Multi-robot cooperation method and system for ice hockey training

Abstract

The invention relates to a multi-robot cooperation method and system for ice hockey training, and belongs to the technical field of ice hockey training.The method comprises the steps that a cooperation control platform dynamically allocates training roles for each training partner robot according to a preset training mode; all partner training robots sense environment information in real time and send the environment information to the cooperative control platform; the cooperative control platform fuses all the environment information to generate a global training situation map used for reflecting the position and motion relation of all the recognition targets in the training field; and the cooperative control platform generates a cooperative motion control instruction for each partner training robot according to the global training situation map and the distributed training role, so that the partner training robots obtain and execute the cooperative motion control instruction. According to the ice hockey training method, a real competition scene can be simulated, environment perception and intelligent decision-making capabilities are realized, and multi-agent cooperation can be realized.

Description

Technical Field

[0001] This application relates to the field of ice hockey training technology, and in particular to a multi-robot collaborative method and system for ice hockey training. Background Technology

[0002] Ice hockey is a high-speed, highly competitive team sport, and the quality of its training heavily depends on the skill level of the sparring partners and the realism of the training environment. Traditional ice hockey training relies primarily on the coach's subjective experience and a limited number of live sparring partners, which results in high costs, difficulty in standardization, and an inability to provide a high-intensity, dynamic, and data-driven competitive environment.

[0003] To overcome the aforementioned problems, auxiliary training equipment such as hockey puck machines have emerged in the existing technology. However, these devices are limited in function, typically only capable of serving from fixed points or along fixed trajectories. They lack the ability to perceive the environment (such as players and the hockey puck) and cannot simulate the dynamic positioning, passing, and cutting plays of teammates and opponents in real matches. Consequently, they suffer from insufficient training interactivity and tactical depth, and therefore require improvement. Summary of the Invention

[0004] In order to provide an ice hockey training solution that can simulate real game scenarios, has environmental perception and intelligent decision-making capabilities, and enables multi-agent collaborative cooperation, this application provides a multi-robot collaborative method and system for ice hockey training.

[0005] Firstly, this application provides a multi-robot collaborative method for ice hockey training. The executing entity of the multi-robot collaborative method for ice hockey training is a multi-robot collaborative system for ice hockey training. The multi-robot collaborative system includes a collaborative control platform and several training robots communicatively connected to the collaborative control platform. The multi-robot collaborative method for ice hockey training includes the following steps: The collaborative control platform dynamically assigns training roles to each training robot according to a preset training mode; wherein, the training roles include at least an offensive player with the ball and a defensive interceptor. All training robots perceive environmental information in real time and send the environmental information to the collaborative control platform; wherein, the environmental information includes at least the relative pose and motion state of the identified targets appearing within the perception range of the training robots; the identified targets include at least ice hockey, the trained player, and the training robots; The collaborative control platform integrates all the environmental information to generate a global training situation map that reflects the position and motion relationship of all identified targets in the training field. The collaborative control platform generates collaborative motion control commands for each training robot based on the global training situation map and the assigned training roles, so that the training robot can acquire and execute the collaborative operation control commands and realize collaborative hockey training.

[0006] By adopting the above technical solutions, firstly, by dynamically assigning training roles such as "puck attacker" and "defender interceptor" to the training robots, the system can simulate multi-role tactical cooperation in real matches (such as numerical superiority, offensive-defensive transitions, and passing and cutting combinations), completely overcoming the shortcomings of existing single-robot training modes, which are characterized by limited functionality and lack of tactical depth. Secondly, through the fusion of distributed perception and a global training situation map, the system gains full-time, precise perception capabilities of all dynamic targets (hockey puck, players, and robots) within the training area. This transforms training from a pre-programmed, fixed process into a dynamic interactive process that responds to athlete behavior in real time. Thirdly, based on the global training situation map and assigned roles, collaborative motion control commands are generated for each robot, enabling multiple robots to move as an organic whole, rather than as independent entities. This achieves complex team tactical cooperation, such as flanking maneuvers, support plays, and zone defense. Finally, the entire training process is perceived and recorded in real time, and all interactive events (such as passes, hits, and interceptions) can be quantified, providing the possibility of generating data-driven training reports and helping coaches and athletes conduct accurate performance evaluations and tactical analyses.

[0007] Optionally, the collaborative control platform generates collaborative motion control commands for each training robot based on the global training situation map and the assigned training roles, including the following steps: The collaborative control platform assigns each training robot as a subordinate object based on its training role, and matches each subordinate object with a master object, establishing a master-slave relationship between each subordinate object and its matched master object. The master object is either the training robot or the player being trained. The collaborative control platform determines the collaborative motion trajectory for the training robot in real time based on the relative positional relationship between subordinate and dominant objects in the preset master-slave formation, and the real-time movement trajectory of the dominant object with a master-slave relationship with the training robot in the global training situation map. Based on the collaborative motion trajectory, it generates collaborative operation control instructions. The collaborative motion trajectory is used to limit the real-time position of subordinate and dominant objects with a master-slave relationship to maintain the master-slave formation. The collaborative operation control instructions are configured to drive the training robot to move along the collaborative motion trajectory and perform role tasks corresponding to the assigned training role, so as to realize collaborative hockey training with the trained player.

[0008] By adopting the above technical solutions, firstly, by pre-setting master-slave formations (such as triangle offense and diamond defense) for each master-slave pair, the system can drive multiple robots to automatically form and maintain complex tactical formations. This overcomes the limitations of existing equipment that can only perform irregular or simple trajectory movements, making the training content highly realistic at the tactical level. Secondly, the cooperative movement trajectory of the subordinate objects is dynamically generated based on the real-time movement trajectory of their master objects (especially the trained players). This means that the robot's behavior can respond in real time to every action of the player (such as changing direction and accelerating), creating an interactive experience of "intelligent coaching" rather than mechanically executing preset programs. Furthermore, by adopting a hierarchical structure of master-slave relationships, the complex "many-to-many" collaborative problem is decomposed into multiple simple "one-to-one" tracking problems. This decomposition greatly simplifies the complexity of the path planning algorithm, reduces the computational burden on the central control platform, and enables the system to run efficiently and in real time. Finally, the generated collaborative operation control commands simultaneously include movement commands and role task commands. This ensures that the robot executes the correct tactical actions (such as passing the ball immediately after running to the correct position) at the correct time and place, achieving a seamless connection between movement and tactical execution, and improving the overall integrity and smoothness of training.

[0009] Optionally, the collaborative control platform dynamically assigns training roles to each training robot according to a preset training mode, including: The collaborative control platform determines all training roles included in a preset training mode; and assigns a training robot to each training role based on a preset optimization strategy; wherein the optimization strategy includes at least one of the following: assigning according to a preset priority for the training role and / or training robot; evaluating the overall benefit of swapping the training roles of any two training robots, and if the overall benefit increases by more than a preset threshold, then swapping the training roles of the corresponding training robots; and assigning a training robot whose actual state can meet the preset capability requirements according to the preset capability requirements required by the task corresponding to the training role. The collaborative control platform assigns each training robot as a subordinate object based on its training role, and matches each subordinate object with a dominant object, including: The collaborative control platform, based on the training role of each training robot, sequentially assigns each training robot as a subordinate object according to the priority of each training role from high to low, and matches a dominant object to the subordinate object. The process of matching a subordinate object with a dominant object includes: A candidate object set is determined for the subordinate object, and the initial cooperative motion trajectory of each candidate object in the candidate object set is calculated. The initial cooperative motion trajectory is compared with the cooperative motion trajectories of all successfully matched training robots to perform spatiotemporal conflict detection. The target object corresponding to the initial cooperative motion trajectory does not produce spatiotemporal conflict is selected from the candidate object set, and the target object is used as the dominant object for successful matching of the subordinate object.

[0010] By adopting the above technical solutions, firstly, through an optimized allocation strategy based on priority and capability requirements, the system ensures that the most important tactical roles are filled by robots in optimal condition, in the best position, and with the greatest expertise. This avoids the "capability mismatch" problem that may result from random or rotating allocations, ensuring that the initial configuration for each training session is the optimal solution under the current conditions, thereby maximizing training effectiveness and overall team performance. Secondly, the introduction of a role-swapping strategy gives the system a "reflection and optimization" capability. Even if the initial allocation is reasonable, the system can still discover a better global configuration by evaluating the overall benefits of role swapping (such as shortening the total path and increasing the safety space). This demonstrates a higher level of system intelligence, enabling it to dynamically find and lock in the globally optimal or even better collaborative solutions. Furthermore, during the matching phase of establishing master-slave relationships, the feasibility of preliminary solutions is screened through spatiotemporal conflict detection. This is equivalent to solving future "traffic conflict" problems at the "tactical blueprint" stage, preventing the risk of collisions, blockages, or mutual interference between robots during training from the root, ensuring high safety, high smoothness, and uninterrupted training. Finally, the dominant targets are matched according to the priority of the training roles from high to low. This ensures that, given the limited system resources (especially high-quality dominant targets), the core tactical roles (such as ball-handling attackers) can obtain the most suitable configuration resources first, thereby ensuring that the core tactical intentions can be executed most effectively.

[0011] Optionally, the method further includes: During collaborative training, each training robot continuously monitors its own status information and sends the status information to the collaborative control platform; The collaborative control platform is used to determine that a spatiotemporal conflict is about to occur when it detects that the collaborative motion trajectories of two first training robots intersect or the minimum distance is less than a preset safety distance in the future. It then triggers and executes a trajectory correction strategy to adjust the collaborative motion trajectory of the corresponding training robots before the spatiotemporal conflict occurs, thereby resolving the spatiotemporal conflict. Herein, the first training robot refers to any training robot. The collaborative control platform is used to determine that a runtime capability conflict has occurred when it detects that the second training robot cannot meet the robot capability requirements of the assigned training role due to a state change, and triggers the execution of a distributed negotiation resolution strategy for the second training robot; wherein, the second training robot refers to any training robot. The distributed negotiation solution strategy includes: selecting the training robot that will undertake the role task of the second training robot as the winning training robot, assigning the role task of the second training robot to the winning training robot, and updating the cooperative motion control instructions of the bidding training robot so that the winning training robot executes the updated cooperative motion control instructions.

[0012] By adopting the above technical solutions, firstly, through real-time trajectory prediction and correction, the system can intelligently adjust the robot's path before a physical collision occurs. This avoids unexpected interruptions, pauses, or resets during training due to robot collisions or jamming, enabling high-intensity continuous tactical training and greatly enhancing the athlete's training immersion and experience. Secondly, when any robot suddenly "goes offline" due to power depletion, mechanical failure, or other reasons (i.e., a runtime capability conflict occurs), the distributed negotiation resolution strategy can automatically and quickly find a "substitute" among the remaining robots, achieving a smooth transfer of tasks. This makes the system fault-tolerant to individual failures, preventing single-point failures from paralyzing the entire training system and ensuring the successful completion of training. Furthermore, this solution, combined with the aforementioned solutions, forms a triple safety net: the first layer (prevention): initial conflict detection during the matching phase; the second layer (avoidance during operation): real-time trajectory correction during runtime (spatiotemporal conflict resolution in this solution); and the third layer (remediation afterward): task transfer in the event of a serious capability failure (capability conflict resolution as claimed in this claim), thus covering risk management throughout the entire lifecycle from the planning phase to the execution phase. Finally, the distributed negotiation mechanism makes the robot team a dynamically reconfigurable system. Theoretically, during training, faulty robots can be removed or fully charged robots can be added at any time, and the system can automatically integrate resources and redistribute tasks, achieving "never-ending" training services.

[0013] Optionally, the matching of the subordinate object with the dominant object further includes: When multiple target objects are selected from the candidate object set, a master-slave condition is preset for each target object. The master-slave condition is configured to ensure that, at any given time, the master-slave condition of one and only one target object is satisfied, based on mutually exclusive state variables or preset priorities. The method further includes: During the training process, for a training robot with multiple target objects, the master-slave conditions of all target objects are evaluated in real time. The target object that meets the master-slave conditions is used as the master object to establish a master-slave relationship with the corresponding training robot as the subordinate object, so as to realize the real-time update of the master-slave relationship of the training robot with multiple target objects.

[0014] By adopting the above technical solutions, a training robot is no longer fixed to a single dominant entity. It can dynamically switch its dominant entity during training based on preset master-slave conditions. This allows a robot to continuously play different tactical roles in a single offensive or defensive round (e.g., first following a player to provide support, then switching to following another robot for cross-cutting), thus simulating extremely complex and varied tactical combinations with a limited number of robots; the robot's "leadership" is designed as a dynamically switchable resource driven by the battlefield situation. Secondly, the robot's behavior is no longer merely tracking, but conditional and intelligent tracking. The system can automatically select the dominant entity that best matches the current tactical intent based on key state variables such as puck position, training stage, and player movements. This allows the robot's behavior to intelligently adapt to changes in the game's pace, achieving a leap from "mechanical execution" to "tactical understanding." Finally, through carefully designed mutually exclusive master-slave conditions, the system ensures that in any complex battlefield environment, a robot will only obey the instructions of a single "leader" at any given moment. This fundamentally avoids the problems of "decision paralysis" or "behavioral conflict" caused by multiple conditions being met simultaneously, and ensures the coordination, stability and predictability of the behavior of the entire multi-robot system.

[0015] Optionally, assigning the role and task of the second training robot to the winning training robot and updating the cooperative motion control instructions of the bidding training robot includes: The role and task of the second training robot are assigned to the winning training robot. The winning training robot enters a multi-role fusion state. The winning training robot needs to maintain its own original role and task as well as the role and task of the second training robot at the same time. During the training process, the master-slave conditions corresponding to all roles and tasks maintained by the winning training robot are continuously monitored in parallel. Based on the evaluation results of the master-slave conditions, the primary task to be executed and the master-slave relationship corresponding to the primary task to be executed are determined from all roles and tasks maintained by the winning training robot. The cooperative motion control commands of the winning training robot are updated based on the master-slave relationship until the training ends.

[0016] By adopting the above technical solutions, firstly, when a robot undertakes a new task, it may face the dilemma of conflicting task instructions from two roles (for example, the original role requires it to defend, while the new role requires it to attack). This solution dynamically determines the "primary task," forcing the robot to follow only one clear tactical objective at any given moment, thus ensuring that its action path and tactical intentions are clear and predictable. Secondly, after a robot malfunctions, the system does not simply find a "substitute," but through role fusion, allows a healthy robot to simultaneously shoulder more tactical responsibilities. This decision-making mechanism ensures that even in the event of attrition, the system can prioritize the execution of the most critical tactical tasks (i.e., the highest priority tasks) without interruption, thereby maximizing the tactical strength and integrity of the original training plan. Next, the master-slave relationship of the winning robot is dynamically updated as its primary task changes. This means that it always follows the most important "leader," ensuring that its position and behavior in the entire collaborative network always match the current most critical tactical phase. Finally, the continuous operation design "until training ends" makes this decision-making mechanism a permanent daemon process. It can handle complex situations such as role overlap caused by multiple failovers during training, ensuring that the system maintains decision-making robustness throughout its entire lifecycle, and providing core support for achieving long-term, unattended automated training.

[0017] Optionally, determining the primary task and the corresponding master-slave relationship from all role tasks maintained by the winning training robot based on the evaluation results of the master-slave conditions includes: Generate and include the role tasks corresponding to all satisfied master-slave conditions into a valid task set; Based on the preset global task priority list, the highest priority role task is selected from the current set of valid tasks and determined as the primary task to be executed in the current control cycle. Recall the character action instructions corresponding to the primary task from the preset behavior library; The target object corresponding to the master-slave condition of the primary execution task is taken as the master object of the winning training robot, and a master-slave relationship is established.

[0018] By adopting the above technical solution, firstly, by generating an effective set of tasks and sorting them according to global priority, the system transforms a complex behavior selection problem into a clear and computable optimization problem. This ensures that the robot's behavior automatically aligns with the most tactically valuable target at any given moment, making optimal decisions without external intervention. Secondly, by strongly binding the primary task to its master-slave relationship, it ensures that the robot's tactical actions (called from the behavior library) and mobile collaboration (following the dominant object) originate from the same tactical intent. This avoids the possibility of actions and movements becoming disconnected, making all behavioral components of the robot coordinated and consistent, serving the same highest-priority task. A deterministic mapping chain of "task → behavior → dominant object" is established, eliminating inconsistencies in internal decision-making.

[0019] Secondly, this application provides a multi-robot collaborative system for ice hockey training, characterized in that the multi-robot collaborative system for ice hockey training includes a collaborative control platform and several training robots communicatively connected to the collaborative control platform. The collaborative control platform is used to dynamically assign training roles to each training robot according to a preset training mode; wherein, the training roles include at least an offensive player with the ball and a defensive interceptor. All training robots are used to perceive environmental information in real time and send the environmental information to the collaborative control platform; wherein, the environmental information includes at least the relative pose and motion state of the identified targets appearing within the perception range of the training robots; the identified targets include at least ice hockey, the trained player, and the training robots; The collaborative control platform is used to integrate all the environmental information to generate a global training situation map that reflects the position and motion relationship of all identified targets in the training field. The collaborative control platform is used to generate collaborative motion control instructions for each training robot based on the global training situation map and the assigned training role, so that the training robot can obtain and execute the collaborative operation control instructions and realize collaborative hockey training.

[0020] Thirdly, this application provides a multi-robot collaborative device for ice hockey training, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any of the first aspects.

[0021] Fourthly, this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and execute the method described in any of the first aspects.

[0022] In summary, this application includes the following beneficial technical effects: This application combines dynamic role allocation, global environment perception, and multi-agent collaborative control to create a high-intensity, highly variable, and interactive ice hockey training environment that can simulate the intensity and complexity of real games. It effectively solves the core pain points of traditional training methods, such as high cost, difficulty in standardization, and weak tactical effectiveness, and represents a significant technological advancement in the field of ice hockey training. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0024] Figure 1 This is a flowchart illustrating a multi-robot collaborative method for ice hockey training disclosed in an embodiment of this application. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1 This application will be described in further detail.

[0026] This application discloses a multi-robot collaborative method for ice hockey training. The implementing entity is a multi-robot collaborative system for ice hockey training, which specifically includes a collaborative control platform and several training robots communicatively connected to the collaborative control platform.

[0027] Regarding the training companion robot, it should be noted that: The training robot comprises a robot body, a drive unit for moving the robot body, and an environmental perception module and a communication module mounted on the robot body. The drive unit includes a two-wheel differential structure and servos, which work together to achieve precise drive and control of the robot body. The hardware configuration of the two-wheel differential structure is existing technology and will not be elaborated upon. Its core lies in supporting and moving the robot body through two independent wheels, each driven by a motor (left drive wheel and right drive wheel), and one or more omnidirectional driven wheels (such as Mecanum wheels or ball wheels). The behavior decision control module issues high-level movement commands to the drive unit. These commands typically include a target linear velocity vector V (including magnitude and direction) and a target angular velocity ω. The low-level controller within the drive unit (such as a microcontroller or motor drive board) calculates the high-level (V, ω) command into the target rotational speeds of the two drive wheels based on the two-wheel differential kinematic model. The calculation formulas are as follows: Left wheel speed ω_L = (V - ω * L / 2) / R, Right wheel speed ω_R = (V + ω * L / 2) / R. Where ω_L and ω_R are the target speeds (radians per second or converted to RPM) of the left and right drive wheels, respectively; V is the target linear velocity of the robot's center of mass; ω is the target angular velocity of the robot around its center of mass; L is the wheelbase between the two drive wheels (fixed mechanical parameter); and R is the radius of the drive wheels (fixed mechanical parameter). The underlying controller uses a PID control algorithm to compare the actual speed feedback from the drive wheel motor encoders with the calculated target speeds, dynamically adjusting the duty cycle of the PWM (Pulse Width Modulation) signals output to the left and right wheel motors, thereby precisely controlling the motors to reach and maintain the target speeds. Specific movement actions include forward / backward (when ω = 0 and V ≠ 0, ω_L = ω_R, the two wheels rotate at the same speed and in the same direction (or in opposite directions), translation (this is existing technology and will not be elaborated on. It usually requires special wheel systems (such as Mecanum wheels, omnidirectional wheels) in conjunction with differential control. Its logic is to control the velocity vectors generated by all wheels to combine into a pure lateral velocity), and zero-radius rotation (when V = 0 and ω ≠ 0, ω_L = -ω_R, the two wheels rotate at equal speeds but in opposite directions, and the robot rotates in place with its center of mass as the center).

[0028] The servo motor's drive logic is as follows: 1. Rotation control: Method one is controlling the driven wheels (the servo motor's output shaft is connected to the steering mechanism of one or more directional driven wheels. When the behavior decision module requires the robot to make precise pointing adjustments with a non-zero radius (e.g., accurately aligning the hitting surface with a teammate before passing the ball), it sends a target angle pulse signal to the servo motor. The control circuit inside the servo motor drives the motor to rotate and rotates the output shaft to a specified angle through the gear set, thereby causing the driven wheels to deflect. This generates a deflection torque during the robot's movement, assisting it in quickly and accurately adjusting the robot's orientation); Method two is controlling the rotating platform (the servo motor drives a small rotating platform equipped with a vision or hitting mechanism. This is existing technology and will not be elaborated further. Its logic is to receive angle commands and rotate the platform relative to the robot body, thereby achieving the orientation of specific components without changing the overall robot's movement trajectory). 2. Braking control includes differential reverse braking (this is existing technology and will not be elaborated on further. After the behavior decision module issues an emergency stop command, the drive unit briefly controls the two drive wheels to reverse at high speed, generating braking force opposite to the direction of movement to achieve an emergency stop) and mechanical braking (the servo controls a mechanical brake pad through a linkage or cam mechanism. When braking is required, the servo receives a signal and drives the brake pad to press against the drive wheel or a dedicated brake disc, achieving braking through friction).

[0029] The environmental perception module included in the training robot is a hardware and software integrated module, comprising a camera, an IMU (Inertial Measurement Unit), and an image processing chip (such as an embedded GPU or VPU). The environmental perception module identifies targets (such as ice hockey pucks, players, and other training robots) within its field of view, calculates the three-dimensional position (X, Y, Z coordinates) of the identified target relative to the camera coordinate system, and then processes the position data of the same identified target in a continuous frame sequence using a Kalman filter to smooth noise and estimate the target's motion state, including instantaneous velocity vectors (Vx, Vy, Vz) and direction of motion. The final output is the environmental information described in S102, which is then sent to the collaborative control platform.

[0030] Specifically: A camera is used to acquire RGB images at a preset rate (e.g., 30fps) after training begins. An IMU measures the robot's linear acceleration (ax, ay, az) and angular velocity (wx, wy, wz) in three axes at a higher frequency (e.g., 200Hz). An image processing chip ensures that each frame is precisely aligned in time with a set of IMU data using hardware timestamps. The image processing chip also identifies targets (including hockey pucks, players, and goals) using a pre-trained and running convolutional neural network model (e.g., a YOLO or SSD-based object detection model). The convolutional neural network model is trained on a large dataset containing images of hockey pucks, players (assuming players wear the same uniform, using the uniform as a reference for recognition), and goals. The input to the model is each frame of the image captured by the camera, and the output is a specific visual recognition result. The visual recognition result is the bounding box (i.e., selecting the target in the image with a rectangular structure) and class probability of each recognition target (i.e., hockey puck, player, goal) in the image. For example, the model will output [hockey puck, confidence 98%, bounding box coordinates (x1, y1, x2, y2)] and [player, confidence 95%, bounding box coordinates ...].

[0031] For the target identified from the image, the image processing chip also uses a preset algorithm (such as the PnP algorithm mentioned in the embodiments of this application) to calculate the geometric relationship between the pixel position of the target in the image and its real-world size, and calculates the three-dimensional position (x_rel, y_rel, z_rel) and pose of the target relative to the camera. This is the relative pose of the target (i.e., the position information mentioned in S101). It should be noted that the default PnP algorithm knows the internal parameters of the camera (such as focal length and optical center, which can be obtained through pre-calibration) and the physical size of the target (such as the standard diameter of an ice hockey puck, the length, width and height of a goal, and the height of a player). Define a state vector [X, Y, θ, Vx, Vy, ω]^T for the robot body, that is: the robot's global position (X, Y), orientation θ, velocity (Vx, Vy) and turning angular velocity ω; where T represents transpose, changing it from a row vector to a column vector.

[0032] The image processing chip also pre-runs a filtering algorithm (such as extended Kalman filtering). This algorithm is used to predict the global position and attitude of the robot body at the current moment based on the robot body's global position and velocity at the previous moment and the current IMU data (acceleration a, angular velocity ω). The specific prediction steps are as follows: using the optimal state estimate of the previous moment and the acceleration measured by the IMU in the current time period, predict the robot state (including position and velocity) at the current moment. This prediction result is called the prior state estimate.

[0033] Based on the state X_{k-1} of the previous time step (e.g., k-1), the predicted state vector X_k of the current time step (e.g., k) can be obtained using the formula: X_k = X_{k-1} + V_{x, k-1} * Δt + 0.5 * a_x * Δt^2. Here, X_k refers to the predicted position of the robot body in the X-axis direction of the global coordinate system at the current time k. This is a priori estimate because it is calculated only based on past motions and IMU input. X_{k-1} represents the optimal estimated position of the robot body in the X-axis direction of the global coordinate system at the previous time step k-1, which is the most reliable position obtained after the previous round of Kalman filtering correction. V_{x, k-1} represents the optimal estimate of the robot's velocity in the X-axis direction at the previous time k-1, Δt is the time interval between the previous and current times, and a_x is the average acceleration in the X-axis direction measured by the IMU from time k-1 to time k. Correspondingly, X_0, i.e., the position and velocity corresponding to k-1=0, is the initial position (X0, Y0) of the moving robot before training begins, with a velocity of 0. The formula X_k = X_{k-1} +V_{x, k-1} * Δt + 0.5 * a_x * Δt^2 is the update formula for the first element X (i.e., the X-axis position) in the state vector. Similarly, the formulas for Y (i.e., the Y-axis position), Vx (X-axis velocity), and Vy (Y-axis velocity) in the state vector are: Y_k = Y_{k-1} + V_{y, k-1} * Δt + 0.5 * a_y * Δt^2; V_{x, k} = V_{x, k-1} + a_x * Δt; V_{y, k} = V_{y, k-1} + a_y * Δt.

[0034] After obtaining the prior state estimate X_k^-, the image processing chip needs to correct it to obtain the optimal state estimate. The correction steps are as follows: First, the observation value Z_k needs to be obtained. The method of obtaining Z_k includes two aspects: one is to calculate the pose change through vision. Assume that at time k-1 and time k, the robot's camera sees the same site marker (such as the apex of the boundary line marked by the training site). Visual algorithms (such as feature point matching) can be used to calculate the movement of the marker in the image. Using this movement information and the camera model, the robot's displacement and rotation from time k-1 to time k can be directly calculated. That is, the pixel coordinates (u1, v1) of the site marker in frame k-1, the pixel coordinates (u2, v2) in frame k, and the camera model (known internal parameters, such as focal length and principal point) are used to calculate, through geometric calculations (e.g., essential matrix or homography matrix decomposition), the rotation and translation that the camera (i.e., the robot itself) must have undergone to cause this pixel change are calculated. The result is the robot's displacement (Δx, Δy, Δz) and rotation (Δα, Δβ, Δγ) from time k-1 to time k. This displacement can be used as the observation value Z_k.

[0035] Secondly, it involves observing stationary targets: the robot identifies a stationary target, such as a goal. Using the PnP algorithm, the position (x_rel, y_rel) of the goal relative to the robot can be calculated. If the absolute position (X_goal, Y_goal) of the goal in the global coordinate system is known, then through coordinate transformation, the global position (X_cam, Y_cam) of the robot itself can be deduced. This calculated position (X_cam, Y_cam) can then be used as the observation value Z_k.

[0036] Finally, the extended Kalman filter algorithm weighted and fused the predicted value X_k^- and the observed value Z_k to correct the prediction error and obtain a smoother and more accurate final estimate, thus updating the prediction result. The specific fusion calculation formula is: X_k = X_k^- + K * (Z_k - H * X_k^-); where H is the observation matrix, a preset fixed matrix used to map the state vector X_k^- to the space where the observation value is located; the X_k obtained after this weighted fusion is the optimal and smoothest state estimate at the current time k, which is used to calculate the robot's body state at the next time k+1, realizing the cyclic real-time calculation of the robot's body state (i.e., position and velocity).

[0037] Finally, combining the robot's own state vector X_k calculated above with the relative position of the target identified by visual recognition, the absolute position (X_puck, Y_puck) and (X_player, Y_player) of the target (hockey puck, trained player, goal) in the global field coordinate system are calculated through coordinate transformation. By differentiating the absolute positions of consecutive frames (e.g., v = (pos_t - pos_{t-1}) / Δt), the target's velocity (Vx_puck, Vy_puck) and direction of motion are estimated, which are used as the target's state information. In other words, the environmental perception module ultimately outputs a quantized dynamic situation map containing timestamps, including: the target's position, velocity vector, and direction vector relative to the robot, as well as the robot's position and posture after filtering (i.e., the prediction results mentioned above).

[0038] The training robot connects to the collaborative control platform via a wireless ad hoc network through a communication module (such as Wi-Fi, Zigbee, or ESP-NOW). The collaborative control platform then works in tandem with the training robot to execute a multi-robot collaborative method for ice hockey training. The corresponding execution steps are as follows: S101, the collaborative control platform dynamically assigns training roles to each training robot according to a preset training mode; wherein, the training roles include at least an offensive player with the ball and a defensive interceptor; Specifically, S101 includes the following sub-steps: The collaborative control platform determines all training roles included in a preset training mode; and assigns a training robot to each training role based on a preset optimization strategy. The optimization strategy includes at least one of the following: allocation based on preset priorities for training roles and / or training robots; evaluation of the overall benefit of swapping any two training robot roles; if the overall benefit increases beyond a preset threshold, swapping the corresponding training robot roles; and assigning a training robot whose actual state meets the preset capability requirements of the task corresponding to the training role.

[0039] During implementation, the following preparatory work will be carried out before the training begins: staff will pre-deploy multiple fixed vision modules (such as wide-angle or panoramic cameras, specifically mobile phone cameras fixed on the sidelines using a mobile phone holder) on the training field to collect image information on the training field; at the same time, staff will select the required number of training robots and place the corresponding training robots in known locations on the training field (such as the center circle, left flank, right flank, and shooting area).

[0040] All training robots and fixed vision modules at the field are powered on and started. The training robots and fixed vision modules are connected to the collaborative control platform via a wireless self-organizing network through a communication module (such as Wi-Fi, Zigbee, or ESP-NOW). Coaches or other users access the collaborative control platform by accessing a pre-installed APP on their monitoring terminal (such as a mobile phone or tablet) and interact with the corresponding APP to select the training mode for the current training session (hereinafter referred to as the current training mode) from the pre-loaded training mode software library, thereby forming a selection command. The collaborative control platform learns the current training mode contained in the selection command and loads the preset configuration file corresponding to the current training mode. The training modes include: multi-on-one offense and defense, group confrontation, fixed tactical formation training, etc. The preset configuration file corresponding to each training mode contains the training roles (such as ball-handling attackers, off-ball receivers, defensive interceptors, etc.), the role tasks of each training role (such as running and receiving, coordinating passing and cooperation, defensive interception, pressing and stealing, etc.), and the preset capability requirements required to perform the corresponding role tasks. The preset capability requirements are specifically reflected in the power consumption required for the training robot to perform the corresponding role tasks, or the absence of predetermined faults during training (such as being unable to communicate with other training robots and / or collaborative control platforms due to disconnection).

[0041] After completing the above preparations, the collaborative control platform assigns a training robot to each training role based on the current training mode and the corresponding configuration file. The specific assignment logic is based on a preset optimization strategy, which includes the following three sub-strategies: 1. Priority sorting and allocation; 2. Role swapping strategy; 3. Role demotion strategy; The specific process of allocation using these three sub-strategies is as follows: The collaborative control platform first sorts the training roles from highest to lowest priority according to the preset priorities for each training role in the configuration file (e.g., the ball-handling attacker has a priority of 5, and the defensive interceptor has a priority of 3). Then, for the training role with the highest current priority, the collaborative control platform selects a training robot from all available training robots (available training robots refer to training robots that have not yet been assigned a training role) whose current state best meets the preset ability requirements of that role (e.g., remaining battery is higher than 80%) and has the highest priority score. The priority score is calculated as follows: Priority Score = 0.5 * (Remaining Battery Percentage) + 0.3 * (1 / Distance between the current position of the training robot and the preset activity area of ​​the role) + 0.2 * (1 - Recent Task Load); where the remaining battery percentage is internal status data of the training robot. The power management system of each training robot can calculate the remaining battery percentage in real time. This data is periodically reported to the collaborative control platform through the communication module of the training robot (e.g., once per second), or actively retrieved by the platform when role allocation is required. Furthermore, in the configuration file of the training mode, one or more main activity areas are predefined for each training role. For example, the activity area of ​​a defensive interceptor role might be defined as a rectangular area of ​​"home defense zone". For ease of calculation, the collaborative control platform usually simplifies this area to a core reference point (X_zone, Y_zone), such as the center point of the rectangle. The distance between the current position of the training robot (measured by the environmental perception module contained in the training robot) and the preset activity area of ​​the role is the distance to this core reference point (X_zone, Y_zone). Furthermore, regarding recent task load, it's important to note that the collaborative control platform maintains a "task load value" for each training robot. This value can be an exponentially decaying weighted sum. For example, each time a role is assigned to a robot, a "base weight" is added to its load value (e.g., 1.0 for high-energy-consuming roles, 0.5 for low-energy-consuming roles). This total load value then decays exponentially over time (e.g., multiplied by 0.9 every minute). During a new round of allocation, the platform reads the current task load value of each robot and normalizes it to the [0, 1] interval. Normalization can be based on the maximum load value among all robots: Recent Task Load = Current Training Robot Load Value / Maximum Load Value among All Training Robots, thus obtaining the specific numerical value of the recent task load. This ultimately achieves priority-based allocation.

[0042] In the priority allocation process, if it is found that no training role can be matched with a training robot that meets the preset ability requirements of the training role (such as requiring battery > 90%), the role demotion strategy is triggered: the collaborative control platform replaces the training role with a substitute role with similar tactical intentions but lower ability requirements according to the predefined demotion path (e.g., ball-handling attacker → off-ball support player). The platform then re-executes the priority allocation process for this new role. If the role demotion strategy is still triggered, the next substitute role is found based on the preset demotion path, and so on, until a training robot is matched for all training roles.

[0043] S102, all training robots perceive environmental information in real time and send the environmental information to the collaborative control platform; the environmental information includes at least the relative pose and motion state of the identified targets appearing within the perception range of the training robots; the identified targets include at least ice hockey, the trained player, and the training robots. S103, the collaborative control platform integrates all environmental information to generate a global training situation map that reflects the position and motion relationship of all identified targets in the training field; S104, the collaborative control platform generates collaborative motion control instructions for each training robot based on the global training situation map and the assigned training roles, so that the training robot can acquire and execute collaborative operation control instructions and realize collaborative hockey training.

[0044] Specifically, S104 includes the following steps: S1041, The collaborative control platform, based on the training role of each training robot, treats each training robot as a subordinate object and matches a dominant object to each subordinate object, establishing a master-slave relationship between each subordinate object and its matched dominant object, wherein the dominant object is the training robot or the trained player. S1042, the collaborative control platform determines the collaborative motion trajectory for the training robot in real time based on the relative positional relationship between subordinate and dominant objects in the preset master-slave formation and the real-time motion trajectory of the dominant object with master-slave relationship with the training robot in the global training situation map. Based on the collaborative motion trajectory, it generates collaborative operation control instructions. The collaborative motion trajectory is used to limit the real-time position of subordinate and dominant objects with master-slave relationship to maintain the master-slave formation. The collaborative operation control instructions are configured to drive the training robot to move along the collaborative motion trajectory and perform role tasks corresponding to the assigned training role in order to achieve collaborative hockey training with the trained player.

[0045] S1041 further includes: S10411, the collaborative control platform, based on the training role of each training robot and in descending order of priority of each training role, sequentially assigns each training robot as a subordinate object and matches a dominant object to the subordinate objects; wherein, matching a dominant object to a subordinate object includes: A candidate object set is determined for the subordinate object, and the initial cooperative motion trajectory of each candidate object in the candidate object set is calculated. Spatiotemporal conflict detection is performed between the initial cooperative motion trajectory and the cooperative motion trajectories of all successfully matched training robots. The target object corresponding to the initial cooperative motion trajectory that does not produce spatiotemporal conflict is selected from the candidate object set, and the target object is used as the dominant object for successful matching of the subordinate object.

[0046] In implementation, after the assignment of training roles and practice robots is completed, each practice robot will be assigned a subordinate role in descending order of priority based on its training role. A master robot will then be matched with a subordinate robot to establish a master-slave relationship between the subordinate robot and its matched master robot. The process of matching a master robot with each subordinate robot is as follows: First, the collaborative control platform queries the master-slave grouping rule base bound to the current training mode. This active grouping rule base, in the form of a data structure, explicitly defines the allowed master-slave relationships between different training roles; example rules are as follows: Rule a: If the training role of the subordinate object is an off-ball receiver THEN, its candidate objects can be [ball-carrying attacker (bot), trained player]; Rule b: If the training role of the subordinate object is a defensive interceptor THEN, its candidate object can be [the trained player (as the defensive target)]; Rule c: If the training role of the subordinate object is the ball-handling attacker THEN, its candidate object can be [the trained player (as a passing receiver)].

[0047] Based on the above rules, the collaborative control platform will select a set of candidate objects that conform to the tactical logic for the current subordinate object. The candidate objects in this set are specific training robots and / or trained players that conform to the rules.

[0048] Then the collaborative control platform will iterate through the above set of candidate objects and perform the following steps for each candidate object: The collaborative control platform first determines the relative positional relationship between subordinate and dominant objects in a predefined master-slave formation. The relative positional relationship is specifically represented by a polar coordinate point defined by a distance d (in meters) and an angle θ (in degrees, relative to the direction of movement of the dominant object). For example, a triangular attack formation includes one dominant object and two subordinate objects, and one subordinate object should be located at a 45-degree angle to the left front of the dominant object and a distance of 2 meters (the corresponding relative positional relationship can be defined as (d=2, θ=45), and the other subordinate object should be located at a 45-degree angle to the right front of the dominant object and a distance of 2 meters). Based on the relative positional relationship defined in the master-slave formation, and using the candidate object's real-time position (X_l, Y_l) and direction of movement φ_l in the global training situation map as a reference, the collaborative control platform calculates the initial desired position that the current subordinate object must move to beforehand if it follows the candidate object, according to the relative positional relationship (d, θ): (X_f_desired, Y_f_desired) = (X_l + d * cos(φ_l + θ), Y_l + d * sin(φ_l + θ)). The initial desired position satisfies the following condition: the relative positional relationship formed between the desired position and the candidate object's current position satisfies the requirements of the master-slave formation.

[0049] The movement path of the subordinate object from its current position to the initial expected position is then calculated (i.e., the initial cooperative motion trajectory, which is a straight line segment starting from the current position of the subordinate object and ending at the initial expected position). This initial cooperative motion trajectory is then compared with the cooperative motion trajectories of all training robots that have successfully matched the dominant object before the current time for a future period of time (e.g., 3 seconds) to detect spatiotemporal conflicts. Specifically, it is detected whether the minimum distance between the initial cooperative motion trajectory and the cooperative motion trajectory of any matched training robot is less than a preset safety threshold (e.g., 0.8 meters) in the future period of time. If so, the position in the initial cooperative motion trajectory where the minimum distance to other cooperative motion trajectories is less than the preset safety threshold is recorded. This position is taken as the conflict point, and the candidate object is determined to have a spatiotemporal conflict and is then eliminated.

[0050] If all candidate objects are found to be in conflict after iteration, the system will initiate a conflict resolution cost assessment. For each conflicting candidate object, it calculates an "alternation solution," which is a new cooperative motion trajectory that avoids all conflict points (e.g., modifying the original trajectory segment passing through the conflict point into an arc trajectory centered on the conflict point with a specified radius). Subsequently, the system evaluates the "resolution cost" of executing each detour trajectory. The cost calculation formula can be defined as: Resolution Cost = Length of the new detour trajectory / Robot's current remaining battery percentage. Finally, the candidate object with the lowest resolution cost is selected as the target object, and the target object is made the dominant object of the current subordinate object to achieve matching.

[0051] If, after traversal, more than one candidate object is found that does not have a spatiotemporal conflict, then all non-conflicting candidate objects are taken as target objects. In other words, a subordinate object can match more than one target object; a master-slave relationship can contain one subordinate object and multiple target objects. Master-slave conditions can be set for all target objects within the same master-slave relationship. Only when the corresponding master-slave condition is met will the corresponding target object establish a master-slave relationship with the subordinate object as the master object. This limits a subordinate object to having only one master object with a master object at any given time. These conditions are Boolean logic expressions, and the corresponding master-slave relationship is only activated when evaluated as "true" in real-time during training. This means that at any given moment, a subordinate object has exactly one master object directing its actions, but it can intelligently switch between multiple preset master objects based on the battlefield situation.

[0052] Furthermore, it should be noted that the master-slave condition aims to ensure that, for a subordinate object with multiple target objects, only one target object establishes a master-slave relationship with that subordinate object at any given time. Therefore, to satisfy this constraint, the master-slave condition setting in this application follows the principle of mutual exclusion. Specifically, this is achieved by dividing the core state variables (such as puck position, training phase) upon which each master-slave condition depends into non-overlapping decision intervals, ensuring that only one condition is true at any given time. Further, a static priority is preset for all master-slave conditions. When multiple conditions are simultaneously evaluated as true in rare cases, the system will only activate the master-slave relationship with the highest priority. For example, uniqueness is ensured by constructing a decision tree. This tree first divides the system into primary partitions based on the training phase, then into secondary partitions based on key tactical positions within the regular phase, and finally into tertiary partitions based on possession. The state variables (such as time, position, possession) upon which the partition conditions at the same level depend are logically mutually exclusive, thus ensuring that the path from the root node to any leaf node is unique. In addition, establishing the highest priority system monitoring rules can interrupt the current path and trigger a re-decision at any time. For example: First Tier Division (Based on Training Phase): If remaining time < 1 / 3 of total time, then activate the master-slave relationship with the defensive interceptor. ELSE (Regular Training Phase): Proceed to the Second Tier Division. Second Tier Division (Based on Key Tactical Positions): If the trained player enters the penalty area, then activate the master-slave relationship with the trained player. ELSE (Player not in the penalty area): Proceed to the Third Tier Division. Third Tier Division (Based on Puck Ownership): If the puck is controlled by the trained player and located in the offensive half, then activate the master-slave relationship with the trained player; if the puck is controlled by the attacking robot, then activate the master-slave relationship with that robot. Furthermore, in any phase, if the current active dominant player's battery is < 15%, then the relationship is forcibly terminated, and the system is immediately re-evaluated from the first tier.

[0053] After completing the above operations, the role swap strategy will be executed. Specifically, the collaborative control platform will evaluate whether the overall benefit of swapping the roles of any two assigned robots (the overall benefit is used to assess whether the "role swap strategy" is worthwhile, with the goal of simultaneously improving system efficiency and safety) exceeds a preset threshold (e.g., 10%). If the condition is met, the swap will be executed. The overall benefit is calculated as: total expected path shortening percentage + total safety margin increase percentage. This completes the role swap strategy. Before the role swap, the collaborative control platform calculates the expected path length for all training robots to reach their respective initial expected positions (i.e., the expected position corresponding to the current position of the dominant training robot) based on the current master-slave relationship matching results. These lengths are then summed to obtain the total path length before. Then, a role swap is simulated (e.g., the training roles assigned to training robots A and B are swapped), and the master-slave matching step and collaborative motion path calculation step (i.e., S1041 and S1042) are executed once based on the swapped training roles and their corresponding training robots. Calculate the sum of the expected path lengths of all training robots under the new configuration (i.e., the new master-slave relationship matching result) (i.e., the total path length after). Then calculate: Total expected path shortening percentage = (Total path length before - Total path length after) / Total path length before * 100%. The calculation steps for the total safety margin increase percentage are as follows: For any pair of training robots (i, j), its safety margin SM_ij is defined as the difference between the minimum predicted distance D_min_ij between their initial cooperative motion trajectories over a future period of time (e.g., 3 seconds) and the safety threshold D_safe (e.g., 0.8 meters), i.e., SM_ij = D_min_ij - D_safe. This value can be negative, indicating that a collision will occur. Calculate the safety margin of any pair of training robots before the swap (i.e., any two training robots), but only sum the positive values ​​(i.e., the safe case), ignoring negative values ​​(because a conflict has already occurred, the focus is on improvement). The total safety margin is calculated as follows: _before_ = Σ max(0, SM_ij). Similarly, under the assumption of a new configuration after role swapping (i.e., the new master-slave relationship matching result), the safety margins of all robot pairs are calculated, and the positive values ​​are summed to obtain the total safety margin _after_. To avoid division by zero, the formula can be adjusted to: percentage increase in total safety margin = (total safety margin _after_ - total safety margin _before_) / (total safety margin _before_ + ε) * 100% (where ε is a very small positive number used to prevent calculation errors when the total safety margin _before_ is zero). If this value is positive, it indicates that the overall safety buffer space of the entire robot team during movement has increased after the swap, the risk of collision has decreased, and the system safety has been improved.Finally, the overall benefit is obtained by adding the percentage reduction of the total expected path to the percentage increase of the total safety margin. If the overall benefit exceeds the preset threshold (e.g., 10%), the efficiency and safety improvement brought about by this role swap is considered to be significant, and the role swap operation is executed (i.e., the training roles of the corresponding training robots are swapped, and the master-slave matching and collaborative motion path calculation steps are re-performed).

[0054] After completing the above operations, training will officially begin (the collaborative control platform can send a start training command to the training robot, so that the training robot can display the countdown time of the training process on the digital display screen on its body, and use the LED light strip on its body to display the corresponding color to indicate the current status of the robot, such as green, to inform the trained player that the training robot is ready and training can begin).

[0055] After training begins, the collaborative control platform acquires environmental information transmitted by the training robot in real time. It also acquires environmental images detected by the fixed vision module at the rinkside, analyzes all acquired environmental images, and identifies the relative poses and motion states of the targets (hockey puck, trained player, training robot) appearing in the environmental images. Here, "targets" refers to the objects appearing in the environmental images. The specific implementation scheme for the collaborative control platform to identify the relative poses and motion states of the targets in the environmental images is the same as the method used by the environmental perception module of the training robot to output environmental information, as described above. That is, it uses a deep learning-based target detection network (such as YOLO or Faster R-CNN) to process each frame of the environmental image, identifying the hockey puck, the player (distinguished by specific colored training uniforms), and the training robot (through visual markers on its body, such as ArUco codes). For the identified targets, the platform calculates their three-dimensional position (X, Y, Z coordinates) relative to the coordinate system of the fixed vision module at the rinkside using binocular visual stereo matching or visual odometry technology combined with IMU data. The position data of the same target in a continuous frame sequence is processed by a Kalman filter to smooth noise and estimate its motion state, including instantaneous velocity vector (Vx, Vy, Vz) and motion direction.

[0056] Then, a global training situation map is obtained based on all environmental images and the relative poses and motion states of the identified targets. The global training situation map is updated and determined in real time, and it is used to reflect the position and motion relationships of all identified targets in the training field.

[0057] The specific method for determining the global training situation map is as follows: The collaborative control platform acts as a data fusion center, running an extended Kalman filter. This filter transforms the target poses and motion states reported by all sensing nodes (training robots and fixed vision modules at the rink's edge) in their respective local coordinate systems into a global coordinate system with the center of the ice rink as the origin. The extended Kalman filter then performs weighted fusion of data from different sources concerning the same target, optimally estimating the absolute position, velocity, orientation, and uncertainty covariance of each target in the global coordinate system. The resulting global training situation map is essentially a real-time updated data structure containing the aforementioned precise state information (absolute position, velocity, and orientation) of all dynamically identified targets.

[0058] Then, based on the real-time updated global training situation map and the master-slave relationship given above, a cooperative motion trajectory is determined for each subordinate object according to the relative positional relationship specified in the predefined master-slave formation. The specific method for determining the cooperative motion trajectory during the training process differs from the calculation method of the initial cooperative motion trajectory in the step of matching the subordinate object with the master object described above. The specific method for determining the cooperative motion trajectory during the training process is generated by the Model Predictive Control (MPC) algorithm. Every preset control cycle, this MPC algorithm, based on the differential kinematics model of the training robot, aims to minimize the tracking error between the subordinate object and the desired position, while maximizing motion smoothness. Its constraints include the maximum speed and maximum acceleration of the training robot, and incorporate real-time obstacle avoidance constraints from the global training situation map to ensure a safe distance from other dynamic targets in the field.

[0059] After determining the coordinated motion trajectory, coordinated operation control instructions are generated based on it. These instructions specifically include movement instructions and task instructions. Movement instructions drive the training robot to move along the corresponding coordinated motion trajectory; task instructions execute the role task corresponding to the assigned training role. For example, a task instruction might be: for a training role of an attacker with the ball, the task instruction triggers a pre-set ball-serving mechanism on the training robot to pass the ball with a specific striking force and angle towards the receiver without the ball or towards the goal; for a training role of a defender / interceptor, the task instruction controls them to perform a rapid lateral slide and extend the interception bar of the ball-serving mechanism to simulate a tackle.

[0060] The execution of the task instruction is achieved by the underlying controller of the training robot, which retrieves a pre-programmed sequence of actions from a pre-stored motion library that matches the task instruction. This process involves precise control of specialized actuators (such as servo motors controlling the ball-hitting mechanism for passing and servo motors controlling the interception stick for interception), and mapping the logical parameters in the task instruction (such as force, angle, and target) to corresponding physical control quantities of the actuators (such as PWM pulse width, motor rotation angle, and current magnitude). The underlying controller ensures that these action sequences are executed synchronously in time and space with the trajectory planned by the movement instruction, thereby completely reproducing the tactical behavior corresponding to the training role. Specifically, When the underlying controller of the training robot receives a task instruction (such as "pass" or "intercept") from the collaborative control platform, it will first parse the task instruction, identify the task type (such as action type: pass), and extract action parameters (such as target object: receiver without the ball, hitting power: 80%, hitting angle: 15 degrees horizontally upward). These parameters will be mapped from the tactical logic to the drive unit contained in the training robot.

[0061] For example, in a passing task: the underlying controller controls the drive unit to drive the servo motor or linear motor, adjusts the hitting plate or lever of the serving mechanism on the robot body to a precise angle corresponding to the hitting angle parameters, controls the electromagnet or servo motor to compress the energy storage spring of the serving mechanism, the degree of compression being proportional to the hitting force parameters, and through the environmental perception module preset on the training robot body, confirms the relative position of the passing target (the receiver without the ball or the goal) and makes fine adjustments. Finally, the drive unit sends an electrical pulse signal to the serving mechanism, triggering the electromagnetic release device, causing the energy storage spring to be released instantaneously, and hitting the puck.

[0062] For interception missions, the underlying controller calculates the differential wheel speed combination required to execute a "lateral slide" based on movement commands and sends it to the drive unit. The drive unit drives the left and right wheels to generate a speed difference, achieving rapid lateral movement (slide). Simultaneously, it controls a dedicated servo motor to drive a simulated interceptor bar to quickly extend from the side of the robot body to a predetermined position. The stability and distance of the slide are ensured by onboard IMU (Inertial Measurement Unit) and encoders. After the interception action is completed, the drive unit controls the servo motor to retract the interceptor bar.

[0063] Throughout the execution of the task instructions, the robot's underlying controller continuously monitors the status of the actuators (such as motor current, servo angle, and sensor triggering) to ensure that the action is completed as expected.

[0064] For any training role, the mission instructions also include controlling the preset LED light strip on the training robot to switch to a specific color (such as red when attacking), or displaying a specific tactical code on the digital display screen.

[0065] The collaborative control platform sends collaborative operation control commands to the training robot, which then receives and executes these commands.

[0066] In addition, during collaborative training, the training robot will monitor its own status information in real time and transmit the status information to the collaborative control platform in real time. The status information here can specifically include the remaining battery percentage, motor driver temperature, calculated core load rate, and its own real-time pose and covariance calculated by the body IMU and visual odometry.

[0067] Optionally, the multi-robot collaborative method for hockey training also includes the following steps: S201, during collaborative training, each training robot continuously monitors its own status information and sends the status information to the collaborative control platform; S202, the collaborative control platform determines that a spatiotemporal conflict is about to occur when it detects that the collaborative motion trajectories of two first training robots intersect or the minimum distance between them is less than a preset safety distance within a future time. It then triggers and executes a trajectory correction strategy to adjust the collaborative motion trajectory of the corresponding training robots before the spatiotemporal conflict occurs, thereby resolving the conflict. Here, "first training robot" refers to any training robot. S203, the collaborative control platform determines that a runtime capability conflict has occurred when it detects that the second training robot cannot meet the robot capability requirements of the assigned training role due to a state change, and triggers the execution of a distributed negotiation resolution strategy for the second training robot; whereby the second training robot refers to any training robot. The distributed negotiation solution strategy includes: selecting the training robot that will take on the role and task of the second training robot as the winning training robot, assigning the role and task of the second training robot to the winning training robot, and updating the cooperative motion control instructions of the bidding training robot so that the winning training robot can execute the updated cooperative motion control instructions.

[0068] The "matching the dominant object to the subordinate object" in S10411 also includes the following steps: When multiple target objects are selected from the candidate object set, a master-slave condition is preset for each target object. The master-slave condition is configured to ensure that, at any given time, the master-slave condition of one and only one target object is satisfied, based on mutually exclusive state variables or preset priorities.

[0069] Multi-robot collaborative methods for ice hockey training also include: S301 During the training process, for a training robot with multiple target objects, the master-slave conditions of all target objects are evaluated in real time. The target object that meets the master-slave conditions is used as the master object to establish a master-slave relationship with the corresponding training robot as the slave object, so as to realize the real-time update of the master-slave relationship of the training robot with multiple target objects.

[0070] The distributed negotiation solution strategy, which involves "assigning the role and task of the second training robot to the winning training robot and updating the cooperative motion control instructions of the bidding training robot," specifically includes: The second training robot's role and tasks are assigned to the winning training robot. The winning training robot enters a multi-role fusion state. The winning training robot must maintain its own original role and tasks as well as the role and tasks of the second training robot. During training, the master-slave conditions corresponding to all roles and tasks maintained by the training robot are continuously monitored in parallel. All roles and tasks corresponding to satisfied master-slave conditions are generated and included in the effective task set. Based on a preset global task priority list, the highest-priority role and task is selected from the current effective task set as the primary execution task for the current control cycle. The corresponding role action instructions are retrieved from the preset behavior library. The target object corresponding to the master-slave conditions of the primary execution task is designated as the master object of the training robot, establishing a master-slave relationship. The cooperative motion control instructions of the training robot are updated based on this master-slave relationship until training ends.

[0071] In implementation, the collaborative control platform determines the presence of a first training robot based on the real-time status information of the training robots. If present, it triggers a trajectory correction strategy. The method for determining the first training robot is as follows: the collaborative control platform predicts the collaborative motion trajectory of all training robots within the next two seconds based on the latest global training situation map. Specifically, the prediction method uses the most accurate state vector of each dynamic target (robot, player, hockey puck) provided by the global training situation map at the current moment, typically including position (X, Y), velocity (Vx, Vy), orientation (φ), and angular velocity (ω) as basic data. For the training player, the trajectory is predicted using a uniform motion model based on the current motion state, according to the player's current position, velocity, and orientation.

[0072] For each training robot, this application needs to clarify that all the master-slave relationships established above are actually one or more hierarchical trees. The root node of the tree is the trained player. Training robots that directly dominate the player constitute first-level subordinate objects (child nodes). Training robots that directly dominate the first-level subordinate objects constitute second-level subordinate objects (grandchild nodes), and so on. Furthermore, this application assumes that all training robots are assigned a dominant object and are located at a certain level in this hierarchical tree. This means that even for robots playing the role of opponents, their master-slave relationships are explicitly defined. For example, their dominant object may be set as "the trained player," and their master-slave formation is used to describe a defensive or confrontational relative position (such as "maintaining a position 3 meters in front of the player to intercept").

[0073] Therefore, after determining the trajectory of the trained player within the next 2 seconds, the collaborative control platform will predict the trajectory of each training robot in a hierarchical order from high to low (e.g., first level, then second level). Specifically, for any training robot (subordinate object S) in the hierarchy tree, the complete predicted trajectory Trajectory_D(t) of the parent node D in the hierarchy tree is first determined. Then, based on the predefined master-slave formation (d, θ), and the position (X_D(t), Y_D(t)) and direction of motion φ_D(t) of the dominant object D at each time point t on the predicted trajectory, the expected position of the subordinate object S at each time point t is calculated: (X_S_desired(t), Y_S_desired(t))=(X_D(t)+d*cos(φ_D(t)+θ), Y_D(t) + d * sin(φ_D(t) + θ)). Starting from the current position of the subordinate object S, and using the calculated expected position time series as the tracking target, the system employs the kinematic model of the training robot (such as a differential drive model, which describes the geometric relationship between the linear velocity v and angular velocity ω of the robot's chassis center point and the rotational speeds nleft and nright of the two drive wheels, and how these velocities are integrated to obtain the robot's pose (position x, y and orientation ϕ)) and control strategy (such as PID tracking or model predictive control) for forward simulation. Specifically, during trajectory prediction, the current position and orientation of the subordinate object are used as the initial state, and the expected linear velocity and angular velocity calculated based on the expected position time series are used as control inputs. Forward numerical integration is performed using the differential drive model to deduce its future pose sequence, which constitutes the predicted cooperative motion trajectory Trajectory_S(t) of the subordinate object S. This process is repeated until all training robots in the hierarchy have been processed. Finally, the system obtains a complete set of predicted trajectories for the next 2 seconds, containing the trained player and all training robots, all interconnected.

[0074] Next, the minimum relative distance between the predicted cooperative motion trajectories of any two training robots is calculated. If the minimum distance is less than the safety threshold (0.8 meters), the relevant robot is marked as the "first training robot," indicating an impending spatiotemporal conflict. The corresponding trajectory correction strategy is to initiate local path replanning for one or both parties involved in the conflict. This is typically achieved by finding the conflict point (i.e., the location on the original cooperative motion trajectory where the minimum distance is less than the preset safety distance) and superimposing a temporary, mutually repelling correction (e.g., modifying a trajectory segment of a specified length at the conflict point on the cooperative motion trajectory into an arc trajectory segment centered on the conflict point with a specified radius), causing it to deviate slightly from the original trajectory to avoid the conflict. The correction strategy prioritizes the option with the least impact on the tactical formation and guides the robots back to their original formation after the conflict is resolved.

[0075] The collaborative control platform also determines the presence of a second training robot based on the real-time status information of the first training robot. If a second training robot is found, a distributed negotiation strategy is triggered. The method for determining a second training robot is as follows: the collaborative control platform continuously checks the status data of each training robot. When it detects that a robot's remaining battery power is below the threshold required to perform its current task (e.g., the task requires >20% battery power, but the robot only has 15% battery power remaining), or when a fault code is reported (e.g., motor overheating), that robot is identified as a "second training robot." The corresponding distributed negotiation strategy is as follows: The collaborative control platform sends task announcements to all training robots except the second training robot. These announcements include not only the role task to be reassigned (i.e., the role task of the second training robot's current training role), but also a detailed description of the task's complete context, including: the role's preset activity area, master-slave formation requirements, other target objects requiring interaction (i.e., all target objects corresponding to the second training robot as a subordinate object, and all subordinate objects corresponding to the second training robot as the master object), the minimum capability standards required to successfully execute the task (e.g., minimum required battery level > 30%, maximum required movement speed > 5m / s), and key information extracted from the latest global training situation map (e.g., the current position of the hockey puck, the current position of the trained player, etc.). Upon receiving the task announcement, the training robot first performs a self-qualification check, that is, checks whether its current state meets the capability requirements specified in the task announcement. If it does not meet the requirements, it automatically abandons the bid; if it does, it is considered qualified. Qualified training robots calculate a bid value to quantify their "cost-effectiveness" or "suitability" for performing the task. This is a core intelligent decision point. The bid value Bid is calculated based on the following formula: Bid = w1 * (remaining battery power) + w2 * (1 / Euclidean distance between the robot's position and the task area) + w3 * (professional matching degree between the robot and the required role in the task); where w1, w2, and w3 are preset weight coefficients. Professional matching degree is a value between 0 and 1, determined by historical data or configuration files (such as configuration files storing the main role tasks performed by different training roles, and the competency values ​​of the main role tasks performed by the training roles, which are the professional matching degrees mentioned above). It represents the degree to which the training robot model or historical performance is proficient in the role task of the second training robot (for example, a robot with agile turning is more suitable for the "defense interceptor" role). Each training robot encapsulates the calculated Bid value and its robot ID into a bidding message and sends it back to the task manager. Correspondingly, after issuing the task announcement information, the collaborative control platform waits for a preset bidding window time (e.g., 500 milliseconds), then collects all returned bidding messages. Subsequently, the collaborative control platform selects the bidder with the highest Bid value as the successful bidder, and then sends a winning notification to the training robot corresponding to the successful bidder, informing it that it has been assigned a new role task. Based on the new role allocation result, the role task of the training robot is updated (i.e., the original role task + the newly assigned role task). Based on the original master-slave relationship of the successful bidder and the master-slave relationship corresponding to the second training robot, the correspondence between the master-slave relationship and the role task is updated. For example, the original master-slave relationship of the successful bidder corresponds to its original training role, and the role task inherited by the successful bidder from the second training robot corresponds to the master-slave relationship of the second training robot.

[0076] However, when a successful bidder needs to inherit the role tasks of a faulty robot, it actually enters a "multi-role fusion" state. At this time, its original role tasks (denoted as Role A) and the newly inherited role tasks (denoted as Role B) may correspond to different or even conflicting sets of master-slave relationships and master-slave conditions. To ensure that the system's decisions are clear and consistent at this time, an arbitration mechanism must be introduced. This application proposes a master-slave condition arbitration mechanism based on task priority. Specifically, the collaborative control platform maintains a globally identical task priority list, which is determined when the training mode is loaded. The priority is set based on tactical importance, for example: ball-handling attacker (priority: 5) > defensive interceptor (priority: 4) > off-ball receiver (priority: 3). After a successful bidder merges multiple role tasks, the collaborative control platform does not select a primary task and then remain fixed. Instead, the collaborative control platform will enter the following high-frequency cyclical role process: Step 1 (Parallel Condition Evaluation): During the training process, at specified control cycles (e.g., every 100 milliseconds), the collaborative control platform evaluates in parallel and in real time the master-slave conditions bound to all roles and tasks of the winning bidder (role A, role B, etc.).

[0077] Step Two: Constructing the Effective Task Set. All role tasks corresponding to master-slave conditions with evaluation results of "true" are included in the current "effective task set." This means that a role task will only participate in the competition if its master-slave condition is met. Step Three: Dynamically Arbitrating the Primary Task. The collaborative control platform compares the priorities of all tasks in the effective task set and always selects the task with the highest static priority as the "primary execution task" for the current cycle.

[0078] Once the primary task is determined, the winning bidder immediately switches its behavior mode to the state corresponding to that primary task. This means that the winning bidder's training robot calls the role action instructions corresponding to the primary task from its built-in preset behavior library (e.g., if the primary task is an attacker with the ball, then its tactical behavior is centered on passing and shooting). The winning bidder's training robot only focuses on the master-slave relationship corresponding to the training role bound to the primary role task (i.e., the training role that takes the primary task as its main role task), and temporarily ignores the master-slave relationship corresponding to other role tasks.

[0079] For example, the winning robot acts as both the puck-carrying attacker (priority 5) and the defensive interceptor (priority 4). In scenario A (our team controls the puck): the attacker's master-slave condition (e.g., the puck is under our control) is true (i.e., the corresponding master-slave condition is met), and it enters the valid task set. The interceptor's condition (e.g., the opponent controls the puck) is false (i.e., the corresponding master-slave condition is not met), and it does not enter the task set. In this case, the primary task is the attacker, and the robot performs offensive actions. In scenario B (the opponent steals the puck): the attacker's master-slave condition becomes false (i.e., the corresponding master-slave condition is not met), and it exits the valid task set. Simultaneously, the interceptor's master-slave condition becomes true (i.e., the corresponding master-slave condition is met), and it enters the valid task set. In this case, the valid task set only includes the interceptor, and the collaborative control platform immediately switches, the primary task becomes the interceptor, and the robot performs defensive actions.

[0080] After determining the primary task and its corresponding master-slave relationship, if there are multiple target objects within the corresponding master-slave relationship, the successful bidder needs to identify the single dominant object from among these target objects. The decision-making process is as follows: Case A (Single condition satisfied): If only one master-slave condition is satisfied among all the master-slave conditions corresponding to all target objects, then the target object corresponding to that master-slave condition is taken as the current sole master object, and a master-slave relationship is established with that successful bidder. Scenario B (Multiple Conditions Met): If multiple master-slave conditions are met simultaneously for all target objects, the system will initiate secondary arbitration. The secondary arbitration strategy can be: Strategy 1 (Fixed Priority): Based on the priority of the training role corresponding to the target object, select the target object with the highest priority (where the priority of the trained player is higher than the priority of all training roles) as the current sole dominant object. For example, trained player > training robot with the ball-handling attacking role. Or Strategy 2 (Dynamic Evaluation): Calculate the "utility value" of each target object that meets the master-slave conditions, and select the target object with the highest utility as the sole dominant object. Utility value = 1 / (Euclidean distance between the current position of the winner and the current position of the target object) + cos(angle between the target object's current direction of movement and the tactical reference direction). cos(angle between the target object's current direction of movement and the tactical reference direction) refers to the cosine of the target object's velocity direction vector and the tactical reference direction vector; the tactical reference direction is a preset simple vector. In offensive mode, it can be a direction vector "from the center line of the field towards the center of the opponent's goal." In defensive mode, it can be the opposite.

[0081] This application also discloses a multi-robot collaborative system for ice hockey training. It includes a collaborative control platform and several training robots communicatively connected to the collaborative control platform. The collaborative control platform is used to dynamically assign training roles to each training robot according to a preset training mode; the training roles include at least an offensive player with the ball and a defensive interceptor. All training robots are used to perceive environmental information in real time and send the environmental information to the collaborative control platform; the environmental information includes at least the relative pose and motion state of the identified targets appearing within the perception range of the training robot; the identified targets include at least ice hockey, the trained player, and the training robot; The collaborative control platform is used to integrate all environmental information to generate a global training situation map that reflects the position and motion relationships of all identified targets within the training field; The collaborative control platform generates collaborative motion control commands for each training robot based on the global training situation map and the assigned training role, so that the training robot can acquire and execute collaborative operation control commands and achieve collaborative hockey training.

[0082] This application also discloses a multi-robot collaborative device for ice hockey training. The multi-robot collaborative device for ice hockey training includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed as described above for the multi-robot collaborative method for ice hockey training.

[0083] This application also discloses a computer-readable storage medium that stores a computer program that can be loaded by a processor and executed as described above for a multi-robot collaborative method for hockey training. The computer-readable storage medium includes, for example, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0084] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0085] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit the scope of protection of the application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

Claims

1. A multi-robot collaborative method for ice hockey training, characterized in that, The execution entity of the multi-robot collaborative method for ice hockey training is a multi-robot collaborative system for ice hockey training. The multi-robot collaborative system includes a collaborative control platform and several training robots communicatively connected to the collaborative control platform. The multi-robot collaborative method for ice hockey training includes the following steps: The collaborative control platform dynamically assigns training roles to each training robot according to a preset training mode; wherein, the training roles include at least an offensive player with the ball and a defensive interceptor. All training robots perceive environmental information in real time and send the environmental information to the collaborative control platform; wherein, the environmental information includes at least the relative pose and motion state of the identified targets appearing within the perception range of the training robots; the identified targets include at least ice hockey, the trained player, and the training robots; The collaborative control platform integrates all the environmental information to generate a global training situation map that reflects the position and motion relationship of all identified targets in the training field. The collaborative control platform generates collaborative motion control commands for each training robot based on the global training situation map and the assigned training roles, so that the training robot can acquire and execute the collaborative operation control commands and realize collaborative hockey training.

2. The multi-robot collaborative method for ice hockey training according to claim 1, characterized in that, The collaborative control platform generates collaborative motion control commands for each training robot based on the global training situation map and the assigned training roles, including the following steps: The collaborative control platform assigns each training robot as a subordinate object based on its training role, and matches each subordinate object with a master object, establishing a master-slave relationship between each subordinate object and its matched master object. The master object is either the training robot or the player being trained. The collaborative control platform determines the collaborative motion trajectory for the training robot in real time based on the relative positional relationship between subordinate and dominant objects in the preset master-slave formation, and the real-time movement trajectory of the dominant object with a master-slave relationship with the training robot in the global training situation map. Based on the collaborative motion trajectory, it generates collaborative operation control instructions. The collaborative motion trajectory is used to limit the real-time position of subordinate and dominant objects with a master-slave relationship to maintain the master-slave formation. The collaborative operation control instructions are configured to drive the training robot to move along the collaborative motion trajectory and perform role tasks corresponding to the assigned training role, so as to realize collaborative hockey training with the trained player.

3. The multi-robot collaborative method for ice hockey training according to claim 2, characterized in that, The collaborative control platform dynamically assigns training roles to each training robot according to a preset training mode, including: The collaborative control platform determines all training roles included in a preset training mode; and assigns a training robot to each training role based on a preset optimization strategy; wherein the optimization strategy includes at least one of the following: assigning according to a preset priority for the training role and / or training robot; evaluating the overall benefit of swapping the training roles of any two training robots, and if the overall benefit increases by more than a preset threshold, then swapping the training roles of the corresponding training robots; and assigning a training robot whose actual state can meet the preset capability requirements according to the preset capability requirements required by the task corresponding to the training role. The collaborative control platform assigns each training robot as a subordinate object based on its training role, and matches each subordinate object with a dominant object, including: The collaborative control platform, based on the training role of each training robot, sequentially assigns each training robot as a subordinate object according to the priority of each training role from high to low, and matches a dominant object to the subordinate object. The process of matching a subordinate object with a dominant object includes: A candidate object set is determined for the subordinate object, and the initial cooperative motion trajectory of each candidate object in the candidate object set is calculated. The initial cooperative motion trajectory is compared with the cooperative motion trajectories of all successfully matched training robots to perform spatiotemporal conflict detection. The target object corresponding to the initial cooperative motion trajectory does not produce spatiotemporal conflict is selected from the candidate object set, and the target object is used as the dominant object for successful matching of the subordinate object.

4. A multi-robot collaborative method for ice hockey training according to claim 3, characterized in that, The method further includes: During collaborative training, each training robot continuously monitors its own status information and sends the status information to the collaborative control platform; The collaborative control platform is used to determine that a spatiotemporal conflict is about to occur when it detects that the collaborative motion trajectories of two first training robots intersect or the minimum distance is less than a preset safety distance in the future. It then triggers and executes a trajectory correction strategy to adjust the collaborative motion trajectory of the corresponding training robots before the spatiotemporal conflict occurs, thereby resolving the spatiotemporal conflict. Herein, the first training robot refers to any training robot. The collaborative control platform is used to determine that a runtime capability conflict has occurred when it detects that the second training robot cannot meet the robot capability requirements of the assigned training role due to a state change, and triggers the execution of a distributed negotiation resolution strategy for the second training robot; wherein, the second training robot refers to any training robot. The distributed negotiation solution strategy includes: selecting the training robot that will undertake the role task of the second training robot as the winning training robot, assigning the role task of the second training robot to the winning training robot, and updating the cooperative motion control instructions of the bidding training robot so that the winning training robot executes the updated cooperative motion control instructions.

5. A multi-robot collaborative method for ice hockey training according to claim 3, characterized in that, The process of matching a subordinate object with a dominant object also includes: When multiple target objects are selected from the candidate object set, a master-slave condition is preset for each target object. The master-slave condition is configured to ensure that, at any given time, the master-slave condition of one and only one target object is satisfied, based on mutually exclusive state variables or preset priorities. The method further includes: During the training process, for a training robot with multiple target objects, the master-slave conditions of all target objects are evaluated in real time. The target object that meets the master-slave conditions is used as the master object to establish a master-slave relationship with the corresponding training robot as the subordinate object, so as to realize the real-time update of the master-slave relationship of the training robot with multiple target objects.

6. A multi-robot collaborative method for ice hockey training according to claim 5, characterized in that, The step of assigning the role and task of the second training robot to the winning training robot and updating the cooperative motion control instructions of the bidding training robot includes: The role and task of the second training robot are assigned to the winning training robot. The winning training robot enters a multi-role fusion state. The winning training robot needs to maintain its own original role and task as well as the role and task of the second training robot at the same time. During the training process, the master-slave conditions corresponding to all roles and tasks maintained by the winning training robot are continuously monitored in parallel. Based on the evaluation results of the master-slave conditions, the primary task to be executed and the master-slave relationship corresponding to the primary task to be executed are determined from all roles and tasks maintained by the winning training robot. The cooperative motion control commands of the winning training robot are updated based on the master-slave relationship until the training ends.

7. A multi-robot collaborative method for ice hockey training according to claim 6, characterized in that, Based on the evaluation results of the master-slave conditions, determining the primary task and the corresponding master-slave relationship from all the role tasks maintained by the winning training robot includes: Generate and include the role tasks corresponding to all satisfied master-slave conditions into a valid task set; Based on the preset global task priority list, the highest priority role task is selected from the current set of valid tasks and determined as the primary task to be executed in the current control cycle. Recall the character action instructions corresponding to the primary task from the preset behavior library; The target object corresponding to the master-slave condition of the primary execution task is taken as the master object of the winning training robot, and a master-slave relationship is established.

8. A multi-robot collaborative system for ice hockey training, characterized in that, A multi-robot collaborative system for ice hockey training includes a collaborative control platform and several training robots that are communicatively connected to the collaborative control platform. The collaborative control platform is used to dynamically assign training roles to each training robot according to a preset training mode; wherein, the training roles include at least an offensive player with the ball and a defensive interceptor. All training robots are used to perceive environmental information in real time and send the environmental information to the collaborative control platform; wherein, the environmental information includes at least the relative pose and motion state of the identified targets appearing within the perception range of the training robots; the identified targets include at least ice hockey, the trained player, and the training robots; The collaborative control platform is used to integrate all the environmental information to generate a global training situation map that reflects the position and motion relationship of all identified targets in the training field. The collaborative control platform is used to generate collaborative motion control instructions for each training robot based on the global training situation map and the assigned training role, so that the training robot can obtain and execute the collaborative operation control instructions and realize collaborative hockey training.

9. A multi-robot collaborative device for ice hockey training, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images