis. Based on the installation parameters and rolling direction of the four Mecanum wheels, the inverse kinematics algorithm is used to calculate the target speed and rotation direction of each drive wheel, and the target speed and rotation direction of each drive wheel are sent to the corresponding driver to control the four Mecanum wheels to work together to achieve three-degree-of-freedom omnidirectional motion of the robot in the plane.

7. A mobile robot control system for implementing the mobile robot control method as described in any one of claims 1 to 6, characterized in that, include: Data acquisition module: Collects feature data of the environment in which the positioning object is located and motion state data of the robot body, and obtains environmental feature data and motion state data respectively; Data preprocessing and feature fusion module: preprocesses and fuses environmental feature data and motion state data to generate standardized state data; Reinforcement learning policy model training module: Based on standardized state data as input, a reinforcement learning training environment is built in an offline simulation environment, and a converged reinforcement learning policy model is obtained through iterative training; Reinforcement learning reasoning and instruction generation module: collects and preprocesses the current robot standardized state data in real time, inputs it into the reinforcement learning policy model, and obtains and outputs preliminary motion control instructions and reference trajectories; Predictive control model optimization module: Establish a predictive control model MPC based on robot kinematics and dynamic constraints. The predictive control model MPC takes the initial motion control command and reference trajectory as reference inputs, and simultaneously inputs the current standardized state data of the robot to perform short-term motion prediction and optimization, and solves to obtain the final control command that satisfies the constraint conditions. Chassis drive distribution and execution module: Distributes the final control commands to the four Mecanum wheel drives, adjusts the speed and direction of each wheel, and realizes the robot's three-degree-of-freedom omnidirectional motion.

Images