Badminton intelligent robot system and control method thereof

Abstract

The present application belongs to the technical field of mobile robot control and human-computer interaction, and provides a badminton intelligent robot system and a control method thereof to solve the problem that the current badminton training robot cannot meet the diversified, personalized and high dynamic requirements in the national fitness scene. The badminton intelligent robot system comprises a robot body, a multi-source cooperative visual perception unit and a control unit. The multi-source cooperative visual perception unit is used for collecting a badminton global image sequence and a badminton near-field local image sequence, and then detecting a badminton global flight trajectory, solving a robot global pose, and predicting a badminton local flight trajectory and transmitting them to the control unit. The control unit is used for judging the current scene state based on the badminton global flight trajectory and executing a matching action accordingly, thereby constructing a complete closed loop from environment perception, motion decision, accurate execution to intelligent interaction, and improving the response speed, control accuracy and training intelligence of the robot.

Description

Technical Field

[0001] This invention belongs to the field of mobile robot control and human-computer interaction technology, and particularly relates to a badminton intelligent robot system and its control method. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, badminton training robots have shortcomings in terms of serving methods, environmental perception, operational decision-making, and execution, making it difficult to meet the diverse, personalized, and highly dynamic needs of the national fitness campaign.

[0004] In the serving phase, current badminton training robots are mainly fixed serving machines or simple wheeled serving devices, resulting in a lack of mobility and an inability to simulate the changes in the trajectory and randomness of the ball's landing point in real rallies; or poor maneuverability and slow response, making it difficult to achieve full-court coverage and continuous multi-shot interaction. In the environmental perception phase, the visual perception module and motion execution module of existing badminton robots are designed independently. The ball position information acquired by the perception end is difficult to transmit to the motion control end in real time and accurately, resulting in delayed robot decision-making, inaccurate timing of hitting the ball, and an inability to form a rapid closed loop from perception to execution, thus affecting the smoothness and stability of continuous multi-shot rallies. In the operation decision and execution phase, existing badminton robots mostly use traditional methods based on inverse kinematics plus PID (proportional-integral-derivative) controllers, or trajectory prediction methods based on aerodynamic models. It is difficult to achieve millisecond-level coordination between the mobile platform and the robotic arm in high-speed ball scenarios, and the prediction accuracy is limited; moreover, the training process is a one-way output, which cannot dynamically adjust the difficulty according to the user's level (from beginner to intermediate), and it is difficult to provide real-time data feedback and action guidance. Summary of the Invention

[0005] To address the technical problems mentioned above, this invention provides a badminton intelligent robot system and its control method. The system constructs a complete closed loop from environmental perception, motion decision-making, precise execution to intelligent interaction, controlling the robot to achieve rapid decision-making and precise completion of various hitting actions, thereby improving the robot's response speed, control accuracy, and training intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a badminton intelligent robot system.

[0007] A badminton intelligent robot system includes: a robot body, a multi-source collaborative visual perception unit, and a control unit; the robot body includes a mobile platform and a single-arm multi-axis racket robotic arm and a serving mechanism mounted thereon, wherein a racket is disposed at the end of the single-arm multi-axis racket robotic arm; the serving mechanism is connected to the control unit. The multi-source collaborative visual perception unit is used to: acquire global image sequences and near-field local image sequences of badminton shuttlecocks, and then detect the global flight trajectory of the badminton shuttlecocks, calculate the global pose of the robot, and predict the local flight trajectory of the badminton shuttlecocks and transmit them to the control unit. The control unit is used to: determine the current scene state based on the global flight trajectory of the badminton shuttlecock and execute matching actions accordingly; the scene state includes continuous rally state, waiting for the shuttlecock to arrive state, and needing to serve state; When the scene is in the state of needing to serve, control the serving mechanism to perform the serving action. After the serve is completed, automatically switch to the waiting state to prepare for the next round of continuous fighting. When the scenario is in a continuous rally state, the type of incoming shuttlecock is determined by combining the global and local flight trajectories of the shuttlecock with the flight parameters, and the hitting strategy is adaptively matched. Based on the calculated global pose of the robot, a two-stage collaborative control of the mobile platform and the single-arm multi-axis racket robot is adopted to execute the matched hitting strategy and obtain the corresponding execution results, thereby dynamically adjusting the hitting strategy for the next cycle.

[0008] In one implementation, the control unit comprises two stages: a pre-positioning stage and a precise control stage. The first stage is the pre-positioning stage. When the predicted local flight trajectory of the badminton shuttlecock has not yet converged, the control platform moves to the approximate area of ​​the return point with adaptive proportional control, and the control of each joint of the single-arm multi-axis racket robotic arm moves to the ready position. The second stage is the precision control stage. After the predicted local flight trajectory of the badminton shuttlecock converges, based on the robot's global pose, the hitting point, and the hitting time, a nonlinear model predictive control algorithm is used to calculate the target acceleration of each joint of the single-arm multi-axis racket robot arm and the target acceleration of the mobile platform, so as to achieve the timing coordination between the omnidirectional movement of the mobile platform and the precise hitting of the single-arm multi-axis racket robot arm.

[0009] In one implementation, the cost function corresponding to the nonlinear model predictive control algorithm in the control unit is a weighted sum of the weighted compensation terms for the ball position error, ball speed error, and ball angle error in the prediction time domain and the smoothing term of the control quantity in the control time domain.

[0010] In one implementation, the control unit has a training mode database that stores multiple training modes. The control unit is also used to acquire user hitting data and evaluate the user's skill level based on a multi-dimensional weighted scoring model. Then, it adaptively matches training modes from the training mode database, generates corresponding serving parameter adjustment instructions and mobile platform target pose instructions, and sends them to the serving mechanism and the mobile platform.

[0011] In one implementation, the control unit is also used to: acquire data of multiple consecutive shots by the user, calculate the comprehensive score of each shot based on a multi-dimensional weighted scoring model, and then compare it with the difficulty adjustment threshold to dynamically adjust the training difficulty; The difficulty adjustment threshold is dynamically determined based on the statistical characteristics of the sliding window of the user's historical shot data: a sliding window of preset length is maintained to record the comprehensive score of the user's most recent shots, and the mean and standard deviation of the comprehensive score within the sliding window are calculated; the difficulty adjustment threshold includes a difficulty increase threshold and a difficulty decrease threshold. The difficulty increase threshold is the mean plus the product of the sensitivity coefficient and the standard deviation; the difficulty decrease threshold is obtained by subtracting the product of the sensitivity coefficient and the standard deviation from the mean, and then limiting it by a preset minimum effective threshold.

[0012] As one implementation method, the control unit employs a continuous counter mechanism to dynamically adjust the training difficulty, the process of which is as follows: When the combined score of the most recent preset number of shots is higher than the difficulty increase threshold, the cumulative upgrade counter is incremented. If the score is lower than the difficulty increase threshold, it is reset to zero. When the preset number of triggers is reached, the difficulty is automatically increased or the proportion of high-difficulty shots is increased. When the combined score of the most recent consecutive preset number of shots is lower than the difficulty reduction threshold, the downgrade counter is incremented. If the score is higher than the difficulty reduction threshold, it is reset to zero. When the preset number of triggers is reached, the difficulty is automatically reduced or the proportion of high-difficulty shots is reduced.

[0013] In one implementation, the mobile platform is a four-wheeled X-shaped omnidirectional mobile platform, which adopts an X-shaped symmetrical frame structure. It consists of four main longitudinal beams and a central load-bearing platform forming an overall load-bearing skeleton. The four main longitudinal beams extend outward in an X-shape with the geometric center of the robot body as the intersection point. The included angle between adjacent main longitudinal beams is 90°. The ends of each main longitudinal beam are rigidly connected to the wheel hub module mounting base by bolts.

[0014] In one embodiment, the single-arm multi-axis racket robotic arm is mounted on a robotic arm base, the serving mechanism is located inside the robotic arm base, and the robotic arm base is mounted on the center of the mobile platform.

[0015] In one embodiment, the serving mechanism includes a dual-wheel friction drive assembly, a ball supply trajectory, a ball storage tube, and a ball distribution wheel; the left and right wheels of the dual-wheel friction drive assembly are driven independently, imparting different spin modes to the shuttlecocks through differential rotation; the ball storage tube is connected to the ball supply trajectory at the end via a flexible ball delivery tube; the ball distribution wheel is located at the outlet of the ball storage tube, enabling the orderly feeding of shuttlecocks one by one from within the ball storage tube.

[0016] A second aspect of the present invention provides a control method for a badminton intelligent robot system.

[0017] A control method for a badminton intelligent robot system includes the following steps performed in the control unit: The system receives the global flight trajectory of the badminton shuttlecock detected by multi-source collaborative visual perception, calculates the global pose of the robot, and predicts the local flight trajectory of the badminton shuttlecock. Based on the global flight trajectory of the badminton shuttlecock, the current scene state is determined and the matching action is executed accordingly; the scene state includes the continuous rally state, the waiting for the shuttlecock, and the need to serve the shuttlecock. When the scene is in the state of needing to serve, control the serving mechanism to perform the serving action. After the serve is completed, automatically switch to the waiting state to prepare for the next round of continuous fighting. When the scenario is in a continuous rally state, the type of incoming shuttlecock is determined by combining the global and local flight trajectories of the shuttlecock with the flight parameters, and the hitting strategy is adaptively matched. Based on the calculated global pose of the robot, a two-stage collaborative control of the mobile platform and the single-arm multi-axis racket robot is adopted to execute the matched hitting strategy and obtain the corresponding execution results, thereby dynamically adjusting the hitting strategy for the next cycle.

[0018] Compared with the prior art, the beneficial effects of the present invention are: The badminton intelligent robot system of the present invention utilizes a multi-source collaborative visual perception unit to collect global image sequences and near-field local image sequences of the badminton shuttlecock, thereby detecting the global flight trajectory of the shuttlecock, calculating the global pose of the robot, and predicting the local flight trajectory of the shuttlecock. Based on the global flight trajectory of the shuttlecock, the control unit determines the current scene state and executes matching actions accordingly. Furthermore, in the continuous rally state, the system dynamically adjusts the hitting strategy for the next cycle based on the execution results, thus realizing a complete closed-loop control from environmental perception, motion decision-making, precise execution to intelligent interaction.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a schematic diagram of the structure of a badminton intelligent robot system without a multimodal human-computer interaction module according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a badminton intelligent robot system including a multimodal human-computer interaction module according to an embodiment of the present invention; Figure 3 This is a three-dimensional schematic diagram of the mobile platform according to an embodiment of the present invention; Figure 4 This is a bottom view of the mobile platform according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the serving mechanism according to an embodiment of the present invention; Figure 6 This is a control flowchart of the badminton intelligent robot system according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the multi-dimensional weighted scoring model according to an embodiment of the present invention; Figure 8 This is a schematic diagram of user skill level, matching training mode, and ball path type according to an embodiment of the present invention; Figure 9 This is a schematic diagram of multimodal human-computer interaction according to an embodiment of the present invention.

[0022] The components include: 1. Robot body; 2. Mobile platform; 3. Single-arm multi-axis racket robotic arm; 4. Serving mechanism; 5. Racket; 6. Multimodal human-machine interaction module; 21. Main longitudinal beam; 22. Central bearing platform; 23. Steering module; 24. Wheel hub module; 41. Ball storage cylinder; 42. Ball distribution wheel; 43. Ball supply trajectory; and 44. Dual-wheel friction drive assembly. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] The example environment of this invention includes a user, a robot body, and a control unit. The user is a person participating in badminton training who interacts with the robot body through voice commands or touch operations, and can view training data and evaluation reports in real time. The robot body can also be extended to other types of ball sports training robots. The control unit is built into the robot body and undertakes the end-to-end hard real-time control tasks. It can also be extended to an external edge computing node or cloud server to execute the control method of this invention and feed back the execution results to the multimodal human-machine interaction module. The robot body and the control unit communicate at high speed in real time through a CAN-FD (CAN with Flexible Data-Rate) bus or an EtherCAT (Ethernet Control Automation Technology) bus. In the embodiments of this invention, data acquisition and processing strictly comply with relevant laws and regulations. The acquisition of user hitting data requires authorization and is only used for training evaluation and difficulty adjustment, and not for other unauthorized purposes. It should be understood that the above example environment is described for illustrative purposes only and does not imply any limitation on the scope of this disclosure. This invention can also be applied to other ball sports training, physical education, and other scenarios, and the environment may also include components such as sensors and communication modules not shown.

[0027] according to Figure 1 An embodiment of the present invention provides a badminton intelligent robot system, comprising a robot body 1, a multi-source collaborative visual perception unit, and a control unit. The robot body 1 includes a mobile platform 2, a single-arm multi-axis racket robotic arm 3 mounted on the mobile platform 2, and a serving mechanism 4, with a racket 5 disposed at the end of the single-arm multi-axis racket robotic arm 3.

[0028] In one implementation, such as Figure 3 and Figure 4As shown, the mobile platform 2 is a four-wheeled X-shaped omnidirectional mobile platform. The four-wheeled X-shaped omnidirectional mobile platform adopts an X-shaped symmetrical frame structure, consisting of four main longitudinal beams 21 and a central support platform 22 forming the overall load-bearing skeleton. The four main longitudinal beams 21 extend outwards in an X-shape with the geometric center of the robot body 1 as the intersection point. The included angle between adjacent main longitudinal beams is 90°, and the ends of each main longitudinal beam are rigidly connected to the wheel hub module mounting base via bolts. The central support platform 22 is located at the geometric center of the X-shaped symmetrical frame structure and is used to support the battery module, control unit, and robotic arm base. The bottom of the mobile platform 2 is equipped with reinforcing ribs to improve overall torsional stiffness. The X-shaped symmetrical frame structure is constructed from hollow aluminum alloy profiles, reducing the self-weight of the mobile platform while ensuring structural strength.

[0029] The mobile platform 2 also includes a wheel hub module 24 and a steering module 23. Both the wheel hub module 24 and the steering module 23 adopt an integrated quasi-direct drive module that combines a motor-reducer-driver. This module integrates a brushless DC motor, a low-reduction-ratio planetary reducer, and a field-oriented control driver into the same mechanical housing, forming a three-in-one structure. Each module supports real-time parameter reconfiguration via a CAN bus. The housing of the wheel hub module 24 is mounted below the rotating support plate via a connecting bracket. Its motor rotor output is rigidly connected to a custom-coated wheel hub, and the tire covers the outer edge of the wheel hub, forming a short transmission chain wheel-side drive unit without couplings or intermediate transmission components. The motor stator housing is fixed, and the rotor rotation directly drives the wheel hub rotation, thereby reducing backlash and mechanical losses, and improving transmission efficiency and dynamic response. The steering module 23 is mounted above the rotating support plate, and its housing is fixedly connected to the suspension of the mobile platform. Its output end drives the rotating support plate and the entire wheel assembly to rotate around the vertical steering axis via a slewing bearing. The steering module has a built-in dual magnetic encoder; one channel is used for motor field-oriented vector control, and the other channel directly outputs the absolute steering angle, thus eliminating the need for a separate steering angle encoder, simplifying wiring, and improving system reliability. All four wheels are independently driven, with the wheel axles forming a 45° angle with the robot body's longitudinal or lateral direction. Decoupled control of longitudinal, lateral, and rotational motion is achieved through vector synthesis of the four wheel speeds. Depending on the specific requirements of the field, the maximum linear velocity can be set to no less than 6 m / s, and the maximum acceleration to no less than 10 m / s².

[0030] This invention provides an integrated high-performance quasi-direct drive module that combines a motor, reducer, and driver into one unit. The module integrates a high-density, high-torque motor, a planetary reducer, and an FOC driver into the same housing. Compared with traditional external reducer servo systems, it reduces size, weight, and power density, and can achieve a high-density arrangement of four-wheel independent drive and steering within a limited wheel well space, providing underlying drive force guarantee for high-speed omnidirectional mobile platforms.

[0031] The embodiments of the present invention can achieve high-bandwidth indirect force control without the need for additional joint torque sensors, which simplifies the structure and reduces costs. At the same time, in high-speed impact scenarios such as slamming, the high transparency of the low reduction ratio transmission chain allows the motor end to directly sense external impacts, achieving smooth buffering and collision protection at the current loop level, which significantly improves the safety of human-machine interaction.

[0032] like Figure 2 As shown, the mobile platform 2 is also equipped with a multimodal human-computer interaction module 6, such as Figure 9 As shown, the multimodal human-computer interaction module 6 includes a voice interaction unit, a touch display unit, and a light indicator unit, which are connected to the control unit to form a complete interactive closed loop of command input and status feedback.

[0033] The voice interaction unit employs an offline voice recognition module with a response latency of less than 200ms. Users can control the robot to switch training modes, adjust serving parameters, or check training status via voice commands (such as "start training," "increase ball speed," "serve to the left half of the court"), achieving hands-free operation and natural interaction during training. The touch display unit displays real-time training data such as the current number of rallies, average ball speed, heatmap of ball landing points, user movement distance, and calorie consumption. It also supports training mode selection, parameter settings, score display, robotic arm joint status display, current ball trajectory type display, and training report presentation. The training report includes user skill level trends, return success rates for each ball trajectory type, movement accuracy statistics, and targeted improvement suggestions, enabling users to quantify and track training effectiveness. The light indicator unit uses different colors or flashing frequencies to indicate the current training mode, shot quality level, shot type, or safety status. For example, green / yellow / red indicate excellent / average / shot quality that needs improvement, respectively, and blue / white / orange indicate basic shot / normal shot / high difficulty shot, respectively. This allows users to intuitively obtain feedback information without having to look at the screen while moving quickly.

[0034] The system receives user voice or touch commands through a multimodal human-computer interaction module and outputs multimodal feedback, including shot quality evaluation, training reports, and light indicators. It evaluates the user's skill level based on a multidimensional weighted scoring model, adaptively matches training modes, and dynamically adjusts training difficulty, enabling local storage and cloud synchronization of training data.

[0035] In one embodiment, the serving mechanism 4 is built into the robotic arm base, sharing the same central support platform mounting reference with the single-arm multi-axis racket robotic arm 3, forming a compact integrated layout for serving and hitting. Figure 5As shown, the serving mechanism 4 includes a dual-wheel friction drive assembly 44, a shuttlecock feed trajectory 43, a shuttlecock storage cylinder 41, and a shuttlecock distribution wheel 42. The left and right wheels of the dual-wheel friction drive assembly 44 are driven by independent servo motors, imparting different spin modes to the shuttlecocks through differential rotation, thereby achieving flexible control over the spin characteristics of the served shuttlecock. The shuttlecock storage cylinder 41 is connected to the shuttlecock feed trajectory 43 at the end via a flexible feed tube, which reduces the probability of shuttlecock jamming during the feeding process. The shuttlecock distribution wheel 42 is located at the outlet 41 of the shuttlecock storage cylinder and is driven by a servo motor, ensuring that the shuttlecocks in the storage cylinder are fed one by one in an orderly manner, guaranteeing the controllability of the serving rhythm. A photoelectric sensor is installed in the shuttlecock feed trajectory 43 for real-time monitoring of the feeding status and detection of shuttlecock jamming; when the photoelectric sensor detects a shuttlecock jamming, the control unit controls the shuttlecock distribution wheel to perform a reverse jittering self-recovery action; if it still cannot recover, a shuttlecock jamming prompt is broadcast and fault information is displayed through the multi-modal human-machine interaction module. Because the ball-serving mechanism is built into the area between the bottom of the robotic arm and the central support platform, the overall center of gravity height is lowered compared to an external solution, and the moment of inertia around the vertical axis is reduced, which is beneficial to the high dynamic steering response of the mobile platform.

[0036] The badminton intelligent robot system of this invention integrates an integrated high-performance collimated direct drive module into the four-wheel X-shaped omnidirectional mobile platform and the joints of the robotic arm. While maintaining high torque density, it achieves near-direct drive dynamic transparency, enabling both the mobile platform and the robotic arm to have millisecond-level dynamic response capabilities. The maximum acceleration is not less than 10m / s² and the maximum speed is not less than 6m / s, which can cover high-speed incoming balls across the entire court and complete accurate returns.

[0037] In one embodiment, the robotic arm base of the single-arm multi-axis racket robotic arm is mounted above the geometric center of the central support platform of the mobile platform. The robotic arm extends upward from the base and then unfolds forward to form a cantilever structure, with a racket at the end for hitting the ball. All five joints of the single-arm multi-axis racket robotic arm are driven by joint modules, functionally divided into power joints (P-joints) and adjustment joints (A-joints). The P-joints include joint J1, which provides the racket swing speed; the A-joints include joints J2 to J5, which adjust the various degrees of freedom of the racket posture. J1, as a P-joint, is driven by a module equipped with planetary gears to provide swing speed without requiring extremely high positional accuracy, similar to the human waist; J2 and J3, as A-joints, are driven by modules equipped with harmonic reducers to control the swing height, similar to the human shoulder and elbow; J4 and J5, as A-joint modules, use direct-drive planetary-cycloidal hybrid reducers to control the racket posture within a limited space, similar to the human wrist. Each joint of the robotic arm is driven by an integrated high-performance collimation drive module. The module housing is rigidly connected to the arm structure, forming a transmission chain with high rigidity and low inertia.

[0038] In one embodiment, the multi-source collaborative visual perception unit is used to: acquire a global image sequence of a badminton shuttlecock and a near-field local image sequence of a badminton shuttlecock, and then detect the global flight trajectory of the badminton shuttlecock, calculate the global pose of the robot, and predict the local flight trajectory of the badminton shuttlecock and transmit it to the control unit.

[0039] The multi-source collaborative visual perception unit includes a baseline binocular visual perception unit deployed outside the site and a three-dimensional binocular visual perception unit deployed on the robot body.

[0040] The baseline binocular vision perception unit is deployed on the outer side of the baseline of a standard badminton court. The world coordinate system of the court is constructed with the center of the outer side of the baseline as the origin. It is used to realize the global detection of the badminton flight trajectory and the independent observation of the global pose of the robot body throughout the entire court.

[0041] The baseline binocular vision perception unit utilizes baseline disparity to achieve long-distance, high-precision detection of the three-dimensional position of the badminton shuttlecock across the entire court. In binocular stereo vision, the baseline refers to the physical distance between the optical centers of the two cameras, and the disparity refers to the difference in pixel coordinates between the projection points of the same spatial point on the left and right images.

[0042] The baseline binocular vision perception unit first uses a checkerboard calibration board to simultaneously calibrate the binocular cameras, obtaining the intrinsic parameter matrices, distortion coefficients, and relative pose transformation matrix of the right camera relative to the left camera; the process is as follows: Obtain the left camera intrinsic parameter matrix in one go Right camera intrinsic parameter matrix Left camera distortion coefficient Right camera distortion coefficient The relative pose transformation matrix of the right camera relative to the left camera Among them, the intrinsic parameter matrix The focal length and principal point parameters of a camera are represented in a unified form as follows: (1); In the formula, The camera in the image Focal length of direction, Here are the origin coordinates of the virtual reference camera after stereo calibration, denoted as [reference point] for the left camera. The corresponding right camera is denoted as The virtual reference camera is based on the left camera. Original distortion coefficients. Describing radial and tangential distortion, used to represent actual imaging points Corrected to the ideal imaging point T is the vector transpose. Actual imaging point. With ideal imaging point The relationship is: (2); In the formula, , Radial distortion coefficient; Tangential distortion coefficients. Relative pose transformation matrix. By rotation matrix With translation vector Composition, in the form of: (3); In the formula, satisfy and This describes the pose relationship between the right camera coordinate system and the left camera coordinate system. It is the identity matrix; Representing the rotation matrix Determinant; translation vector This represents the three-dimensional coordinates of the optical center of the right camera in the coordinate system of the left camera. The baseline length between the optical centers of the two cameras. It is determined by the Euclidean magnitude of the translation vector, that is: (4); Based on the calibration results, the stereo correction rotation matrix and projection matrix are calculated to ensure that the epipolar lines of the left and right images are strictly horizontally aligned after correction, and a stereo correction mapping is generated for real-time distortion correction and epipolar line correction of each subsequent acquired image frame.

[0043] Specifically, based on the calibration results , , , and Calculate the stereo correction rotation matrix With projection matrix This ensures that the epipolar lines of the corrected left and right images are strictly horizontally aligned, meaning that corresponding points of the same spatial point in the two corrected images satisfy... Projection matrix Define the unified intrinsic parameters of the stereo-corrected virtual reference camera, including the equivalent focal length. coordinates of the principal point Its form is: (5); In the formula, The equivalent focal length of the virtual reference camera after stereo calibration is given, and the left and right virtual cameras share the same internal parameters after calibration.

[0044] The binocular camera synchronously acquires badminton shuttlecock image sequences at a frame rate of no less than 160fps via hardware triggering, ensuring that the synchronization error of the timestamps of the left and right frames meets the real-time detection requirements in high-speed shuttlecock scenarios.

[0045] After performing real-time distortion and stereo correction on the synchronously acquired left and right images, the position of the badminton shuttlecock pixel is detected in the left image. And along the horizontal polar line, a one-dimensional search is performed in the corresponding row of the right figure to match badminton features and determine... Meanwhile, a parabolic interpolation subpixel-level matching algorithm is used to improve parallax accuracy.

[0046] The system calculates the depth coordinates of the shuttlecock in the left camera coordinate system by substituting parallax into the depth formula. Then, it reconstructs the 3D coordinates of the shuttlecock in the left camera coordinate system using triangulation. This is then transformed to the field world coordinate system via a homogeneous transformation using the field joint calibration extrinsic parameter matrix. Simultaneously, the baseline binocular vision perception unit identifies the QR code marker affixed to the robot body and calculates the robot's global pose in the field world coordinate system using the PnP (Perspective-n-Point) pose calculation algorithm, providing centimeter-level global positioning for mobile platform navigation and positioning control. The system maintains a preset-length observation window, recording the 3D position sequence of the shuttlecock in the most recent frames. The least squares method is used to perform parabolic fitting on the data within the window to estimate the shuttlecock's flight trajectory parameters and predict the center point of the return area and the estimated arrival time.

[0047] Specifically, after stereo correction, the row coordinates of corresponding points in the left and right images are consistent, that is... Horizontal parallax Defined as: (6); Substituting the horizontal parallax into the depth formula, we can calculate the depth coordinates of the badminton shuttlecock in the left camera coordinate system. (Distance along the camera's optical axis): (7); Further, the three-dimensional coordinates of the badminton shuttlecock in the left camera coordinate system are reconstructed using triangulation: (8); In the formula, Here are the three-dimensional coordinates of the badminton shuttlecock in the left camera coordinate system; the left camera coordinate system has its origin at the optical center of the left camera. The axis points inward along the optical axis (at a downward angle to the horizontal plane). The axis runs horizontally to the right along the image. The axis is perpendicular to the image and points downwards. The 3D points in the left camera coordinate system are then subjected to the site-joint calibration extrinsic parameter matrix. Perform a homogeneous transformation to convert to the site's world coordinate system: (9); In the formula, The badminton shuttlecock is in three-dimensional coordinates within the court's world coordinate system; the origin of the court's world coordinate system is the center outside the baseline. , The axis points longitudinally towards the net along the court. The axis runs laterally along the site. The axis is perpendicular to the horizontal ground and points upwards, that is... This represents the height of the badminton shuttlecock relative to the horizontal ground. (External parameter matrix) By rotation matrix With translation vector Composition, encoding the attitude (including tilt angle) and position (including elevation) of the left camera relative to the site's world coordinate system, namely: (10); After joint calibration of the site, the extrinsic parameter matrix has established a rigid mapping between the left camera coordinate system and the site world coordinate system. Therefore, the reconstruction result directly outputs the three-dimensional position in the site world coordinate system.

[0048] Simultaneously, the QR code marker affixed to the robot body is identified, and the global pose of the robot in the world coordinate system of the site is calculated using the PnP pose calculation algorithm: (11); In the formula, This represents the robot's global pose in the world coordinate system of the site. The known 3D coordinates of the QR code corner points in the robot's body coordinate system. These are the pixel coordinates of the QR code corner points detected in the image. and These are the camera intrinsic parameter matrix and distortion coefficients, respectively; This is the algorithm function for PnP pose calculation.

[0049] Maintain an observation window of preset length N frames and record the three-dimensional position sequence of the sphere in the most recent few frames. The least squares method is used to fit a parabolic curve to the data within the window, estimating the badminton shuttlecock's flight trajectory parameters and predicting the center point of the return area and the estimated arrival time. The depth measurement accuracy brought about by the large baseline layout is quantified and evaluated by the error propagation formula: (12); In the formula, This represents the depth measurement error along the optical axis in the left camera coordinate system. This represents sub-pixel level parallax detection error. At focal length... When fixed, increase the baseline. This can increase the error amplification factor. Linear reduction significantly improves the depth reconstruction accuracy of distant targets; After the transformation, this accuracy advantage is simultaneously mapped to the three-dimensional positioning accuracy in the world coordinate system of the site.

[0050] A 3D binocular vision perception unit is deployed on the robot body. With a short baseline and wide field of view, it initiates fine-grained perception after the badminton shuttlecock enters the robot's near-field region, supplementing the blind spots of the baseline binocular vision perception unit when there is near-field occlusion or insufficient resolution, thus achieving high-precision local tracking of the shuttlecock's trajectory. After the 3D binocular vision perception unit and the robotic arm complete hand-eye calibration, it can directly output the 3D coordinates of the shuttlecock in the robotic arm's operating space, providing high-precision near-field data for robotic arm trajectory planning and precise control of the hitting point during the precision control phase. The robotic arm pose and the robot's global pose are used in real-time for hand-eye calibration to calculate the shuttlecock's 3D position in the camera coordinate system and uniformly transform it to the court world coordinate system, ensuring that the visual coordinate system and the court coordinate system maintain consistent mapping during dynamic swings and full-court movement. The control unit communicates with the multi-source collaborative vision perception unit, receiving the shuttlecock's global flight trajectory, the robot's global pose, and the shuttlecock's local flight trajectory. Based on this data, it determines the current scene state and executes matching actions accordingly.

[0051] The multi-source collaborative vision perception unit achieves the synergy of global and local perception through a data fusion strategy: the baseline binocular vision perception unit provides the initial detection of the badminton trajectory across the entire court and the robot's global pose, while the three-dimensional binocular vision perception unit provides high-resolution supplementary detection in the near-field region. After the two are uniformly transformed to the same world coordinate system, the perception-decision task adaptively switches the main detection source or performs weighted fusion based on the relative distance between the badminton and the robot, ensuring continuous and seamless visual coverage from full-court trajectory prediction to accurate near-field shot.

[0052] In the specific implementation process, combined with Figure 6 The control unit is used to: determine the current scene state based on the global flight trajectory of the badminton shuttlecock and execute matching actions accordingly; the scene state includes the continuous rally state, the waiting for the shuttlecock to arrive state, and the need to serve state.

[0053] When the badminton shuttlecock is determined to be in a valid rally, the control platform and robotic arm work together to execute a precise return shot. After the return shot is completed, the system enters a waiting state for the incoming shuttlecock. The multi-source collaborative visual perception module enters a high frame rate detection mode to monitor whether the user's return shot enters the valid court area. When a lost shot event is detected based on visual perception and court boundary determination (including situations where the rally is interrupted due to the shuttlecock going out of bounds, the user not touching the shuttlecock, or the shuttlecock landing), or when the system is in an initial no-spin state and a rally needs to be started, the serve state is triggered. The serve trigger conditions include at least one of the following: the initial opening serve and a re-serve after a lost shot.

[0054] When the scene is in the state of needing to serve, control the serving mechanism to perform the serving action. After the serve is completed, automatically switch to the waiting state to prepare for the next round of continuous fighting.

[0055] When the scenario is in a continuous rally state, the type of incoming shuttlecock is determined by combining the global and local flight trajectories of the shuttlecock with the flight parameters, and the hitting strategy is adaptively matched. Based on the calculated global pose of the robot, a two-stage collaborative control of the mobile platform and the single-arm multi-axis racket robot is adopted to execute the matched hitting strategy and obtain the corresponding execution results, thereby dynamically adjusting the hitting strategy for the next cycle.

[0056] The shuttlecock flight parameters here include flight time, estimated initial velocity within the observation window, and flight elevation angle. The incoming shuttlecock type is automatically classified based on a combined threshold of these three parameters: flight time, estimated initial velocity within the observation window, and flight elevation angle. The combined threshold for these three parameters is preset or calibrated online based on at least one of the following: shuttlecock model, court environment, and historical hitting data.

[0057] A high clear corresponds to a long flight time, low initial velocity, and high elevation angle; a flat clear corresponds to a medium flight time, medium initial velocity, and medium elevation angle; and a smash corresponds to a short flight time, high initial velocity, and low elevation angle. The initial velocity mentioned above specifically refers to the initial velocity estimated by least-squares fitting based on the position data of the most recent frames after the shuttlecock enters the binocular vision observation window, not the velocity of the shuttlecock at the moment of impact. When the discrimination parameters are in the ambiguous zone of the category boundary, the default response is based on a higher dynamic threat level to ensure interception safety. The control mode adaptively switches according to the real-time discrimination results: a high clear triggers an energy-saving positioning mode, a flat clear triggers a standard response mode, and a smash triggers a high-dynamic emergency mode, achieving intelligent matching between the incoming ball type and the robot's response strategy.

[0058] The spherical flight dynamics model is based on rigid body aerodynamics, and parameters are calibrated by combining the force characteristics of the badminton shuttlecock in actual flight with kinematic observation data. The badminton shuttlecock is mainly affected by gravity and air resistance during flight, and its dynamic equations are as follows: (13); In the formula, For the quality of badminton shuttlecocks, Let g be the acceleration vector of the badminton shuttlecock, g=(0,0, 9.81) m / s² is the gravitational acceleration vector. This refers to air resistance. With badminton speed In the opposite direction, its magnitude is proportional to the square of the velocity: (14); In the formula, air density; The drag coefficient of the badminton shuttlecock (calibrated through wind tunnel experiments or trajectory fitting); The cross-sectional area of ​​the badminton shuttlecock facing the wind; badminton shuttlecock flight speed; Let be the unit vector in the direction of velocity. This simplified model, while maintaining prediction accuracy, avoids the difficulties of real-time estimation of rotational parameters and the risk of numerical singularities.

[0059] When the discrimination parameter is in the fuzzy area of ​​the category boundary, the default response is based on the higher dynamic threat level; the hitting strategy is adaptively selected according to the discrimination result, including landing point calculation, hitting point planning and return strategy generation; if in a continuous rally state, the return strategy branch is entered; if it is determined that the current scenario requires a serve, the serve strategy branch is entered; at the same time, the execution result is received and the discrimination sensitivity and strategy parameters of the next cycle are dynamically adjusted.

[0060] By determining whether the ball has crossed the court boundary, touched the ground, or whether the user has successfully touched the ball, a round interruption signal is generated. The evaluation results are distributed through a dual-path feedback mechanism: the first path is a visual feedback closed loop, which directly feeds back the visual perception data (including the ball loss detection results) to the multi-source collaborative visual perception unit for continuous detection in the next round; the second path is a strategy correction feedback, which feeds back the ball quality evaluation and deviation analysis results to the adaptive matching ball strategy step for correcting the incoming ball discrimination threshold and ball strategy parameters.

[0061] Specifically, the two stages are: the pre-positioning stage and the precise control stage. The first stage is the pre-positioning stage. When the predicted local flight trajectory of the badminton shuttlecock has not yet converged, the control platform moves to the approximate area of ​​the return point with adaptive proportional control, and the control of each joint of the single-arm multi-axis racket robotic arm moves to the ready position. The second stage is the precision control stage. After the predicted local flight trajectory of the badminton shuttlecock converges, based on the robot's global pose, the point of impact, and the time of impact, a nonlinear model predictive control algorithm is used to calculate the target acceleration of each joint of the single-arm multi-axis racket robot and the target acceleration of the moving platform. This achieves the time-series coordination between the omnidirectional movement of the moving platform and the precise hitting of the single-arm multi-axis racket robot. Specifically, the power joints maintain a fixed rotational speed to provide the swing speed, while the adjustment joints adjust the racket posture in real time.

[0062] The specific implementation process of the pre-positioning stage is as follows: Based on a multi-source collaborative visual perception unit, badminton shuttlecock image sequences are acquired at a high frame rate. The 3D position of the shuttlecock in the camera coordinate system is calculated through stereo matching and then transformed to the field world coordinate system through hand-eye calibration. An observation window of a preset length is maintained, recording the shuttlecock position sequence of the most recent few frames. The least squares method is used to perform parabolic fitting on the data within the window to obtain the initial parameter estimate of the badminton shuttlecock's flight trajectory, and then predict the center point of the return area and the expected arrival time.

[0063] The rebound zone is defined as a circular area centered on the predicted landing point, with a radius determined by the trajectory prediction uncertainty and the braking distance at the current speed of the moving platform. The prediction uncertainty is obtained by statistical analysis of the trajectory fitting residuals, and the braking distance is proportional to the square of the current speed of the moving platform and inversely proportional to the maximum acceleration.

[0064] In the pre-positioning phase of the mobile platform, adaptive proportional control with speed saturation is employed. The speed command is the position error mapped by an adaptive gain matrix and then limited by a vector saturation function to ensure that the synthesized speed does not exceed the maximum linear speed of the mobile platform. The adaptive gain increases with the distance between the mobile platform and the target point to accelerate the approach and decreases with the distance to avoid overshoot. Meanwhile, motion command smoothing is achieved through a dual mechanism of dead-zone filtering and rate limiting.

[0065] The dead-zone filtering rule is: if the change in the velocity command vector between adjacent control cycles is less than the threshold determined based on the control cycle and the minimum identifiable acceleration, then the velocity command of the previous moment remains unchanged. The rate limiting rule is: the change in the velocity command vector between adjacent control cycles shall not exceed the limit value determined based on the maximum acceleration and the control cycle.

[0066] The above mechanism ensures that the motion commands of the mobile platform fluctuate smoothly within a narrow range during the phase before the prediction converges, effectively tolerating prediction errors and avoiding energy waste and mechanical shock caused by sudden motion changes. During the pre-positioning phase, each joint of the racket arm moves to a preset ready position: the power joints are preset to a preset ratio of the mid-range rotational speed, adjusting the joints to the ready-to-hit posture, laying the foundation for a rapid response during the precise return phase.

[0067] The nonlinear model predictive control algorithm takes the robot's current state, the point of impact, and the time of impact as inputs to calculate the target acceleration of each joint and the target acceleration of the mobile platform. The state vector includes the planar pose of the mobile platform, the linear velocity and spin angular velocity of the mobile platform, and the rotation angle and angular velocity of each joint of the robotic arm; the control vector includes the longitudinal and lateral accelerations of the mobile platform and the angular acceleration of each joint.

[0068] The state transition equation is constructed based on the robot's kinematic model, using forward Euler integrals, and is based on the current state vector. With control vector Calculate the state vector at the next time step The state transition equation is constructed based on the robot's kinematic model, and its specific form is as follows: (15); The state vector is determined by the pose of the mobile platform, the velocity of the mobile platform, the joint angles and angular velocities of the robotic arm, while the control vector is determined by the acceleration of the mobile platform and the angular acceleration of the joints. The sampling period is This is the state transition function. The state vector specifically includes the planar position and orientation angle of the mobile platform in the field coordinate system, the longitudinal and lateral linear velocities and spin angular velocities of the mobile platform, and the rotation angles and angular velocities of each joint of the robotic arm; the control vector specifically includes the longitudinal and lateral accelerations of the mobile platform and the angular accelerations of each joint.

[0069] State transition function It comprises a mobile platform kinematics subsystem and a robotic arm kinematics subsystem. For the mobile platform subsystem, based on the kinematic model of an X-shaped omnidirectional mobile platform, the forward kinematic mapping of the four wheel speeds and the mobile platform velocity is mapped from the three-dimensional velocity vector to the four-dimensional wheel velocity vector via a transformation matrix. Its inverse kinematic mapping is mapped from the four-dimensional wheel velocity vector to the three-dimensional velocity vector via a pseudo-inverse matrix. The elements of the transformation matrix are determined by the wheel radius, wheelbase parameters, and arrangement angles. For the robotic arm subsystem, the updates of each joint angle and angular velocity satisfy discrete kinematic equations containing acceleration terms.

[0070] The cost function corresponding to the nonlinear model predictive control algorithm is a weighted sum of the weighted compensation terms for the hitting position error, hitting velocity error, and hitting angle error in the prediction time domain, and the smoothing term of the control quantity in the control time domain. The hitting angle error is expressed as the inner product of the racket normal unit vector and the target normal unit vector. When both are unit vectors, it is equivalent to the cosine compensation term of the angle between the two normals, avoiding the dimensional problems and magnitude degradation risks associated with direct vector differences. Each weight is set independently and does not need to satisfy normalization constraints.

[0071] Cost function J in the prediction time domain With control time domain The internal definition is as follows: (16); In the formula, For the first The error vector of the ball's position during the strike; For the first Swing speed error; For the first Step control quantity change rate; To predict the time-domain step size; To control the time-domain step size; These represent the weights for position, velocity, angle, and smoothness, respectively; T is the vector transpose.

[0072] Constraints are hard-constrained at each prediction step using a quadratic programming solver, including: joint position limit constraints, joint velocity limit constraints, joint acceleration limit constraints, combined velocity limit constraints of the moving platform, combined acceleration limit constraints of the moving platform, wheel speed limit constraints, and normal unit vector constraints. For the omnidirectional X-shaped moving platform, nonholonomic constraints are not required to preserve its omnidirectional movement capability.

[0073] In the overall control process, the serving mechanism, as a component of the execution layer, is incorporated into the overall collaborative control framework in the second stage of the nonlinear model predictive control algorithm to achieve collaborative optimization of serving timing, serving posture, and wheel speed. At the same time, the serving mechanism interlocks with the return action through scene state judgment to ensure that only one action, either return or serve, is executed at any given time, thus avoiding mechanism conflicts.

[0074] In one or more embodiments, the control unit has a training mode database that stores multiple training modes. The control unit is also used to acquire user hitting data and evaluate the user's skill level based on a multi-dimensional weighted scoring model. Then, it adaptively matches training modes from the training mode database, generates corresponding serving parameter adjustment instructions and mobile platform target pose instructions, and sends them to the serving mechanism and the mobile platform.

[0075] The training modes include Fixed Point Mode, Follow Mode, Movement Mode, Random Mode, Weakness Mode, and Competition Mode. Each mode corresponds to different landing point planning strategies, serve parameter ranges, movement platform rules, and ball trajectory type libraries. The training modes are configured in progressive difficulty: Basic Mode corresponds to fixed positions and basic ball trajectories; Standard Mode corresponds to follow-movement and standard ball trajectories; Advanced Mode corresponds to full-court movement and complex ball trajectories; and Challenge Mode corresponds to competitive interaction and high-difficulty ball trajectories.

[0076] by Figure 8 For example, when a user's skill level is Professional (90-100 points), the matched training mode is the competitive mode, corresponding to the high clear, flat high clear, flat drive, slice drop, net shot, or smash in the ball type library; when a user's skill level is Advanced (75-90 points), the matched training mode is the moving mode, corresponding to the high clear, flat high clear, flat drive, slice drop, or net shot in the ball type library; when a user's skill level is Intermediate (60-75 points), the matched training mode is the follow mode, corresponding to the high clear, flat high clear, or flat drive in the ball type library; when a user's skill level is Beginner (0-60 points), the matched training mode is the fixed-point mode, corresponding to the high clear or flat high clear in the ball type library.

[0077] The ball-hitting data includes at least one of the following: ball landing point, return speed, return angle, positioning rate, and return height. A multi-dimensional weighted scoring model quantifies the quality of each dimension of the shot according to preset weights, generates a comprehensive score, and divides the user's skill level into several mutually exclusive levels based on preset thresholds, with no overlap between adjacent levels.

[0078] like Figure 7 As shown, the multi-dimensional weighted scoring model is defined as follows: (17); In the formula, The score is the overall score for each shot. For the first Sub-scores for each dimension of shot quality; The weights are as follows. The five-dimensional sub-scores are calculated based on the landing point deviation, the ratio of the return ball speed to the target speed, the contact time deviation, the robot's positioning deviation, and the return ball height deviation, respectively. The smaller the deviation, the higher the score. The closer it is to 100; the weights of each dimension can be set according to the actual situation, for example... , , , , .

[0079] The control unit is also used to: acquire data on multiple consecutive shots by the user, calculate the comprehensive score for each shot based on a multi-dimensional weighted scoring model, and then compare it with the difficulty adjustment threshold to dynamically adjust the training difficulty.

[0080] The difficulty adjustment threshold is dynamically determined based on the statistical characteristics of the sliding window of the user's historical shot data: a sliding window of preset length is maintained to record the comprehensive score of the user's most recent shots, and the mean and standard deviation of the comprehensive score within the sliding window are calculated; the difficulty adjustment threshold includes a difficulty increase threshold and a difficulty decrease threshold. The difficulty increase threshold is the mean plus the product of the sensitivity coefficient and the standard deviation; the difficulty decrease threshold is obtained by subtracting the product of the sensitivity coefficient and the standard deviation from the mean, and then limiting it by a preset minimum effective threshold.

[0081] The sensitivity coefficient can be adjusted according to user preferences. A larger value makes the threshold range wider and the difficulty adjustment more conservative, while a smaller value makes the threshold range narrower and the difficulty adjustment more sensitive.

[0082] In the control unit, a continuous counter mechanism is used to dynamically adjust the training difficulty. The process is as follows: When the combined score of the most recent preset number of shots is higher than the difficulty increase threshold, the cumulative upgrade counter is incremented. If the score is lower than the difficulty increase threshold, it is reset to zero. When the preset number of triggers is reached, the difficulty is automatically increased or the proportion of high-difficulty shots is increased. When the combined score of the most recent consecutive preset number of shots is lower than the difficulty reduction threshold, the downgrade counter is incremented. If the score is higher than the difficulty reduction threshold, it is reset to zero. When the preset number of triggers is reached, the difficulty is automatically reduced or the proportion of high-difficulty shots is reduced.

[0083] In addition, an absolute threshold protection mechanism is implemented: when a user's score repeatedly falls below a preset rest threshold, training is automatically paused and the user is prompted to rest, preventing overtraining and potential sports injuries. All training data is timestamped and supports local storage and cloud synchronization.

[0084] This invention, through a multimodal intelligent human-computer interaction module, not only supports multi-channel information interaction such as voice, touch, and lighting, but also constructs a closed loop of intelligent evaluation and adaptive adjustment based on user data: it quantifies user skill level through a multi-dimensional weighted scoring model, automatically matches training modes and ball types, and dynamically adjusts the difficulty, transforming the training process from one-way output to two-way adaptive intelligent coaching, significantly improving the training's relevance, enjoyment, and scientific rigor. This invention can complete incoming ball detection and early prediction in a short time, and quickly complete the mobile platform's positioning and the robotic arm's ball-hitting action after prediction convergence, meeting the real-time requirements of flat ball flight time. For smashes, provided the system has entered the return zone during the pre-positioning stage, it can quickly complete emergency mobile platform positioning and robotic arm emergency return, achieving a high smash return success rate.

[0085] In one or more embodiments, a control method for a badminton intelligent robot system includes the following steps performed in the control unit: Step 1: Receive the global flight trajectory of the badminton shuttlecock detected by multi-source collaborative visual perception, the calculated global pose of the robot, and the predicted local flight trajectory of the badminton shuttlecock; Step 2: Based on the global flight trajectory of the badminton shuttlecock, determine the current scene state and execute the matching action accordingly; scene states include continuous rally state, waiting for the shuttlecock to arrive state, and needing to serve state; When the scene is in the state of needing to serve, control the serving mechanism to perform the serving action. After the serve is completed, automatically switch to the waiting state to prepare for the next round of continuous fighting. When the scenario is in a continuous rally state, the type of incoming shuttlecock is determined by combining the global and local flight trajectories of the shuttlecock with the flight parameters, and the hitting strategy is adaptively matched. Based on the calculated global pose of the robot, a two-stage collaborative control of the mobile platform and the single-arm multi-axis racket robot is adopted to execute the matched hitting strategy and obtain the corresponding execution results, thereby dynamically adjusting the hitting strategy for the next cycle.

[0086] This embodiment senses the global image sequence and the near-field local image sequence of the badminton shuttlecock, thereby detecting the global flight trajectory of the shuttlecock, calculating the global pose of the robot, and predicting the local flight trajectory of the shuttlecock. Based on the global flight trajectory of the shuttlecock, it determines the current scene state and executes matching actions accordingly. Furthermore, in the continuous rally state, it dynamically adjusts the hitting strategy for the next cycle based on the execution results, thus realizing a complete closed-loop control from environmental perception, motion decision-making, precise execution to intelligent interaction.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A badminton intelligent robot system, characterized in that, include: The robot body (1), multi-source collaborative visual perception unit and control unit; the robot body (1) includes a mobile platform (2) and a single-arm multi-axis racket robotic arm (3) and a serving mechanism (4) mounted thereon, with a racket (5) disposed at the end of the single-arm multi-axis racket robotic arm (3); the serving mechanism (4) is connected to the control unit; The multi-source collaborative visual perception unit is used to: acquire global image sequences and near-field local image sequences of badminton shuttlecocks, and then detect the global flight trajectory of the badminton shuttlecocks, calculate the global pose of the robot, and predict the local flight trajectory of the badminton shuttlecocks and transmit them to the control unit. The control unit is used to: determine the current scene state based on the global flight trajectory of the badminton shuttlecock and execute matching actions accordingly; the scene state includes continuous rally state, waiting for the shuttlecock to arrive state, and needing to serve state; When the scene state is the state of needing to serve, control the serving mechanism (4) to perform the serving action. After the serving is completed, automatically switch to the state of waiting for the ball to arrive, and prepare for the next round of continuous fighting. When the scene is in a continuous rally state, the type of incoming ball is determined by combining the global flight trajectory and local flight trajectory of the badminton ball with the flight parameters of the badminton ball, and the hitting strategy is adaptively matched. Based on the calculated global pose of the robot, the two-stage collaborative control of the mobile platform (2) and the single-arm multi-axis racket robot (3) in the pre-positioning stage and the precise control stage is adopted to execute the matched hitting strategy and obtain the corresponding execution results, and then dynamically adjust the hitting strategy for the next cycle.

2. The badminton intelligent robot system as described in claim 1, characterized in that, In the control unit, the two stages of pre-positioning and precise control are as follows: The first stage is the pre-positioning stage. When the predicted local flight trajectory of the badminton shuttlecock has not yet converged, the control platform (2) moves to the approximate area of ​​the return point with adaptive proportional control, and the control of each joint of the single-arm multi-axis racket robot arm (3) moves to the ready position. The second stage is the precision control stage. After the predicted local flight trajectory of the badminton shuttlecock converges, based on the robot's global pose, hitting point and hitting time, a nonlinear model predictive control algorithm is used to calculate the target acceleration of each joint of the single-arm multi-axis racket robot arm (3) and the target acceleration of the mobile platform (2) in order to achieve the timing coordination between the omnidirectional movement of the mobile platform (2) and the precise hitting of the single-arm multi-axis racket robot arm (3).

3. The badminton intelligent robot system as described in claim 2, characterized in that, In the control unit, the cost function corresponding to the nonlinear model predictive control algorithm is: the weighted sum of the weighted compensation terms for the ball position error, ball speed error, and ball angle error in the prediction time domain and the smoothing term of the control quantity in the control time domain.

4. The badminton intelligent robot system as described in claim 1, characterized in that, The control unit is equipped with a training mode database, which contains a variety of training modes. The control unit is also used to acquire user hitting data and evaluate the user's skill level based on a multi-dimensional weighted scoring model. Then, it adaptively matches the training mode from the training mode database, generates corresponding serving parameter adjustment instructions and target pose instructions for the mobile platform (2), and sends them to the serving mechanism (4) and the mobile platform (2).

5. The badminton intelligent robot system as described in claim 1, characterized in that, The control unit is also used to: acquire data on multiple consecutive shots by the user, calculate the comprehensive score of each shot based on a multi-dimensional weighted scoring model, and then compare it with the difficulty adjustment threshold to dynamically adjust the training difficulty; The difficulty adjustment threshold is dynamically determined based on the statistical characteristics of the sliding window of the user's historical shot data: a sliding window of preset length is maintained, the comprehensive score of the user's most recent shots is recorded, and the mean and standard deviation of the comprehensive score within the sliding window are calculated; the difficulty adjustment threshold includes a difficulty increase threshold and a difficulty decrease threshold, and the difficulty increase threshold is the product of the mean, the sensitivity coefficient and the standard deviation. The difficulty reduction threshold is obtained by subtracting the product of the sensitivity coefficient and the standard deviation from the mean, and then applying a preset minimum effective threshold for amplitude limiting.

6. The badminton intelligent robot system as described in claim 5, characterized in that, In the control unit, a continuous counter mechanism is used to dynamically adjust the training difficulty. The process is as follows: When the combined score of the most recent preset number of shots is higher than the difficulty increase threshold, the cumulative upgrade counter is incremented. If the score is lower than the difficulty increase threshold, it is reset to zero. When the preset number of triggers is reached, the difficulty is automatically increased or the proportion of high-difficulty shots is increased. When the combined score of the most recent consecutive preset number of shots is lower than the difficulty reduction threshold, the downgrade counter is incremented. If the score is higher than the difficulty reduction threshold, it is reset to zero. When the preset number of triggers is reached, the difficulty is automatically reduced or the proportion of high-difficulty shots is reduced.

7. The badminton intelligent robot system as described in claim 1, characterized in that, The mobile platform (2) is a four-wheeled X-shaped omnidirectional mobile platform. It adopts an X-shaped symmetrical frame structure and is composed of four main longitudinal beams (21) and a central bearing platform (22) forming an overall bearing skeleton. The four main longitudinal beams (21) extend outward in an X-shape with the geometric center of the robot body (1) as the intersection point. The included angle between adjacent main longitudinal beams (21) is 90°. The ends of each main longitudinal beam (21) are rigidly connected to the wheel hub module mounting seat by bolts.

8. The badminton intelligent robot system as described in claim 1, characterized in that, The single-arm multi-axis racket robotic arm (3) is mounted on the robotic arm base, the ball-serving mechanism (4) is located inside the robotic arm base, and the robotic arm base is mounted on the center (2) of the mobile platform.

9. The badminton intelligent robot system as described in claim 1, characterized in that, The serving mechanism (4) includes a dual-wheel friction drive assembly (44), a ball supply trajectory (43), a ball storage tube (41), and a ball distribution wheel (42). The left and right wheels of the dual-wheel friction drive assembly (44) are driven independently, and different rotation modes are given to the badminton shuttlecocks through differential rotation. The ball storage tube (41) is connected to the ball supply trajectory (43) at the end through a flexible ball supply tube. The ball distribution wheel (42) is located at the outlet of the ball storage tube, and the badminton shuttlecocks in the ball storage tube (41) are supplied one by one in an orderly manner through the ball distribution wheel (42).

10. A control method for a badminton intelligent robot system, characterized in that, Based on the badminton intelligent robot system as described in any one of claims 1-9, the following steps are performed in the control unit: The system receives the global flight trajectory of the badminton shuttlecock detected by multi-source collaborative visual perception, calculates the global pose of the robot, and predicts the local flight trajectory of the badminton shuttlecock. Based on the global flight trajectory of the badminton shuttlecock, the current scene state is determined and the matching action is executed accordingly; the scene states include the continuous rally state, the waiting for the shuttlecock, and the need to serve the shuttlecock. When the scene state is the state of needing to serve, control the serving mechanism (4) to perform the serving action. After the serving is completed, automatically switch to the state of waiting for the ball to arrive, and prepare for the next round of continuous fighting. When the scene is in a continuous rally state, the type of incoming ball is determined by combining the global flight trajectory and local flight trajectory of the badminton ball with the flight parameters of the badminton ball, and the hitting strategy is adaptively matched. Based on the calculated global pose of the robot, the two-stage collaborative control of the mobile platform (2) and the single-arm multi-axis racket robot (3) in the pre-positioning stage and the precise control stage is adopted to execute the matched hitting strategy and obtain the corresponding execution results, and then dynamically adjust the hitting strategy for the next cycle.

Images