Reinforcement learning-based game physics collision precision detection and performance optimization system

Abstract

The application discloses a game physics collision precision detection and performance optimization system based on reinforcement learning, which comprises a physical interaction semantic decoupling module, an inter-frame continuity memory module, a collision opportunity window planning module, a detection evidence adaptive delivery module, a differentiated detection kernel scheduling module, a physical consistency review module and a steady-state performance shaping module.

Description

Technical Field

[0001] This invention relates to the field of game physics collision detection and intelligent optimization technology, and in particular to a game physics collision accuracy detection and performance optimization system based on reinforcement learning. Background Technology

[0002] Existing game physics collision detection solutions typically rely on bounding box determination, continuous collision detection, trajectory interpolation analysis, or hierarchical detection kernel scheduling to identify the contact relationships between game objects. They also combine fixed rules to control the detection frequency, detection accuracy, and computational resources to meet the real-time collision processing requirements during game operation.

[0003] In scenarios with significant changes in the number of objects, complex interaction relationships, or high-speed motion, existing technologies mostly rely on single-frame states or local rules to perform detection. They lack the joint utilization of candidate collision evolution states, missed detection correction information, penetration compensation information, and collision propagation effects in continuous game frames. It is difficult to dynamically plan the spatiotemporal window of candidate collisions, and it is also difficult to verify the consistency of collision detection results based on physical response results. As a result, the accuracy and stability of the adjustment of detection evidence delivery and detection kernel call plan in continuous game frames are limited.

[0004] Therefore, how to provide a system for accurate collision detection and performance optimization in game physics based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a game physics collision accuracy detection and performance optimization system based on reinforcement learning. This invention comprehensively utilizes reinforcement learning decision-making, cross-frame temporal memory, collision opportunity window planning, adaptive delivery of detection evidence, differentiated detection kernel scheduling, and physical consistency verification methods. It describes in detail the dynamic identification of candidate collisions in game scenes, detection resource allocation, collision determination, and continuous frame smoothing optimization process. It has the advantages of high collision detection accuracy, low false negative rate, good operational stability, and high resource utilization efficiency.

[0006] The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to embodiments of the present invention includes:

[0007] The physical interaction semantic decoupling module is used to obtain the state data of each object in the game scene, perform semantic cluster division, and obtain collision semantic clusters;

[0008] The inter-frame continuity memory module is used to record information in consecutive game frames to form the temporal memory results of candidate collisions;

[0009] The collision opportunity window planning module is used to generate collision prior information based on collision semantic clusters and temporal memory results, and obtain a set of high-value collision opportunity windows through an improved R2D2 decision unit.

[0010] The adaptive evidence delivery module is used to assign differentiated detection evidence strengths to candidate collisions within a set of high-value collision opportunity windows, thereby obtaining evidence delivery results.

[0011] The differential detection kernel scheduling module is used to perform collision determination and output collision detection results based on the collision semantic cluster, the set of high-value collision opportunity windows and the evidence delivery results.

[0012] The physical consistency verification module is used to verify the consistency of the collision detection results based on the physical response results corresponding to the collision detection results, and generate compensation and correction results.

[0013] The strategy feedback update module is used to update the decision parameters and candidate collision memory state of the improved R2D2 decision unit by compensating for correction results, collision detection results, and operational performance results.

[0014] The steady-state performance shaping module is used to smoothly shape the evidence delivery strength and detection kernel call plan in the target game frame.

[0015] Optionally, the physical interaction semantic decoupling module specifically includes:

[0016] Obtain the state data of each object in the game scene within the current game frame and the preset historical game frames;

[0017] Calculate the motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance of each object based on the state data;

[0018] The physical interaction semantic vectors of each object are constructed based on motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance.

[0019] Semantic clustering is performed based on the physical interaction semantic vectors of each object to obtain the corresponding collision semantic clusters.

[0020] Optionally, the inter-frame continuity memory module specifically includes:

[0021] Obtain object state data, physical interaction semantic vectors, and collision semantic clusters from consecutive game frames, and determine candidate collision object pairs;

[0022] Construct candidate collision evolution states based on the position and velocity changes of candidate collision objects in consecutive game frames;

[0023] Record historical false detection correction information based on contact relationship records and position correction records in consecutive game frames;

[0024] Record historical missed detection correction information based on positional relationship records, contact relationship records, and penetration relationship records in consecutive game frames;

[0025] Record penetration compensation information based on penetration relationship records and position correction records in consecutive game frames;

[0026] The temporal memory results of candidate collisions are constructed based on the candidate collision evolution state, historical false detection correction information, historical missed detection correction information, and penetration compensation information.

[0027] Optionally, the collision opportunity window planning module specifically includes:

[0028] Obtain collision semantic clusters, object historical trajectory data, scene topology, object interaction relationships and temporal memory results from the current game frame and consecutive historical game frames, and construct candidate collision spatiotemporal fragments;

[0029] Collision prior information is generated based on the object's historical trajectory data, scene topology, object interaction relationships, and time sequence memory results.

[0030] Calculate the collision value, missed detection risk, and propagation impact of candidate collision spatiotemporal segments based on prior collision information;

[0031] Candidate collision spatiotemporal windows are generated based on collision value, missed detection risk, and propagation impact. The candidate collision spatiotemporal windows are then filtered through the opportunity window screening gate in the improved R2D2 decision unit to obtain the filtered candidate collision spatiotemporal windows.

[0032] The candidate collision spatiotemporal windows after screening are state-encoded through the collision evidence memory pathway and the physical consistency memory pathway, and the replay priority is determined by the multi-factor priority replay unit to form a window replay sequence.

[0033] By using the dual-output decision head in the improved R2D2 decision unit to dynamically plan the window replay sequence, the collision detection benefit value, performance cost value, and comprehensive window decision value are obtained.

[0034] The candidates are sorted according to their comprehensive window decision values ​​after screening to construct a set of high-value collision opportunity windows.

[0035] Optionally, the step of performing state encoding processing on the selected candidate collision spatiotemporal windows through the collision evidence memory pathway and the physical consistency memory pathway, and determining the replay priority through the multi-factor priority replay unit to form the window replay sequence specifically includes:

[0036] The candidate collision evolution state, historical missed detection correction information and penetration compensation information are obtained, and the collision evidence is encoded through the collision evidence memory path in the improved R2D2 decision unit to obtain the collision evidence encoding result.

[0037] The physical consistency violation degree and position correction association state corresponding to the candidate collision spatiotemporal window after filtering are obtained, and physical consistency encoding is performed through the physical consistency memory path to obtain the physical consistency encoding result;

[0038] The collision evidence coding result and the physical consistency coding result are fused to obtain the state coding result;

[0039] The replay priority of each candidate collision spatiotemporal window after screening is determined by the multi-factor priority replay unit.

[0040] The playback priorities are sorted, and a window playback sequence is formed according to the sorting results.

[0041] Optionally, the adaptive evidence delivery module specifically includes:

[0042] Obtain the set of high-value collision opportunity windows, time-series memory results, and detection load information, and determine the target of detection evidence delivery for candidate collisions within the window;

[0043] Calculate the required detection evidence value for candidate collisions based on the high-value collision opportunity window, temporal memory results, and detection load information.

[0044] The improved R2D2 decision unit outputs collision detection benefit value, performance cost value, detection evidence requirement value, state coding result and detection evidence placement priority, and calculates the basic placement value of trajectory interpolation depth, basic placement value of continuous detection length, basic placement value of boundary sampling density, basic placement value of historical frame call range and basic placement value of high-precision detection kernel trigger times respectively.

[0045] The load constraint adjustment of the basic deployment value is performed based on the detection load information to obtain the adjusted deployment value corresponding to the trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame call range, and high-precision detection kernel trigger count.

[0046] The strength of detection evidence corresponding to candidate collisions is constructed based on the adjusted delivery values, and the evidence delivery results are generated based on the strength of detection evidence.

[0047] Optionally, the differential detection kernel scheduling module specifically includes:

[0048] Obtain the collision semantic cluster to which the object belongs, the set of high-value collision opportunity windows, the evidence delivery results, and the strength of the detected evidence, and determine the initial candidate collisions;

[0049] Based on the object interaction relationships, force propagation characteristics, and event triggering importance corresponding to the initial candidate collisions, an extended analysis is performed on the associated contact objects, force transmission objects, and event triggering objects to obtain the set of associated contact objects, the set of force transmission objects, and the set of event triggering objects.

[0050] The impact of collision propagation is calculated based on the set of related contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the results of evidence deployment.

[0051] Calculate the kernel matching value corresponding to each collision detection kernel based on the collision propagation impact results and the strength of detection evidence.

[0052] The collision detection kernels that match the candidate collisions are determined based on the kernel matching values ​​corresponding to each collision detection kernel.

[0053] The collision detection kernel is invoked to perform collision determination on the initial candidate collisions, the set of associated contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the strength of detection evidence, and the collision detection results are obtained.

[0054] Optionally, the physical consistency verification module specifically includes:

[0055] Obtain the collision detection results and the corresponding physical response results;

[0056] Calculate the response velocity coefficient, response displacement coefficient, response direction coefficient, and response duration coefficient based on the physical response results;

[0057] The physical consistency check value is calculated based on the collision detection results, response speed coefficient, response displacement coefficient, response direction coefficient, and response duration coefficient.

[0058] Calculate the average physical consistency check value based on the physical consistency check values ​​corresponding to all initial candidate collisions in the current game frame;

[0059] Compensation and correction results are generated based on the physical consistency check value and the average physical consistency check value.

[0060] Optionally, the policy feedback update module specifically includes:

[0061] Obtain the performance results, which include at least the detection time, detection load information, and resource usage.

[0062] The candidate collision memory state is updated based on the compensation correction result, collision detection result, and time sequence memory result to obtain the updated candidate collision memory state.

[0063] Calculate the collision detection benefit feedback value and performance cost feedback value based on the compensation correction results, collision detection results, collision propagation impact results and operational performance results;

[0064] The timing difference error is calculated based on the collision detection benefit feedback value, the performance cost feedback value, and the collision detection benefit value and performance cost value output by the dual-output decision head.

[0065] The updated empirical sequence replay priority is obtained by updating the empirical sequence replay priority based on the comprehensive timing difference error, historical missed detection correction information, collision propagation level, physical consistency violation degree and detection load fluctuation value.

[0066] The decision parameters of the improved R2D2 decision unit are updated based on the comprehensive time-series difference error and the empirical sequence playback priority, resulting in the updated decision parameters.

[0067] The updated candidate collision memory state is fed back to the inter-frame continuous memory module, and the updated experience sequence replay priority and updated decision parameters are fed back to the collision opportunity window planning module and the adaptive delivery module of detection evidence.

[0068] Optionally, the steady-state performance shaping module specifically includes:

[0069] Obtain the evidence delivery results, collision detection results, collision propagation impact results, compensation and correction results, and running performance results for each game frame within the statistical game frame interval, and obtain the detection evidence strength and high-precision detection kernel trigger number adjustment delivery value in the target game frame;

[0070] The execution pressure value is calculated based on the evidence delivery results and running performance results corresponding to each game frame within the statistical game frame interval, and the detection maintenance value is calculated based on the collision detection results, collision propagation impact results, and compensation correction results.

[0071] The steady-state performance shaping factor is calculated based on the execution pressure value and detection maintenance value in the statistical game frame interval;

[0072] The strength of detection evidence in the target game frame is smoothed and shaped according to the steady-state performance shaping coefficient to obtain the smoothed detection evidence strength.

[0073] The detection kernel call plan value of the target game frame is calculated based on the strength of detection evidence after smoothing and the impact of collision propagation in the target game frame;

[0074] The detection kernel call plan in the target game frame is determined based on the detection kernel call plan value, and the smoothed and shaped detection evidence strength and detection kernel call plan are output to the detection evidence adaptive delivery module and the differential detection kernel scheduling module.

[0075] The beneficial effects of this invention are:

[0076] This invention introduces physical interaction semantic decoupling, inter-frame continuity memory, and collision opportunity window planning mechanisms. This allows the candidate collision identification process to move beyond local information within the current game frame and incorporate object motion relationships, collision evolution states, and historical correction information across consecutive game frames. Based on this technical solution, the determination of the candidate collision spatiotemporal window more closely reflects the actual object interaction process, thereby improving the targeting of collision detection in complex scenarios.

[0077] This invention further utilizes an adaptive evidence delivery module to differentiate the allocation of trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame recall range, and high-precision detection kernel trigger count. Combined with a differentiated detection kernel scheduling module, it performs extended analysis on associated contact objects, force-transmitting objects, and event-triggered objects. Based on this technical solution, detection resources can be centrally allocated to candidate collisions with high collision value, high risk of missed detection, or strong propagation impact. This ensures the integrity of collision determination while reducing unnecessary detection overhead and improving the matching degree between collision detection and resource allocation.

[0078] This invention also employs a physical consistency verification module to perform consistency verification based on the physical response results corresponding to the collision detection results, and feeds back the compensation and correction results to the physical consistency memory path and multi-factor priority playback unit in the improved R2D2 decision unit. Based on this technical solution, a closed-loop correction relationship is formed between the collision detection results and the physical response process. This allows the collision detection process to not only focus on whether contact has occurred, but also to correct the detection results by incorporating velocity changes, displacement corrections, and response duration, thereby improving the consistency between the collision detection results and the actual physical process.

[0079] Furthermore, this invention utilizes a strategy feedback update module and a steady-state performance shaping module to jointly apply compensation correction results, collision detection results, runtime performance results, evidence deployment results, and collision propagation impact results to the smooth shaping of the detection evidence strength and detection kernel call plan in the target game frame. Based on this technical solution, fluctuations in the detection load across multiple consecutive game frames can be controlled, and the adjustment of detection strength between consecutive game frames is more stable, thereby helping to maintain a stable balance between collision detection accuracy and game runtime performance. Attached Figure Description

[0080] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0081] Figure 1 This is a schematic diagram of the functional module structure of the game physics collision accuracy detection and performance optimization system based on reinforcement learning proposed in this invention.

[0082] Figure 2 This is a flowchart of the inter-frame continuity memory module in the reinforcement learning-based game physics collision accuracy detection and performance optimization system proposed in this invention.

[0083] Figure 3 This is a schematic diagram of the improved R2D2 decision unit in the reinforcement learning-based game physics collision accuracy detection and performance optimization system proposed in this invention. Detailed Implementation

[0084] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0085] refer to Figures 1-3 A game physics collision accuracy detection and performance optimization system based on reinforcement learning, including:

[0086] The physical interaction semantic decoupling module is used to obtain the state data of each object in the game scene, perform semantic cluster division, and obtain collision semantic clusters;

[0087] The inter-frame continuity memory module is used to record information in consecutive game frames to form the temporal memory results of candidate collisions;

[0088] The collision opportunity window planning module is used to generate collision prior information based on collision semantic clusters and temporal memory results, and obtain a set of high-value collision opportunity windows through an improved R2D2 decision unit.

[0089] The adaptive evidence delivery module is used to assign differentiated detection evidence strengths to candidate collisions within a set of high-value collision opportunity windows, thereby obtaining evidence delivery results.

[0090] The differential detection kernel scheduling module is used to perform collision determination and output collision detection results based on the collision semantic cluster, the set of high-value collision opportunity windows and the evidence delivery results.

[0091] The physical consistency verification module is used to verify the consistency of the collision detection results based on the physical response results corresponding to the collision detection results, and generate compensation and correction results.

[0092] The strategy feedback update module is used to update the decision parameters and candidate collision memory state of the improved R2D2 decision unit by compensating for correction results, collision detection results, and operational performance results.

[0093] The steady-state performance shaping module is used to smoothly shape the evidence delivery strength and detection kernel call plan in the target game frame.

[0094] In this embodiment, the physical interaction semantic decoupling module specifically includes:

[0095] Acquire the state data of each object in the game scene within the current game frame and preset historical game frames. The state data includes position coordinates, velocity vector, acceleration vector, object type, contact record, force record, and event trigger record.

[0096] Calculate the motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance of each object based on the state data;

[0097] Calculating motion continuity involves obtaining the amount of velocity change of the target object between the current game frame and the previous game frame, and determining the motion continuity contribution value corresponding to the velocity change based on the velocity change, so that the smaller the velocity change, the larger the corresponding motion continuity contribution value.

[0098] Obtain the change in acceleration of the target object between the current game frame and the previous game frame, and determine the motion continuity contribution value corresponding to the change in acceleration based on the change in acceleration. The smaller the change in acceleration, the larger the corresponding motion continuity contribution value.

[0099] Obtain the angle between the target object's motion direction in the current game frame and the previous game frame, and determine the motion continuity contribution value corresponding to the change in direction based on the angle between the motion directions. The smaller the angle between the motion directions, the greater the corresponding motion continuity contribution value.

[0100] The motion continuity contribution values ​​corresponding to velocity changes, acceleration changes, and direction changes are weighted to obtain the motion continuity of the target object. Each weight is a motion continuity calculation weight, and the sum of all motion continuity calculation weights is one.

[0101] Collision sensitivity is determined by the speed of the target object, the number of candidate contact objects in the neighborhood, and the complexity of the collider. Specifically, the speed of the target object in the current game frame, the number of candidate contact objects in the neighborhood, and the complexity of the target object's collider are obtained and normalized respectively. Then, the normalized speed, the normalized number of candidate contact objects, and the normalized collider complexity are weighted according to the collision sensitivity calculation weight to obtain the collision sensitivity of the target object. The sum of the weights of each collision sensitivity calculation is one.

[0102] The interaction value is determined by the target object's role priority, its criticality in the current task chain, and the frequency of user interaction. Specifically, the target object's role priority, its criticality in the current task chain, and the frequency of user interaction are obtained. The role priority, criticality, and frequency of user interaction are then weighted according to the interaction value calculation weight to obtain the interaction value of the target object. The sum of the weights of each interaction value calculation is one.

[0103] The force propagation characteristics are jointly determined by the force response intensity of the target object, the transmission level in the associated contact link, and the historical frequency of triggering chain responses. Specifically, the force response intensity of the target object in the current game frame, the transmission level of the target object in the associated contact link, and the historical frequency of the target object triggering chain responses are obtained and normalized respectively. Then, the normalized force response intensity, normalized transmission level, and normalized historical frequency are weighted according to the force propagation characteristics to obtain the force propagation characteristics of the target object. The sum of the weights of each force propagation characteristic is one.

[0104] The importance of an event trigger is determined by the probability that the target object triggers the target event, the level of impact after the target event is triggered, and the urgency of the timing of the target event. Specifically, the probability that the target object triggers the target event, the level of impact after the target object triggers the target event, and the urgency of the timing of the target object triggering the target event are obtained. The probability, level of impact, and urgency of the target event are then weighted according to the weight of the event trigger importance calculation to obtain the event trigger importance of the target object. The sum of the weights of each event trigger importance calculation is one.

[0105] The physical interaction semantic vectors of each object are constructed based on motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance.

[0106] Specifically, the motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance of the target object are used as five semantic features to represent the physical interaction state of the target object. These features are then combined in a fixed order to obtain the physical interaction semantic vector corresponding to the target object.

[0107] Semantic clustering is performed based on the physical interaction semantic vectors of each object to obtain the corresponding collision semantic clusters;

[0108] Specifically, the process involves: obtaining the physical interaction semantic vectors of each object, and calculating the differences between any two objects in five dimensions: motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance. Then, assigning corresponding semantic distance calculation weights to the differences in the five dimensions, and summing the weighted differences in each dimension to obtain the semantic distance between the two objects. The sum of the semantic distance calculation weights is one.

[0109] Semantic relationships between objects are constructed based on the semantic distance between them. Based on these semantic relationships, objects are clustered and divided into the same collision semantic cluster. Objects whose semantic distance satisfies the same cluster merging condition are grouped into the same collision semantic cluster, while objects whose semantic distance satisfies the different cluster separation condition are grouped into different collision semantic clusters. After completing the clustering of each object, the collision semantic cluster to which the target object belongs is determined.

[0110] In this embodiment, the inter-frame continuity memory module specifically includes:

[0111] Obtain object state data, physical interaction semantic vectors, and collision semantic clusters in consecutive game frames, and determine candidate collision object pairs based on the object position relationships, velocity relationships, and collision semantic cluster affiliation relationships in consecutive game frames;

[0112] Specifically, the process involves: acquiring the state data of each object in the current game frame and previous consecutive game frames, the physical interaction semantic vectors corresponding to each object, and the collision semantic cluster to which each object belongs; then, for any two objects, extracting the positional relationship, velocity relationship, and collision semantic cluster affiliation relationship between the two objects in consecutive game frames; and identifying the two objects as candidate collision object pairs when they meet the conditions of close positional relationship, similar velocity relationship, and belong to the same collision semantic cluster or have an associated contact relationship in consecutive game frames.

[0113] Construct candidate collision evolution states based on the position and velocity changes of candidate collision objects in consecutive game frames;

[0114] Specifically, for each candidate collision object pair, extract the object distance, relative speed, approach trend value, and number of consecutive approach frames in the current game frame, and use the object distance, relative speed, approach trend value, and number of consecutive approach frames together as the candidate collision evolution state of the candidate collision object pair.

[0115] Among them, the distance between objects is used to represent the spatial proximity of candidate collision object pairs in the current game frame, the relative velocity magnitude is used to represent the relative motion intensity of candidate collision object pairs in the current game frame, the proximity change trend value is used to represent the proximity change of candidate collision object pairs between adjacent game frames, and the number of consecutive proximity frames is used to represent the continuous proximity state of candidate collision object pairs in consecutive game frames.

[0116] The proximity trend value is determined by the change in distance between objects and the change in relative speed between adjacent game frames. Specifically, it is obtained by: obtaining the difference between the distance between candidate collision objects in the previous game frame and the distance between objects in the current game frame, and obtaining the difference between the relative speed of candidate collision objects in the current game frame and the relative speed of objects in the previous game frame. Then, the difference between the distance between objects and the difference between the relative speed are weighted according to the proximity trend calculation weight to obtain the proximity trend value. The sum of the weights of each proximity trend calculation is one.

[0117] Based on the contact relationship records and position correction records in consecutive game frames, historical false detection correction information is recorded. Specifically, the contact relationship records and position correction records of candidate collision object pairs in the current game frame and previous consecutive game frames are obtained.

[0118] Based on the contact relationship record, determine whether the candidate collision object pair has a contact representation in the current game frame, and based on the position correction record, determine whether the position correction result of the candidate collision object pair in the current game frame satisfies the contact direction consistency condition and displacement response condition;

[0119] Among them, when a candidate collision object pair has a contact representation, but the corresponding position correction result does not meet the contact direction consistency condition or the displacement response condition, it is determined that the candidate collision object pair has a false detection correction requirement.

[0120] Based on the contact relationship record, the position correction consistency result, and the false detection correction amplitude corresponding to the false detection correction requirement, the historical false detection correction information is weighted according to the weight of the historical false detection correction information to obtain the historical false detection correction information corresponding to the candidate collision object pair. The sum of the weights of each historical false detection correction information is one.

[0121] Historical false detection correction information is recorded in the temporal memory results of candidate collision object pairs to represent the false detection correction status of candidate collision object pairs in consecutive game frames;

[0122] Based on the position relationship records, contact relationship records, and penetration relationship records in consecutive game frames, historical missed detection correction information is recorded. Specifically, the position relationship records, contact relationship records, and penetration relationship records of candidate collision object pairs in consecutive game frames are obtained.

[0123] Determine whether there is a contact relationship record for the candidate collision object pair in the current game frame; when there is no contact relationship record for the candidate collision object pair in the current game frame, and the position relationship record and the penetration relationship record together indicate that the candidate collision object pair meets the abnormal penetration condition, it is determined that there is a need for missed detection correction for the candidate collision object pair.

[0124] The historical missed detection correction information is determined by the contact relationship missing information, the penetration relationship record information, and the missed detection correction magnitude. The contact relationship missing amount is determined based on the contact relationship record, the penetration relationship amount is determined based on the penetration relationship record, and the contact relationship missing amount, the penetration relationship amount, and the missed detection correction magnitude are weighted according to the historical missed detection correction information calculation weight to obtain the historical missed detection correction information corresponding to the candidate collision object pair. The sum of the calculation weights of each historical missed detection correction information is one.

[0125] Based on the penetration relationship record and position correction record in consecutive game frames, the penetration compensation information is recorded. Specifically, the penetration relationship record and position correction record of the candidate collision object pair in consecutive game frames are obtained. Based on the penetration relationship record, the penetration depth of the candidate collision object pair in the current game frame is determined. Based on the position correction record, the position correction amount and position correction direction consistency of the candidate collision object pair in the current game frame are determined.

[0126] The penetration compensation information is determined by the penetration depth, position correction amount, and position correction direction consistency. The penetration compensation information is weighted according to the penetration depth, position correction amount, and position correction direction consistency according to the calculation weight of the penetration compensation information to obtain the penetration compensation information corresponding to the candidate collision object pair. The sum of the calculation weights of each penetration compensation information is one.

[0127] The temporal memory results of candidate collisions are constructed based on the candidate collision evolution state, historical false detection correction information, historical missed detection correction information and penetration compensation information. Specifically, the temporal memory results of the candidate collision object pair in the previous game frame, as well as the candidate collision evolution state, historical false detection correction information, historical missed detection correction information and penetration compensation information of the candidate collision object pair in the current game frame are obtained.

[0128] The temporal memory results from the previous game frame, the candidate collision evolution state in the current game frame, the historical false detection correction information, the historical missed detection correction information, and the penetration compensation information in the current game frame are fused according to the temporal memory fusion weights to obtain the temporal memory results of the candidate collision object pair in the current game frame. The sum of the temporal memory fusion weights is one.

[0129] In this embodiment, the collision opportunity window planning module specifically includes:

[0130] Obtain collision semantic clusters, object historical trajectory data, scene topology, object interaction relationships and temporal memory results from the current game frame and consecutive historical game frames, and construct candidate collision spatiotemporal fragments;

[0131] Constructing candidate collision spatiotemporal slices involves, for each candidate collision object pair, combining the semantic attribution relationship in the collision semantic cluster, the trajectory evolution relationship in the object's historical trajectory data, the spatial constraint relationship in the scene topology, the interaction association relationship in the object interaction relationship, and the collision evolution state in the temporal memory results, to determine the starting game frame, ending game frame, and corresponding spatial region of the candidate collision object pair in consecutive game frames, and constructing candidate collision spatiotemporal segments; wherein, the candidate collision spatiotemporal segment includes at least the object identifier, the starting game frame, the ending game frame, and the corresponding spatial region;

[0132] Based on object historical trajectory data, scene topology, object interaction relationships, and time-series memory results, collision prior information corresponding to candidate collision spatiotemporal segments is generated. Specifically, this involves: determining trajectory approach information based on object historical trajectory data, which represents the approach trend of candidate collision object pairs in consecutive game frames; determining topological constraint information based on scene topology, which represents the topological blocking relationships, channel relationships, and boundary restriction relationships between candidate collision object pairs; and determining interaction association information based on object interaction relationships, which represents the interaction strength and frequency between candidate collision object pairs.

[0133] Obtain the temporal memory results corresponding to the candidate collision object pairs, and perform weighted processing on trajectory approach information, topological constraint information, interaction association information and temporal memory results according to the collision prior information calculation weight to obtain the collision prior information corresponding to the candidate collision spatiotemporal segment, wherein the sum of the calculation weights of each collision prior information is one;

[0134] The collision value, missed detection risk, and propagation impact of candidate collision spatiotemporal segments are calculated based on prior collision information. Specifically, the collision value of candidate collision spatiotemporal segments is obtained by weighting the collision sensitivity and interaction value of each object in the candidate collision object pair according to the collision value calculation weight. The sum of the calculation weights of each collision value is one.

[0135] Based on the historical missed detection correction information, penetration compensation information and candidate collision evolution status of the candidate collision object, the missed detection risk is obtained by weighting according to the missed detection risk calculation weight. The sum of the weights of each missed detection risk is one.

[0136] Based on the force propagation characteristics of each object in the candidate collision object pair and the interaction association information corresponding to the object interaction relationship, the propagation influence is weighted according to the propagation influence calculation weight to obtain the propagation influence corresponding to the candidate collision spatiotemporal segment, where the sum of the calculation weights of each propagation influence is one.

[0137] Candidate collision spatiotemporal windows are generated based on collision value, missed detection risk, and propagation impact. The candidate collision spatiotemporal windows are then filtered through the opportunity window screening gate in the improved R2D2 decision unit. Specifically, the collision value, missed detection risk, propagation impact, and time-series memory results are weighted according to the window screening calculation weights to obtain the window screening results corresponding to each candidate collision spatiotemporal window. The sum of the window screening calculation weights is one.

[0138] The window filtering results in the current game frame are summed, and the sum is divided by the number of candidate collision spatiotemporal windows in the current game frame to obtain the average window filtering result; the candidate collision spatiotemporal windows whose window filtering result is not less than the average window filtering result are retained as the filtered candidate collision spatiotemporal windows.

[0139] The candidate collision spatiotemporal windows after screening are state-encoded through the collision evidence memory pathway and the physical consistency memory pathway, and the replay priority is determined by the multi-factor priority replay unit to form a window replay sequence.

[0140] By using the dual-output decision head in the improved R2D2 decision unit to perform dynamic planning on each screened candidate collision spatiotemporal window in the window replay sequence, the collision detection benefit value, performance cost value and comprehensive window decision value corresponding to each screened candidate collision spatiotemporal window are obtained.

[0141] Specifically, the candidate collision spatiotemporal windows in the window replay sequence are sequentially input into the dual-output decision head. The first output part of the dual-output decision head performs weighted processing based on the collision value, missed detection risk, propagation impact, and time sequence memory results of the candidate collision spatiotemporal windows, and outputs the collision detection benefit value.

[0142] The second output part of the dual-output decision head performs weighted processing based on the detection load fluctuation value, spatial region range, time span, and detection evidence strength requirements corresponding to the candidate collision spatiotemporal window, and outputs the performance cost value.

[0143] The weights are calculated based on the comprehensive window decision values. The collision detection benefit value and performance cost value are comprehensively processed to obtain the comprehensive window decision value corresponding to each candidate collision spatiotemporal window after screening. The comprehensive window decision value increases with the increase of the collision detection benefit value and decreases with the increase of the performance cost value. The sum of the weights of each comprehensive window decision value is one.

[0144] The candidate collision spatiotemporal windows are sorted according to their comprehensive window decision values ​​after screening, and a set of high-value collision opportunity windows is constructed by selecting candidate collision spatiotemporal windows whose comprehensive window decision values ​​are not less than the average of the comprehensive window decision values.

[0145] Constructing a set of high-value collision opportunity windows involves sorting the comprehensive window decision values ​​corresponding to each selected candidate collision spatiotemporal window, accumulating the comprehensive window decision values ​​corresponding to each selected candidate collision spatiotemporal window, and dividing the accumulated result by the number of selected candidate collision spatiotemporal windows to obtain the average comprehensive window decision value.

[0146] After screening, candidate collision spatiotemporal windows whose comprehensive window decision value is not less than the average of the comprehensive window decision values ​​are selected as high-value collision opportunity windows. The high-value collision opportunity windows are then grouped and organized according to object identifier, start game frame, end game frame and corresponding spatial region to obtain a set of high-value collision opportunity windows.

[0147] In this embodiment, the candidate collision spatiotemporal windows after screening are state-encoded through the collision evidence memory pathway and the physical consistency memory pathway, and the replay priority is determined by the multi-factor priority replay unit to form the window replay sequence, specifically including:

[0148] The candidate collision evolution state, historical missed detection correction information, and penetration compensation information corresponding to the selected candidate collision spatiotemporal window are obtained, and collision evidence is encoded through the collision evidence memory path in the improved R2D2 decision unit.

[0149] Specifically, the following steps are taken: obtain the distance between objects, relative velocity magnitude, proximity trend value, and number of consecutive proximity frames in the candidate collision evolution state corresponding to the candidate collision spatiotemporal window after filtering; and perform weighted processing according to the candidate collision evolution intensity calculation weight based on the distance between objects, relative velocity magnitude, proximity trend value, and number of consecutive proximity frames to obtain the candidate collision evolution intensity, wherein the sum of the calculation weights of each candidate collision evolution intensity is one.

[0150] The historical missed detection correction information and penetration compensation information corresponding to the candidate collision spatiotemporal window after screening are obtained. The candidate collision evolution intensity, historical missed detection correction information and penetration compensation information are input into the collision evidence memory path and weighted according to the collision evidence coding calculation weight to obtain the collision evidence coding result corresponding to the candidate collision spatiotemporal window after screening. The sum of the calculation weights of each collision evidence coding is one.

[0151] The physical consistency violation degree and position correction association state corresponding to the selected candidate collision spatiotemporal window are obtained, and physical consistency encoding is performed through the physical consistency memory path in the improved R2D2 decision unit. Specifically, the position correction consistency result corresponding to the selected candidate collision spatiotemporal window is obtained, and the physical consistency violation degree is determined based on the position correction consistency result; the position correction amount and position correction direction consistency degree corresponding to the selected candidate collision spatiotemporal window are obtained, and the position correction association state is obtained by weighting the position correction amount and position correction direction consistency degree according to the weights calculated by the position correction association state. The sum of the weights calculated for each position correction association state is one.

[0152] The degree of physical consistency violation and the location correction association state are input into the physical consistency memory path, and weighted according to the physical consistency code calculation weight to obtain the physical consistency code result corresponding to the candidate collision spatiotemporal window after screening. The sum of the calculation weights of each physical consistency code is one.

[0153] The collision evidence coding result and the physical consistency coding result are fused to obtain the state coding result corresponding to each candidate collision spatiotemporal window after screening. Specifically, the collision evidence coding result and the physical consistency coding result corresponding to the candidate collision spatiotemporal window after screening are obtained, and the collision evidence coding result and the physical consistency coding result are fused according to the state coding fusion weight to obtain the state coding result corresponding to the candidate collision spatiotemporal window after screening. The sum of the fusion weights of each state coding is one.

[0154] The replay priority of each candidate collision spatiotemporal window after screening is determined by the multi-factor priority replay unit in the improved R2D2 decision unit. Specifically, the state encoding result of the candidate collision spatiotemporal window after screening in the current game frame and the state encoding result in the previous game frame are obtained, and the timing difference error is determined based on the difference between the state encoding result in the current game frame and the state encoding result in the previous game frame.

[0155] The historical missed detection correction information, propagation impact, physical consistency violation degree and detection load fluctuation value corresponding to the selected candidate collision spatiotemporal window are obtained. The temporal difference error, historical missed detection correction information, propagation impact, physical consistency violation degree and detection load fluctuation value are weighted according to the replay priority to obtain the replay priority corresponding to the selected candidate collision spatiotemporal window. The sum of the weights of each replay priority is one.

[0156] The replay priorities corresponding to each selected candidate collision spatiotemporal window are sorted, and a window replay sequence is formed according to the sorting results. Specifically, the replay priorities corresponding to each selected candidate collision spatiotemporal window are obtained and sorted from high to low according to the replay priorities corresponding to each selected candidate collision spatiotemporal window.

[0157] The replay priorities corresponding to each selected candidate collision spatiotemporal window are accumulated, and the accumulated result is divided by the number of selected candidate collision spatiotemporal windows to obtain the average replay priority.

[0158] Selected candidate collision spatiotemporal windows with a replay priority no less than the average replay priority are retained, and window replay sequences are formed in descending order of replay priority.

[0159] In this embodiment, the adaptive evidence delivery module specifically includes:

[0160] The process involves obtaining a set of high-value collision opportunity windows, temporal memory results, and detection load information, and determining the detection evidence delivery targets corresponding to candidate collisions within the windows. Specifically, this includes obtaining a set of high-value collision opportunity windows in the current game frame, temporal memory results for each candidate collision object pair, and detection load information for each candidate collision object pair.

[0161] Traverse each high-value collision opportunity window in the set of high-value collision opportunity windows, extract the candidate collision object pairs corresponding to each high-value collision opportunity window, and determine the candidate collision object pairs as the detection evidence delivery objects; wherein, each detection evidence delivery object establishes an association relationship with the corresponding high-value collision opportunity window, time-series memory results, and detection load information respectively;

[0162] The required detection evidence value for candidate collisions is calculated based on the high-value collision opportunity window, time-series memory results, and detection load information. Specifically, the comprehensive window decision value, time-series memory results, and detection load information for each detection evidence delivery object are obtained. The weights are calculated according to the required detection evidence value. The comprehensive window decision value, time-series memory results, and detection load information after negative transformation are weighted to obtain the required detection evidence value for each detection evidence delivery object. The sum of the weights of each required detection evidence value is one, and the larger the detection load information, the smaller the contribution value of the detection load information after negative transformation.

[0163] The collision evidence memory path and physical consistency memory path in the improved R2D2 decision unit are used to perform state encoding processing on the detection evidence requirement value, and the multi-factor priority playback unit determines the detection evidence delivery priority. The corresponding collision detection benefit value, performance cost value, detection evidence requirement value, state encoding result and detection evidence delivery priority are output through the dual-output decision head.

[0164] Based on the collision detection benefit value, performance cost value, detection evidence requirement value, state coding result and detection evidence delivery priority, the basic delivery value of trajectory interpolation depth, the basic delivery value of continuous detection length, the basic delivery value of boundary sampling density, the basic delivery value of historical frame call range, and the basic delivery value of high-precision detection kernel trigger times are calculated respectively.

[0165] Among them, the collision evidence memory pathway is used to encode collision evidence based on the candidate collision evolution state, historical missed detection correction information and penetration compensation information to obtain the collision evidence encoding result.

[0166] The physical consistency memory path is used to modify the associated state based on the degree and location of physical consistency violation to perform physical consistency encoding processing and obtain the physical consistency encoding result.

[0167] The multi-factor priority replay unit is used to perform weighted processing based on collision evidence coding results, physical consistency coding results, propagation impact and detection load information to obtain the detection evidence delivery priority.

[0168] The first output part of the dual-output decision head is used to output the collision detection benefit value, and the second output part of the dual-output decision head is used to output the performance cost value;

[0169] Based on the detection load information, the basic deployment values ​​are adjusted by load constraints to obtain the adjusted deployment values ​​corresponding to trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame call range, and high-precision detection kernel trigger count. The load constraint adjustment includes adjusting the load constraints of each basic deployment value according to the detection load information so that when the detection load information increases, the adjustment result corresponding to each basic deployment value decreases synchronously, resulting in the trajectory interpolation depth adjustment deployment value, continuous detection length adjustment value, boundary sampling density adjustment value, historical frame call range adjustment value, and high-precision detection kernel trigger count adjustment value.

[0170] The detection evidence strength corresponding to the candidate collision is constructed based on the adjusted delivery values. Specifically, the trajectory interpolation depth adjustment delivery value, continuous detection length adjustment delivery value, boundary sampling density adjustment delivery value, historical frame call range adjustment delivery value, and high-precision detection kernel trigger number adjustment delivery value corresponding to each detection evidence delivery object are obtained. They are then combined in a fixed order of trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame call range, and high-precision detection kernel trigger number to obtain the detection evidence strength corresponding to each detection evidence delivery object.

[0171] Evidence delivery results are generated based on the strength of detected evidence; the strength of detected evidence corresponding to each evidence delivery object is obtained; the weights of the evidence delivery results are calculated, and the trajectory interpolation depth adjustment delivery value, continuous detection length adjustment delivery value, boundary sampling density adjustment delivery value, historical frame call range adjustment delivery value, and high-precision detection kernel trigger count adjustment delivery value in the strength of detected evidence are weighted to obtain the evidence delivery results corresponding to each evidence delivery object, wherein the sum of the weights of each evidence delivery result is one.

[0172] In this embodiment, the differential detection kernel scheduling module specifically includes:

[0173] The collision semantic cluster to which an object belongs, the set of high-value collision opportunity windows, the evidence delivery results and the strength of detection evidence are obtained, and the initial candidate collisions are determined. Specifically, the collision semantic cluster to which each object in the current game frame belongs, the high-value collision opportunity window corresponding to each candidate collision object pair, the evidence delivery results corresponding to each candidate collision object pair, and the strength of detection evidence corresponding to each candidate collision object pair are obtained.

[0174] Traverse each high-value collision opportunity window in the set of high-value collision opportunity windows, extract the candidate collision object pairs corresponding to each high-value collision opportunity window, and determine the candidate collision object pairs as initial candidate collisions; wherein, each initial candidate collision establishes an association relationship with the corresponding collision semantic cluster, high-value collision opportunity window, evidence delivery result, and detection evidence strength respectively.

[0175] Based on the object interaction relationships, force propagation characteristics, and event triggering importance corresponding to the initial candidate collisions, an extended analysis is performed on the associated contact objects, force transmission objects, and event triggering objects to obtain the set of associated contact objects, the set of force transmission objects, and the set of event triggering objects.

[0176] Specifically, for each initial candidate collision, traverse the adjacent objects that have an object interaction relationship with the two objects in the initial candidate collision.

[0177] Based on the object interaction relationships between the two objects in the initial candidate collision and each of their neighboring objects, the associated contact value corresponding to each neighboring object is calculated. The associated contact value corresponding to each neighboring object is obtained by averaging the object interaction relationships between the two objects and their neighboring objects.

[0178] Based on the force propagation characteristics of each adjacent object and the object interaction relationship between each object and its adjacent objects, the force transmission value corresponding to each adjacent object is calculated. The force transmission value corresponding to each adjacent object is obtained by multiplying the force propagation characteristics of the adjacent objects by the average value of the object interaction relationship between each object and its adjacent objects.

[0179] Based on the event triggering importance of each adjacent object and the object interaction relationship between each object and its adjacent objects, the event triggering value corresponding to each adjacent object is calculated. The event triggering value corresponding to each adjacent object is obtained by multiplying the event triggering importance of the adjacent object by the average of the object interaction relationship between each object and its adjacent objects.

[0180] The associated contact value, force transmission value, and event trigger value corresponding to all adjacent objects are accumulated and then divided by the number of adjacent objects to obtain the average associated contact value, average force transmission value, and average event trigger value.

[0181] Adjacent objects with a correlation contact value not less than the average correlation contact value are identified as correlation contact objects, forming a set of correlation contact objects; adjacent objects with a force transmission value not less than the average force transmission value are identified as force transmission objects, forming a set of force transmission objects; adjacent objects with an event trigger value not less than the average event trigger value are identified as event trigger objects, forming a set of event trigger objects.

[0182] The impact of collision propagation is calculated based on the set of related contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the results of evidence deployment.

[0183] Based on the proportion of the number of objects in the associated contact object set to the number of adjacent objects, the average value of all associated contact values ​​in the associated contact object set, and the evidence delivery results, the associated contact impact value is obtained by weighting according to the weight of the associated contact impact value calculation. The sum of the weights of each associated contact impact value calculation is one.

[0184] Based on the proportion of the number of objects in the force transmission object set to the number of adjacent objects, the average value of all force transmission values ​​in the force transmission object set, and the average value of the force propagation characteristics of all objects in the force transmission object set, the force transmission influence value is obtained by weighting according to the weight of the force transmission influence value calculation. The sum of the weights of each force transmission influence value calculation is one.

[0185] The event trigger impact value is obtained by weighting the event trigger impact value according to the proportion of the number of objects in the event trigger object set to the number of adjacent objects, the average of all event trigger values ​​in the event trigger object set, and the average of the event trigger importance of all objects in the event trigger object set. The sum of the weights of each event trigger impact value is one.

[0186] Based on the impact values ​​of associated contact, force transmission, event triggering, and evidence placement, the impact values ​​of collision propagation are weighted according to their respective weights to obtain the impact results of collision propagation. The sum of the weights of each impact result is one.

[0187] The kernel matching value corresponding to each collision detection kernel is calculated based on the collision propagation impact results and the strength of detection evidence. Specifically, the deployment value is adjusted based on the collision propagation impact results and the number of triggers of the high-precision detection kernel after negative conversion. The kernel matching value corresponding to the fast exclusion detection kernel is obtained by weighting according to the calculation weight of the fast exclusion detection kernel matching value. The sum of the calculation weights of each fast exclusion detection kernel matching value is one.

[0188] The deployment value is adjusted based on the impact of collision propagation, the continuous detection length, and the historical frame call range. The deployment value is then weighted according to the weights calculated by the continuous collision detection kernel matching values ​​to obtain the kernel matching value corresponding to the continuous collision detection kernel. The sum of the weights calculated for each continuous collision detection kernel matching value is one.

[0189] Based on the collision propagation impact results, the trajectory interpolation depth adjustment deployment value, and the continuous detection length adjustment deployment value, the weighted processing is performed according to the weight calculated by the trajectory interpolation detection kernel matching value to obtain the kernel matching value corresponding to the trajectory interpolation detection kernel. The sum of the weights calculated for each trajectory interpolation detection kernel matching value is one.

[0190] Based on the collision propagation impact results, the boundary sampling density adjustment value, and the high-precision detection kernel trigger count adjustment value, the kernel matching value corresponding to the boundary high-precision detection kernel is obtained by weighting according to the weight calculated by the boundary high-precision detection kernel matching value. The sum of the weights of each boundary high-precision detection kernel matching value is one.

[0191] Based on the impact results of collision propagation, the impact value of force transmission, and the impact value of event triggering, the kernel matching value corresponding to the propagation response detection kernel is obtained by weighting according to the weight calculated by the kernel matching value of the propagation response detection kernel. The sum of the weights of each kernel matching value is one.

[0192] Among them, the high-precision detection core trigger count adjustment value after negative conversion is obtained by subtracting the high-precision detection core trigger count adjustment value from one;

[0193] Based on the kernel matching value corresponding to each collision detection kernel, the collision detection kernel that matches the candidate collision is determined. Specifically, the kernel matching value corresponding to the fast exclusion detection kernel, the kernel matching value corresponding to the continuous collision detection kernel, the kernel matching value corresponding to the trajectory interpolation detection kernel, the kernel matching value corresponding to the boundary high-precision detection kernel, and the kernel matching value corresponding to the propagation response detection kernel are obtained.

[0194] Then compare the kernel matching values ​​corresponding to each collision detection kernel, and select the collision detection kernel with the largest kernel matching value as the collision detection kernel that matches the candidate collision.

[0195] The collision detection kernel is invoked to perform collision determination on the initial candidate collision, the set of associated contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the strength of detection evidence, and the collision detection result is obtained. Specifically, the collision detection kernel that matches the candidate collision is invoked, and the initial candidate collision, the set of associated contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the strength of detection evidence are used as input to the collision detection kernel.

[0196] Based on the detection method corresponding to the collision detection kernel, collision determination processing is performed on the initial candidate collision and its extended objects, and the collision detection result is output. The collision detection result includes at least the collision occurrence mark, collision contact position, collision contact time, collision contact direction, and the set of associated object identifiers corresponding to the initial candidate collision.

[0197] In this embodiment, the physical consistency verification module specifically includes:

[0198] Obtain collision detection results and corresponding physical response results, wherein the physical response results include at least the change in response velocity, the correction of response displacement, the consistency of response direction, and the duration of response.

[0199] The response speed coefficient, response displacement coefficient, response direction coefficient, and response duration coefficient are calculated based on the physical response results. Specifically, the response speed change corresponding to all initial candidate collisions in the current game frame is obtained, and all response speed changes are accumulated. The accumulated result is divided by the number of initial candidate collisions to obtain the average response speed change in the current game frame.

[0200] Divide the change in response velocity corresponding to each initial candidate collision by the mean change in response velocity to obtain the response velocity coefficient corresponding to each initial candidate collision.

[0201] Obtain the response displacement correction amount corresponding to all initial candidate collisions in the current game frame, accumulate all response displacement correction amounts, and then divide the accumulated result by the number of initial candidate collisions to obtain the average response displacement correction amount in the current game frame.

[0202] Divide the response displacement correction amount corresponding to each initial candidate collision by the mean of the response displacement correction amount to obtain the response displacement coefficient corresponding to each initial candidate collision.

[0203] The consistency of the response direction corresponding to each initial candidate collision is used as the response direction coefficient corresponding to each initial candidate collision.

[0204] Obtain the response duration corresponding to all initial candidate collisions in the current game frame, accumulate all response durations, and then divide the accumulated result by the number of initial candidate collisions to obtain the average response duration in the current game frame.

[0205] Divide the response duration of each initial candidate collision by the average response duration to obtain the response duration coefficient of each initial candidate collision.

[0206] The physical consistency verification value is calculated based on the collision detection results and the response velocity coefficient, response displacement coefficient, response direction coefficient, and response duration coefficient. Specifically, the collision occurrence marker is extracted from each collision detection result, and the weights are calculated according to the physical consistency verification value. The product of the collision occurrence marker and the response velocity coefficient, the product of the collision occurrence marker and the response displacement coefficient, the response direction coefficient, and the response duration coefficient are weighted to obtain the physical consistency verification value corresponding to each initial candidate collision. The sum of the weights of each physical consistency verification value is one.

[0207] The average physical consistency check value is calculated based on the physical consistency check values ​​corresponding to all initial candidate collisions in the current game frame. Specifically, the physical consistency check values ​​corresponding to all initial candidate collisions in the current game frame are obtained, and all physical consistency check values ​​are accumulated. The accumulated result is then divided by the number of initial candidate collisions in the current game frame to obtain the average physical consistency check value of the current game frame.

[0208] The compensation and correction results are generated based on the physical consistency check value and the average physical consistency check value. Specifically, the physical consistency check value corresponding to each initial candidate collision is compared with the average physical consistency check value.

[0209] When the physical consistency check value corresponding to an initial candidate collision is less than the average physical consistency check value, it is determined that the consistency of the initial candidate collision does not meet the check requirements, and a corresponding compensation correction result is generated. The compensation correction result is obtained by weighting the difference between the average physical consistency check value and the physical consistency check value, historical missed detection correction information and penetration compensation information according to the weight of the compensation correction result calculation, and the sum of the weights of each compensation correction result is one.

[0210] The degree of physical consistency violation and the amount of replay correction are determined based on the compensation and correction results. Specifically, the degree of physical consistency violation is obtained by weighting the compensation and correction results and the response direction coefficient according to the weights calculated for the degree of physical consistency violation. The degree of physical consistency violation increases with the increase of the compensation and correction results and decreases with the increase of the response direction coefficient. The sum of the weights calculated for each degree of physical consistency violation is one.

[0211] Based on the compensation correction results and the difference between the average physical consistency check value and the physical consistency check value, the replay correction amount is calculated by weighting according to the weight of the replay correction amount to obtain the replay correction amount. The sum of the weights of each replay correction amount is one.

[0212] The compensation and correction results are input into the physical consistency memory path and the multi-factor priority replay unit in the improved R2D2 decision unit. Specifically, the compensation and correction results and the degree of physical consistency violation are input into the physical consistency memory path to update the physical consistency encoding results of the corresponding candidate collision spatiotemporal window after screening.

[0213] The compensation correction results, replay correction amount, and physical consistency violation degree are input into the multi-factor priority replay unit to update the replay priority of the corresponding candidate collision spatiotemporal window after screening.

[0214] In this embodiment, the strategy feedback update module specifically includes:

[0215] Obtain compensation and correction results, collision detection results, and operational performance results. The operational performance results shall include at least the detection time, detection load information, and resource consumption.

[0216] The candidate collision memory state is updated based on the compensation correction results, collision detection results, and time sequence memory results. Specifically, the collision occurrence marker is extracted from each collision detection result, and the physical consistency check value and time sequence memory result corresponding to each initial candidate collision are obtained.

[0217] According to the update weight of the candidate collision memory state, the temporal memory result, the compensation correction result, the collision occurrence mark and the physical consistency check value are weighted to obtain the updated candidate collision memory state, wherein the sum of the update weights of each candidate collision memory state is one.

[0218] The collision detection benefit feedback value and performance cost feedback value are calculated based on the compensation correction results, collision detection results, collision propagation impact results, and operational performance results. Specifically, the collision detection benefit feedback value is obtained by weighting the collision occurrence marker, physical consistency verification value, collision propagation impact results, and compensation correction results after negative transformation according to the calculation weight of the collision detection benefit feedback value. The sum of the calculation weights of each collision detection benefit feedback value is one, and the larger the compensation correction result, the smaller the corresponding contribution value of the compensation correction result after negative transformation.

[0219] Based on the detection duration, detection load information, and resource usage, the performance cost feedback value is calculated and weighted according to the weights to obtain the performance cost feedback value. The sum of the weights of each performance cost feedback value is one.

[0220] The timing difference error is calculated based on the collision detection benefit feedback value, the performance cost feedback value, and the collision detection benefit value and performance cost value output by the dual-output decision head. Specifically, the collision detection benefit value and performance cost value output by the dual-output decision head are obtained, the absolute value of the difference between the collision detection benefit feedback value and the collision detection benefit value are calculated respectively to obtain the benefit timing difference error, and the absolute value of the difference between the performance cost feedback value and the performance cost value is calculated to obtain the cost timing difference error.

[0221] The revenue time series difference error and the cost time series difference error are weighted according to the weight of the comprehensive time series difference error to obtain the comprehensive time series difference error. The sum of the weights of each comprehensive time series difference error is one.

[0222] The priority of the experience sequence replay is updated based on the comprehensive temporal difference error, historical missed detection correction information, collision propagation level, physical consistency violation degree and detection load fluctuation value. Specifically, the comprehensive temporal difference error, historical missed detection correction information, collision propagation level, physical consistency violation degree and detection load fluctuation value corresponding to each initial candidate collision are obtained.

[0223] The weights are updated according to the priority of the experience sequence replay. The weights of the comprehensive time-series differential error, historical missed detection correction information, collision propagation level, physical consistency violation degree and detection load fluctuation value are weighted to obtain the updated experience sequence replay priority. The sum of the update weights of each experience sequence replay priority is one.

[0224] The decision parameters of the improved R2D2 decision unit are updated based on the comprehensive temporal difference error and the empirical sequence replay priority. Specifically, the parameter update target is constructed based on the comprehensive temporal difference error and the updated empirical sequence replay priority corresponding to all initial candidate collisions in the current game frame.

[0225] The current decision parameters of the improved R2D2 decision unit are updated by gradient according to the parameter update objective to obtain the updated decision parameters. The higher the priority of the updated empirical sequence replay and the larger the comprehensive temporal difference error, the greater the influence of the initial candidate collision on the parameter update objective.

[0226] The updated candidate collision memory state is fed back to the inter-frame continuous memory module, and the updated experience sequence replay priority and updated decision parameters are fed back to the collision opportunity window planning module and the adaptive delivery module of detection evidence. The decision parameters corresponding to the multi-factor priority replay unit and the dual-output decision head in the improved R2D2 decision unit are updated.

[0227] In this embodiment, the steady-state performance shaping module specifically includes:

[0228] Obtain the evidence delivery results, collision detection results, collision propagation impact results, compensation and correction results, and running performance results for each game frame within the statistical game frame interval. Also obtain the detection evidence strength and high-precision detection kernel trigger count adjustment delivery value in the target game frame. Specifically, determine the statistical game frame interval as the consecutive game frames before the current game frame.

[0229] Obtain the evidence delivery results, collision detection results, collision propagation impact results, compensation and correction results, and runtime performance results for each game frame within the statistical game frame interval; among which, the runtime performance results include at least the detection duration, detection load information, and resource consumption.

[0230] Obtain the detection evidence strength corresponding to each candidate collision object pair in the target game frame. The detection evidence strength includes the trajectory interpolation depth adjustment value, the continuous detection length adjustment value, the boundary sampling density adjustment value, the historical frame call range adjustment value, and the high-precision detection kernel trigger count adjustment value.

[0231] The execution pressure value is calculated based on the evidence delivery results and running performance results corresponding to each game frame within the statistical game frame interval, and the detection maintenance value is calculated based on the collision detection results, collision propagation impact results, and compensation correction results.

[0232] Specifically, for each game frame within the statistical game frame interval, the evidence delivery results, detection duration, detection load information and resource consumption corresponding to all initial candidate collisions in the game frame are obtained respectively.

[0233] The execution pressure value for each game frame is calculated by weighting the average evidence delivery results, average detection time, average detection load information, and average resource usage corresponding to all initial candidate collisions in the game frame. The sum of the weights for each execution pressure value is one.

[0234] For each game frame within the statistical game frame interval, the collision occurrence marker is extracted from the collision detection results, and the collision propagation impact results and compensation correction results corresponding to all initial candidate collisions in the game frame are obtained.

[0235] The detection maintenance value for each game frame is obtained by weighting the average collision occurrence marker, the average collision propagation impact result, and the average compensation correction result corresponding to all initial candidate collisions in the game frame according to the weight calculation of the detection maintenance value. The sum of the weights of each detection maintenance value is one.

[0236] The steady-state performance shaping factor is calculated based on the execution pressure value and detection maintenance value in the statistical game frame interval;

[0237] Specifically, the execution pressure values ​​corresponding to each game frame within the statistical game frame interval are accumulated, and the accumulated result is divided by the number of game frames within the statistical game frame interval to obtain the average execution pressure.

[0238] The detection maintenance values ​​corresponding to each game frame within the statistical game frame interval are accumulated, and the accumulated result is divided by the number of game frames within the statistical game frame interval to obtain the detection maintenance mean.

[0239] Based on the ratio between the execution pressure value corresponding to the current game frame and the average execution pressure, and the ratio between the detection maintenance value corresponding to the current game frame and the average detection maintenance value, the steady-state performance shaping coefficient is obtained by combining the weights calculated according to the steady-state performance shaping coefficient. The sum of the weights of each steady-state performance shaping coefficient is one. When the execution pressure value corresponding to the current game frame increases relative to the average execution pressure, the steady-state performance shaping coefficient increases. When the detection maintenance value corresponding to the current game frame increases relative to the average detection maintenance value, the steady-state performance shaping coefficient decreases.

[0240] The detection evidence strength in the target game frame is smoothed and shaped according to the steady-state performance shaping coefficient to obtain the smoothed detection evidence strength. Specifically, the following steps are taken: the trajectory interpolation depth adjustment value, continuous detection length adjustment value, boundary sampling density adjustment value, historical frame call range adjustment value, and high-precision detection kernel trigger number adjustment value are obtained for each candidate collision object pair in the target game frame.

[0241] Based on the steady-state performance shaping coefficients, the trajectory interpolation depth adjustment value, continuous detection length adjustment value, boundary sampling density adjustment value, historical frame call range adjustment value, and high-precision detection kernel trigger count adjustment value are smoothed and shaped to obtain the smoothed trajectory interpolation depth adjustment value, smoothed continuous detection length adjustment value, smoothed boundary sampling density adjustment value, smoothed historical frame call range adjustment value, and smoothed high-precision detection kernel trigger count adjustment value.

[0242] Among them, when the steady-state performance shaping coefficient increases, the corresponding smoothing shaping result decreases, and when the steady-state performance shaping coefficient decreases, the corresponding smoothing shaping result increases.

[0243] The detection kernel call plan value of the target game frame is calculated based on the strength of detection evidence after smoothing and the impact of collision propagation in the target game frame;

[0244] Specifically, this involves: obtaining the impact results of each candidate collision object on the corresponding collision propagation in the target game frame, as well as the adjustment values ​​for trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame call range, and high-precision detection kernel trigger count after smoothing.

[0245] The deployment value is adjusted based on the collision propagation effect and the number of triggers of the high-precision detection kernel after negative transformation and smoothing. The value is then weighted according to the calculation weight of the fast exclusion detection kernel call plan value to obtain the fast exclusion detection kernel call plan value. The sum of the calculation weights of each fast exclusion detection kernel call plan value is one.

[0246] The deployment value is adjusted based on the impact of collision propagation, the continuous detection length after smoothing and shaping, and the historical frame call range after smoothing and shaping. The deployment value is then weighted according to the weights calculated by the continuous collision detection kernel call plan value to obtain the continuous collision detection kernel call plan value. The sum of the weights calculated for each continuous collision detection kernel call plan value is one.

[0247] Based on the collision propagation impact results, the trajectory interpolation depth adjustment deployment value after smoothing and the continuous detection length adjustment deployment value after smoothing, the trajectory interpolation detection kernel call plan value is weighted according to the weights calculated by the trajectory interpolation detection kernel call plan value to obtain the trajectory interpolation detection kernel call plan value. The sum of the weights calculated by each trajectory interpolation detection kernel call plan value is one.

[0248] The deployment value is adjusted based on the collision propagation effect, the boundary sampling density after smoothing and shaping, and the number of triggers of the high-precision detection kernel after smoothing and shaping. The deployment value is then weighted according to the weights calculated for the boundary high-precision detection kernel call plan value to obtain the boundary high-precision detection kernel call plan value. The sum of the weights calculated for each boundary high-precision detection kernel call plan value is one.

[0249] The deployment value is adjusted based on the impact of collision propagation, the number of triggers of the high-precision detection kernel after smoothing and shaping, and the historical frame call range after smoothing and shaping. The deployment value is then weighted according to the weights calculated by the propagation response detection kernel call plan value to obtain the propagation response detection kernel call plan value. The sum of the weights calculated for each propagation response detection kernel call plan value is one.

[0250] Among them, the adjusted deployment value of the high-precision detection kernel trigger count after negative transformation and smoothing is obtained by subtracting the adjusted deployment value of the high-precision detection kernel trigger count after smoothing from one;

[0251] The detection kernel call plan in the target game frame is determined based on the detection kernel call plan value, and the smoothed and shaped detection evidence strength and detection kernel call plan are output to the detection evidence adaptive delivery module and the differential detection kernel scheduling module.

[0252] Specifically, this involves: obtaining the fast elimination detection kernel call plan value, continuous collision detection kernel call plan value, trajectory interpolation detection kernel call plan value, boundary high-precision detection kernel call plan value, and propagation response detection kernel call plan value for each candidate collision object pair in the target game frame;

[0253] Compare the values ​​of the detection kernel call plans and select the detection kernel with the largest detection kernel call plan value as the candidate collision target pair in the target game frame;

[0254] The smoothed trajectory interpolation depth adjustment value, the smoothed continuous detection length adjustment value, the smoothed boundary sampling density adjustment value, the smoothed historical frame call range adjustment value, and the smoothed high-precision detection kernel trigger count adjustment value are combined in a fixed order to obtain the smoothed detection evidence strength.

[0255] The strength of the detection evidence after smoothing and shaping is output to the adaptive delivery module of detection evidence, and the detection kernel call plan is output to the differential detection kernel scheduling module.

[0256] Example 1: To verify the feasibility of this invention in practice, it was applied to the physical interaction test environment of an open-world 3D action game. This test environment included game physics scenarios such as character movement, melee attacks, collisions with remote projectiles, obstacle contact, destructible scene interaction, and chain reaction force propagation. Under the same engine configuration, the same number of objects, and the same rendering load, the performance of a fixed-rule collision detection scheme and the scheme of this invention were compared. The test environment simultaneously deployed high-speed moving objects, continuously displaced objects, densely packed contact objects, and interactive objects with event-triggered attributes. This caused the collision detection process to simultaneously face problems such as high-speed penetration, local congestion, chain reaction collision propagation, and competition for detection resources. This allowed for a more complete demonstration of the problems of missed detections, false detections, local computational power accumulation, and continuous game frame load fluctuations that easily occur in existing technologies relying on single-frame states and fixed detection frequencies.

[0257] In this scenario, the physical interaction semantic decoupling module first reads the position, velocity, acceleration, contact records, and interaction records of game objects in consecutive game frames, and classifies high-speed projectiles, character bodies, destructible objects, scene obstacles, and event triggers into different collision semantic clusters. Subsequently, the inter-frame continuity memory module records the distance changes, approach trends, missed detection correction information, and penetration compensation information between candidate collision object pairs in consecutive game frames, forming a temporal memory result. The collision opportunity window planning module generates candidate collision spatiotemporal windows based on the collision semantic clusters, object historical trajectory data, scene topology, object interaction relationships, and temporal memory results. The window selection result is output through the opportunity window filtering gate in the improved R2D2 decision unit, and then the collision detection benefit value and performance cost value are output through a dual-output decision head. The detection evidence adaptive delivery module dynamically allocates trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame call range, and high-precision detection kernel trigger count based on the high-value collision opportunity window set, ensuring that high-risk objects receive higher detection evidence strength, while low-risk objects maintain lower detection evidence strength. The differentiated detection kernel scheduling module further integrates the collision propagation impact results, switching between the fast exclusion detection kernel, continuous collision detection kernel, trajectory interpolation detection kernel, boundary high-precision detection kernel, and propagation response detection kernel. Next, the physical consistency verification module performs consistency checks using the response speed change, response displacement correction, response direction consistency, and response duration corresponding to the collision detection results. It generates compensation correction results for candidate collisions with insufficient consistency and adjusts the candidate collision memory state and experience sequence replay priority through the policy feedback update module. Finally, the steady-state performance shaping module smooths and shapes the detection evidence strength and detection kernel call plan in subsequent target game frames based on the evidence delivery results, collision detection results, collision propagation impact results, and runtime performance results across multiple consecutive game frames, thereby reducing computational spikes in local frames.

[0258] In multiple rounds of closed testing, the fixed-rule collision detection scheme tends to suffer from concentrated calls to high-precision detection kernels in local frames under scenarios involving high-speed object approach, dense multi-person collisions, and chain propagation of destructible objects. This leads to increased detection time per frame and a lag in compensation after penetration during rapid displacement and staggered contact. The present invention, by introducing collision opportunity window planning and adaptive delivery of detection evidence, can enhance the strength of detection evidence for potentially effective contact pairs before a candidate collision actually occurs. Simultaneously, a physical consistency backtesting module performs closed-loop correction on collision detection results and physical response results, ensuring that collision determination no longer relies solely on the geometric contact results of the current game frame but incorporates response changes across consecutive game frames to correct for abnormal penetration and missed detections. Test results show that, under the same test load, the present invention significantly reduces the missed detection rate in high-speed projectile scenarios, maintains a low false detection rate in chaotic group battle scenarios, exhibits more complete chain collision responses in destructible scenarios, and suppresses fluctuations in detection time between consecutive game frames. This demonstrates that the present invention not only improves collision detection accuracy but also enhances continuous operational stability.

[0259] Furthermore, in the same test environment, repeated runs were performed on six representative scenarios, including high-speed projectile crossing scenarios, multi-player melee combat scenarios, narrow passage congestion scenarios, vehicle collision and obstacle collision scenarios, destructible object chain propagation scenarios, and composite event triggering scenarios. Both the fixed-rule collision detection scheme and the present invention's scheme underwent the same number of replay tests, and data was collected under the same number of objects, map size, rendering configuration, and physics engine step size. Statistical data shows that the present invention's scheme achieved a higher average collision detection accuracy in all six scenarios than the fixed-rule collision detection scheme. The average missed detection rate and the number of abnormal penetration events were significantly reduced. The average single-frame detection time only increased slightly or remained basically consistent in some scenarios, while the inter-frame duration fluctuations, compensation correction lag, and the number of invalid calls to the high-precision detection kernel were significantly reduced. This demonstrates that the present invention, through a continuous processing chain of "candidate collision spatiotemporal window identification—detection evidence delivery—differentiated detection kernel scheduling—physical consistency verification—steady-state performance shaping," can achieve a balance between collision detection accuracy and runtime performance in complex game physics scenarios.

[0260] Table 1 Comparison of the implementation effects of the present invention's scheme and the fixed-rule collision detection scheme

[0261] Test Scenario object size Collision detection accuracy (fixed rule / this invention) False negative rate (fixed rule / this invention) Average single-frame detection time (milliseconds, fixed rule / this invention) Inter-frame duration fluctuation (milliseconds, fixed rule / this invention) Number of abnormal penetration events (fixed rule / this invention) High-precision detection of invalid kernel call counts (fixed rule / this invention) High-speed projectile traversing the scene 320 91.8% / 97.9% 5.7% / 1.3% 5.84 / 6.02 2.31 / 1.06 23 / 4 118 / 39 Multiplayer melee battle scene 540 89.6% / 96.4% 6.4% / 1.9% 8.92 / 9.11 3.76 / 1.85 31 / 7 156 / 52 Narrow passage crowded contact scenario 460 90.7% / 96.8% 5.8% / 1.8% 7.65 / 7.88 3.14 / 1.42 19 / 5 132 / 44 Vehicle collision and obstacle collision scenarios 280 92.4% / 97.2% 4.9% / 1.5% 6.73 / 6.95 2.88 / 1.21 17 / 3 96 / 28 Destructible object chain propagation scenario 610 88.9% / 95.7% 6.9% / 2.2% 10.26 / 10.54 4.42 / 2.03 38 / 9 184 / 63 Composite event triggering scenarios 490 90.1% / 96.1% 6.1% / 2.0% 8.37 / 8.56 3.51 / 1.67 26 / 6 149 / 47 average value 450 90.58% / 96.68% 5.97% / 1.78% 7.96 / 8.18 3.34 / 1.54 25.67 / 5.67 139.17 / 45.50

[0262] As can be seen from the table above, the present invention demonstrates relatively stable optimization effects across multiple key metrics in complex game physics scenarios. Regarding collision detection accuracy, the present invention achieves accuracy rates of 97.9%, 96.4%, 96.8%, 97.2%, 95.7%, and 96.1% in six scenarios: high-speed projectile passage, multi-player melee combat, congested contact in narrow passages, vehicle collisions and obstacle collisions, chain reactions of destructible objects, and compound event triggering, respectively. These accuracy rates are all higher than those of the fixed-rule solutions (91.8%, 89.6%, 90.7%, 92.4%, 88.9%, and 90.1%), with the average accuracy increasing from 90.58% to 96.68%. Simultaneously, the average false negative rate decreased from 5.97% to 1.78%, indicating that the present invention improves the completeness of candidate collision identification under conditions of high-speed approach, continuous contact, and chain reactions.

[0263] From the perspective of operational stability indicators, although the average single-frame detection time of this invention increased from 7.96 milliseconds to 8.18 milliseconds, the increase was relatively limited. However, the inter-frame duration fluctuation decreased from 3.34 milliseconds to 1.54 milliseconds, indicating that the distribution of detection overhead between consecutive game frames was more stable. This improvement was even more pronounced in scenarios involving chain propagation of destructible objects and multiplayer melee combat, where the inter-frame fluctuation decreased from 4.42 milliseconds to 2.03 milliseconds in the former and from 3.76 milliseconds to 1.85 milliseconds in the latter. This demonstrates that this invention, through collision opportunity window planning, adaptive delivery of detection evidence, and steady-state performance shaping mechanisms, makes the allocation of detection resources in high-load scenarios closer to the requirements of continuous operation.

[0264] Furthermore, the number of abnormal penetration events and the number of invalid calls to the high-precision detection kernel also showed a downward trend. The average number of abnormal penetration events in the six scenarios decreased from 25.67 to 5.67, and the average number of invalid calls to the high-precision detection kernel decreased from 139.17 to 45.50, indicating that while improving collision detection accuracy, the present invention has imposed more targeted constraints on the scope of use of high-precision detection resources. In summary, under complex game physics interaction conditions, the present invention can improve collision detection accuracy, reduce missed detections and abnormal penetration, and improve execution stability in consecutive game frames while maintaining acceptable detection duration.

[0265] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A game physics collision accuracy detection and performance optimization system based on reinforcement learning, characterized in that, include: The physical interaction semantic decoupling module is used to obtain the state data of each object in the game scene, perform semantic cluster division, and obtain collision semantic clusters; The inter-frame continuity memory module is used to record information in consecutive game frames to form the temporal memory results of candidate collisions; The collision opportunity window planning module is used to generate collision prior information based on collision semantic clusters and temporal memory results, and obtain a set of high-value collision opportunity windows through an improved R2D2 decision unit. The adaptive evidence delivery module is used to assign differentiated detection evidence strengths to candidate collisions within a set of high-value collision opportunity windows, thereby obtaining evidence delivery results. The differential detection kernel scheduling module is used to perform collision determination and output collision detection results based on the collision semantic cluster, the set of high-value collision opportunity windows and the evidence delivery results. The physical consistency verification module is used to verify the consistency of the collision detection results based on the physical response results corresponding to the collision detection results, and generate compensation and correction results. The strategy feedback update module is used to update the decision parameters and candidate collision memory state of the improved R2D2 decision unit by compensating for correction results, collision detection results, and operational performance results. The steady-state performance shaping module is used to smoothly shape the evidence delivery strength and detection kernel call plan in the target game frame.

2. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The physical interaction semantic decoupling module specifically includes: Obtain the state data of each object in the game scene within the current game frame and the preset historical game frames; Calculate the motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance of each object based on the state data; The physical interaction semantic vectors of each object are constructed based on motion continuity, collision sensitivity, interaction value, force propagation characteristics, and event triggering importance. Semantic clustering is performed based on the physical interaction semantic vectors of each object to obtain the corresponding collision semantic clusters.

3. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The inter-frame continuity memory module specifically includes: Obtain object state data, physical interaction semantic vectors, and collision semantic clusters from consecutive game frames, and determine candidate collision object pairs; Construct candidate collision evolution states based on the position and velocity changes of candidate collision objects in consecutive game frames; Record historical false detection correction information based on contact relationship records and position correction records in consecutive game frames; Record historical missed detection correction information based on positional relationship records, contact relationship records, and penetration relationship records in consecutive game frames; Record penetration compensation information based on penetration relationship records and position correction records in consecutive game frames; The temporal memory results of candidate collisions are constructed based on the candidate collision evolution state, historical false detection correction information, historical missed detection correction information, and penetration compensation information.

4. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The collision opportunity window planning module specifically includes: Obtain collision semantic clusters, object historical trajectory data, scene topology, object interaction relationships and temporal memory results from the current game frame and consecutive historical game frames, and construct candidate collision spatiotemporal fragments; Collision prior information is generated based on the object's historical trajectory data, scene topology, object interaction relationships, and time sequence memory results. Calculate the collision value, missed detection risk, and propagation impact of candidate collision spatiotemporal segments based on prior collision information; Candidate collision spatiotemporal windows are generated based on collision value, missed detection risk, and propagation impact. The candidate collision spatiotemporal windows are then filtered through the opportunity window screening gate in the improved R2D2 decision unit to obtain the filtered candidate collision spatiotemporal windows. The candidate collision spatiotemporal windows after screening are state-encoded through the collision evidence memory pathway and the physical consistency memory pathway, and the replay priority is determined by the multi-factor priority replay unit to form a window replay sequence. By using the dual-output decision head in the improved R2D2 decision unit to dynamically plan the window replay sequence, the collision detection benefit value, performance cost value, and comprehensive window decision value are obtained. The candidates are sorted according to their comprehensive window decision values ​​after screening to construct a set of high-value collision opportunity windows.

5. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 4, characterized in that, The process of performing state encoding on the selected candidate collision spatiotemporal windows through collision evidence memory pathways and physical consistency memory pathways, and determining the replay priority through a multi-factor priority replay unit to form a window replay sequence specifically includes: The candidate collision evolution state, historical missed detection correction information and penetration compensation information are obtained, and the collision evidence is encoded through the collision evidence memory path in the improved R2D2 decision unit to obtain the collision evidence encoding result. The physical consistency violation degree and position correction association state corresponding to the candidate collision spatiotemporal window after filtering are obtained, and physical consistency encoding is performed through the physical consistency memory path to obtain the physical consistency encoding result; The collision evidence coding result and the physical consistency coding result are fused to obtain the state coding result; The replay priority of each candidate collision spatiotemporal window after screening is determined by the multi-factor priority replay unit. The playback priorities are sorted, and a window playback sequence is formed according to the sorting results.

6. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The adaptive evidence delivery module specifically includes: Obtain the set of high-value collision opportunity windows, time-series memory results, and detection load information, and determine the target of detection evidence delivery for candidate collisions within the window; Calculate the required detection evidence value for candidate collisions based on the high-value collision opportunity window, temporal memory results, and detection load information. The improved R2D2 decision unit outputs collision detection benefit value, performance cost value, detection evidence requirement value, state coding result and detection evidence placement priority, and calculates the basic placement value of trajectory interpolation depth, basic placement value of continuous detection length, basic placement value of boundary sampling density, basic placement value of historical frame call range and basic placement value of high-precision detection kernel trigger times respectively. The load constraint adjustment of the basic deployment value is performed based on the detection load information to obtain the adjusted deployment value corresponding to the trajectory interpolation depth, continuous detection length, boundary sampling density, historical frame call range, and high-precision detection kernel trigger count. The strength of detection evidence corresponding to candidate collisions is constructed based on the adjusted delivery values, and the evidence delivery results are generated based on the strength of detection evidence.

7. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The differential detection kernel scheduling module specifically includes: Obtain the collision semantic cluster to which the object belongs, the set of high-value collision opportunity windows, the evidence delivery results, and the strength of the detected evidence, and determine the initial candidate collisions; Based on the object interaction relationships, force propagation characteristics, and event triggering importance corresponding to the initial candidate collisions, an extended analysis is performed on the associated contact objects, force transmission objects, and event triggering objects to obtain the set of associated contact objects, the set of force transmission objects, and the set of event triggering objects. The impact of collision propagation is calculated based on the set of related contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the results of evidence deployment. Calculate the kernel matching value corresponding to each collision detection kernel based on the collision propagation impact results and the strength of detection evidence. The collision detection kernels that match the candidate collisions are determined based on the kernel matching values ​​corresponding to each collision detection kernel. The collision detection kernel is invoked to perform collision determination on the initial candidate collisions, the set of associated contact objects, the set of force-transmitting objects, the set of event-triggered objects, and the strength of detection evidence, and the collision detection results are obtained.

8. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The physical consistency verification module specifically includes: Obtain the collision detection results and the corresponding physical response results; Calculate the response velocity coefficient, response displacement coefficient, response direction coefficient, and response duration coefficient based on the physical response results; The physical consistency check value is calculated based on the collision detection results, response speed coefficient, response displacement coefficient, response direction coefficient, and response duration coefficient. Calculate the average physical consistency check value based on the physical consistency check values ​​corresponding to all initial candidate collisions in the current game frame; Compensation and correction results are generated based on the physical consistency check value and the average physical consistency check value.

9. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The strategy feedback update module specifically includes: Obtain the performance results, which include at least the detection time, detection load information, and resource usage. The candidate collision memory state is updated based on the compensation correction result, collision detection result, and time sequence memory result to obtain the updated candidate collision memory state. Calculate the collision detection benefit feedback value and performance cost feedback value based on the compensation correction results, collision detection results, collision propagation impact results and operational performance results; The timing difference error is calculated based on the collision detection benefit feedback value, the performance cost feedback value, and the collision detection benefit value and performance cost value output by the dual-output decision head. The updated empirical sequence replay priority is obtained by updating the empirical sequence replay priority based on the comprehensive timing difference error, historical missed detection correction information, collision propagation level, physical consistency violation degree and detection load fluctuation value. The decision parameters of the improved R2D2 decision unit are updated based on the comprehensive time-series difference error and the empirical sequence playback priority, resulting in the updated decision parameters. The updated candidate collision memory state is fed back to the inter-frame continuous memory module, and the updated experience sequence replay priority and updated decision parameters are fed back to the collision opportunity window planning module and the adaptive delivery module of detection evidence.

10. The game physics collision accuracy detection and performance optimization system based on reinforcement learning according to claim 1, characterized in that, The steady-state performance shaping module specifically includes: Obtain the evidence delivery results, collision detection results, collision propagation impact results, compensation and correction results, and running performance results for each game frame within the statistical game frame interval, and obtain the detection evidence strength and high-precision detection kernel trigger number adjustment delivery value in the target game frame; The execution pressure value is calculated based on the evidence delivery results and running performance results corresponding to each game frame within the statistical game frame interval, and the detection maintenance value is calculated based on the collision detection results, collision propagation impact results, and compensation correction results. The steady-state performance shaping factor is calculated based on the execution pressure value and detection maintenance value in the statistical game frame interval. The strength of detection evidence in the target game frame is smoothed and shaped according to the steady-state performance shaping coefficient to obtain the smoothed detection evidence strength. The detection kernel call plan value of the target game frame is calculated based on the strength of detection evidence after smoothing and the impact of collision propagation in the target game frame; The detection kernel call plan in the target game frame is determined based on the detection kernel call plan value, and the smoothed and shaped detection evidence strength and detection kernel call plan are output to the detection evidence adaptive delivery module and the differential detection kernel scheduling module.

Images