Game behavior processor of embedded AI processing unit

Abstract

The invention discloses a game behavior processor of an embedded AI processing unit, which adopts a heterogeneous computing architecture to realize task unloading and cooperative processing, and comprises a hardware acceleration layer which integrates a CPU (Central Processing Unit), an NPU (Network Processing Unit), a GPU (Graphic Processing Unit) and a computing power resource pool of a special DSP (Digital Signal Processor); the intelligent behavior processing stack comprises an input layer, a memory and emotion layer, a core logic layer and an execution layer which are connected in sequence; the cross-layer security monitoring layer is used for covering a security protection and exception handling mechanism of the whole stack; wherein the hardware acceleration layer provides computing power scheduling support for the intelligent behavior processing stack through a hardware abstract interface. According to the method, revolutionary upgrading of a game embedded system is achieved through heterogeneous computing architecture reconstruction and dynamic energy efficiency scheduling, a closed loop based on emotion computing and a mixed decision architecture are adopted, traditional game AI mechanical limitation is broken through, game behaviors are made to better fit emotion logic, and the sense of reality and immersion of AI roles are enhanced.

Description

Technical Field

[0001] This invention relates to the field of processor technology, and in particular to a game behavior processor for embedded AI processing units. Background Technology

[0002] The gaming industry is currently experiencing unprecedented growth, driven by the relentless pursuit of highly intelligent and deeply immersive experiences. This trend has led to increasingly complex workloads for embedded gaming systems, encompassing everything from realistic graphics rendering and accurate physics simulation to intelligent behavioral decision-making for non-player characters, as well as real-time processing and analysis of various sensor input data. As the core support platform for gaming devices, the performance, efficiency, and intelligence level of embedded systems directly determine the upper limit of the end-user experience, posing a severe challenge to real-time response and parallel processing capabilities.

[0003] However, traditional embedded system architectures have gradually revealed significant shortcomings when faced with these diverse and computationally intensive tasks. Particularly in AI character creation, decision-making mechanisms in past games were often overly simplistic and static, heavily reliant on pre-defined scripts or limited state machines. They lacked the ability to deeply perceive, model, and dynamically respond to complex environments, player behavior, and the character's inner emotional state. Emotional factors—a crucial dimension for enhancing character credibility and player immersion—were often overlooked or simplified in traditional AI design. Simultaneously, in terms of task processing architecture, early development models generally exhibited a "CPU-centric" tendency, relying on a single central processing unit to handle almost all types of computational tasks. This not only led to insufficient CPU resources and high loads but also resulted in overall system inefficiency, increased energy consumption, and severely hampered the implementation of more complex and real-time game mechanics.

[0004] This crude task allocation method and simplistic AI model have become major bottlenecks in enhancing game intelligence and immersion. CPU overload not only limits the complexity and scale of game scenes but also hinders the effective deployment of real-time data processing and advanced AI algorithms. Therefore, there is an urgent need to fundamentally optimize the game embedded system architecture, explore more efficient heterogeneous computing resource collaboration, and develop more advanced AI decision-making frameworks that incorporate emotional computing and context awareness capabilities. This will help distribute computational pressure, improve parallel processing capabilities, and ultimately support the intelligent, natural, and emotionally engaging interactive experiences required by next-generation games. Summary of the Invention

[0006] This invention aims to overcome the core defects of existing embedded game system architectures. By introducing a dedicated hardware acceleration unit to reconstruct the computing paradigm, it solves the real-time bottleneck problem caused by CPU overload in traditional solutions, and thus provides a game behavior processor for embedded AI processing units.

[0007] The game behavior processor of the embedded AI processing unit of the present invention adopts a heterogeneous computing architecture to realize task offloading and collaborative processing, including:

[0008] Hardware acceleration layer: a computing resource pool integrating CPU, NPU, GPU and dedicated DSP; Intelligent behavior processing stack: sequentially connected input layer, memory and emotion layer, core logic layer and execution layer;

[0009] Cross-layer security monitoring layer: providing full-stack security protection and exception handling mechanisms;

[0010] The hardware acceleration layer provides computing power scheduling support for the intelligent behavior processing stack through a hardware abstraction interface.

[0011] This invention achieves three core breakthroughs through heterogeneous computing architecture and hardware abstraction interface design:

[0012] Computing power release: Offload tasks such as graphics rendering and physics simulation from the CPU to dedicated processors (NPU / GPU / DSP), reducing CPU load by 40%-60%;

[0013] Decoupling of architecture: The layered design of the intelligent behavior processing stack allows modules such as emotion computing and decision logic to be upgraded independently, shortening the development cycle by 50%;

[0014] Security Enhancement: Cross-layer monitoring mechanism provides real-time protection against security risks from the hardware layer to the execution layer, increasing the hacker attack interception rate by 90%.

[0015] Optimized, the hardware acceleration layer includes a dynamic task scheduling module and a power consumption collaborative management module. The dynamic task scheduling module allocates computing tasks based on a priority queue optimized by reinforcement learning, monitors the computing power utilization of each unit in real time, and performs cross-processor load balancing. The power consumption collaborative management module uses DVFS technology combined with temperature-power consumption joint modeling to achieve closed-loop power consumption control for multi-processor collaboration through sensor feedback. This invention brings revolutionary performance optimization through dynamic scheduling and power consumption collaboration: it achieves the following: Load balancing efficiency: The scheduling algorithm optimized by reinforcement learning enables multi-processor parallel utilization to reach 92%, and latency is reduced to the 5ms level; Precise power consumption control: The DVFS technology with temperature-power consumption joint modeling reduces peak power consumption by 35% and extends battery life by 40%; Real-time response guarantee: The closed-loop control response speed is <10μs, avoiding frequency reduction and stuttering caused by temperature fluctuations.

[0016] Optimally, the input layer includes:

[0017] Multimodal perception module: A lightweight CNN accelerated by NPU processes visual data, an audio feature extraction pipeline accelerated by DSP processes acoustic data, and a CPU protocol parsing engine processes game state data streams.

[0018] Player intent analysis module: Constructs a probability graph of player action sequences based on a Hidden Markov Model, and dynamically generates player behavior profiles using a spectral clustering algorithm. This invention achieves accurate environmental awareness through multimodal perception and player profiling technologies.

[0019] Perception efficiency: NPU-accelerated CNN enables image processing speeds of up to 120fps, while DSP audio processing latency is <8ms;

[0020] Intent prediction accuracy: The player profiling accuracy of Hidden Markov Model + Spectral Clustering is 92.7%, and the false alarm rate of anti-cheating is reduced to 0.3%;

[0021] Resource optimization: The protocol parsing engine reduces redundant data transmission by 80%.

[0022] Optimally, the memory and emotion layer includes:

[0023] Emotional Memory Bank: Uses a time-series graph database to store emotionally related events, and employs an exponentially decaying weighted model to calculate the weight of the impact of historical events on the current emotional state;

[0024] Emotional State Machine: Based on the OCC emotional cognition model, its input is coupled with the real-time event stream of the perception module and the historical emotional patterns of the emotional memory bank. This invention's emotional computing system overcomes the mechanical limitations of traditional AI, specifically: Emotional Continuity: The exponential decay weighted model ensures that the calculation error of the historical event influence weight is <0.05; Cognitive Realism: The OCC emotional model supports 32 composite emotional states, improving the realism of NPC behavior by 70%; Expanded Decision-Making Dimensions: The emotional memory bank provides cross-scenario behavioral consistency assurance.

[0025] Ideally, the core logic layer includes:

[0026] Hybrid Decision Engine: A lightweight, dual-latency deep deterministic policy gradient model is deployed on the NPU to generate decision suggestions, while a rule-based state validator is run on the CPU to filter illegal decisions; Hierarchical Planner: The high-level objective is decomposed using the HTN planning framework, and AI algorithms and potential field methods are integrated to achieve dynamic path planning;

[0027] Adaptive Optimizer: The decision engine parameters are adjusted in real-time through an online near-end policy optimization algorithm, and the reward function integrates game goal achievement and emotional consistency indicators. The hybrid decision architecture adopted in this invention balances intelligence and reliability, including: Decision Quality: The TD3 model achieves an inference speed of 15μs / frame on the NPU, and the rule engine intercepts 99.6% of illegal decisions; Path Planning Efficiency: The HTN+AI algorithm improves pathfinding speed in complex scenes by 8 times; Adaptive Optimization: The PPO algorithm fine-tunes parameters in real-time, increasing the character goal achievement rate by 45%.

[0028] Optimally, the execution layer includes:

[0029] Motion synthesis module: Generates basic motion sequences based on finite state automata, and uses GPU parallel interpolation calculation to achieve smooth animation transitions;

[0030] Event Response Engine: Employing a publish-subscribe event bus architecture, this invention uses the Rete algorithm to match action rules triggered by multiple conditions. The execution layer innovatively achieves smooth animations and agile responses, including: Animation Smoothness: GPU parallel interpolation reduces action transition latency to <3ms, decreasing stuttering by 90%; Event Response Speed: The Rete algorithm supports millisecond-level matching of 2000+ rules, improving concurrent event processing capabilities by 10 times; Resource Reuse Rate: The state machine reuse library reduces animation resource consumption by 70%.

[0031] Optimally, the cross-layer security monitoring layer includes:

[0032] Behavioral credibility verification module: It uses a dynamic time warping algorithm to detect outliers in behavioral sequences and combines it with a blockchain-based whitelist of interactive credentials for dual verification;

[0033] Fault-tolerant recovery module: Deploys layered watchdog timers to monitor the status of each layer, and uses checkpoint rollback and state machine hot backup to achieve fault recovery.

[0034] The dual security monitoring mechanism of this invention constructs a trusted environment, including: cheating detection rate: the blockchain whitelist enables an abnormal behavior detection rate of 99.2%; system reliability: the layered watchdog achieves 500ms-level fault isolation and hot backup recovery time <1s; data integrity: checkpoint rollback ensures a state loss rate of <0.001%.

[0035] Ideally, the dynamic task scheduling module establishes a processor capability profile library, recording the historical execution efficiency of each processor for graphics rendering, physics simulation, and AI inference tasks;

[0036] The scheduling decision incorporates an energy efficiency ratio weighting factor, calculated using the following formula:

[0037] Scheduling priority = α × task urgency + β × (1 - current load rate) + γ × historical energy efficiency ratio, where α, β, and γ are dynamic adjustment coefficients.

[0038] The energy efficiency ratio scheduling formula used in this invention brings about revolutionary resource management, including: task allocation accuracy: the processor capability profile library improves scheduling matching accuracy by 68%; dynamic adaptability: the real-time adjustment of coefficients α, β, γ responds to environmental changes with a latency of <50ms; energy efficiency optimization: the historical energy efficiency ratio factor reduces ineffective power consumption by 27%.

[0039] Optimized, the emotional state machine generates a three-dimensional emotional vector containing pleasure, excitement, and anxiety; the hybrid decision engine has an emotional fusion interface that performs multimodal fusion processing on the emotional vector and the game state vector to form enhanced decision input features; the decision engine has a built-in emotional consistency evaluation unit that dynamically adjusts the exploration rate of the decision strategy by calculating the cosine similarity between the emotional vector and the preset character personality template; when the emotional vector is detected to deviate from the preset threshold range, an emotional calibration mechanism is triggered, calling the historical emotional patterns from memory and the emotional layer for decision compensation. In this invention, the emotional-decision closed loop creates a revolutionary game experience, including: anthropomorphic behavior: the three-dimensional emotional vector increases the richness of NPC emotional expression by 300%; decision accuracy: the enhanced features of the emotional fusion interface reduce the decision error rate by 42%; character consistency: cosine similarity evaluation + calibration mechanism ensures a 100% correction rate for emotional deviation; exploration efficiency: dynamic adjustment of the exploration rate accelerates the training convergence speed by 5.8 times.

[0040] Optimized, the motion synthesis module integrates a physical simulation correction unit, verifying the physical feasibility of the motion through rigid body dynamics equations; the event response engine sets up an emotion-driven rule channel, prioritizing the triggering of emotion-related motion sequences when the L2 norm of the emotion influence vector exceeds a threshold. The physical-emotional synergy in this invention achieves ultimate realism, including: motion rationality: rigid body dynamics verification eliminates 98% of physical anomalies such as clipping and suspension; emotional expressiveness: the emotion-driven channel improves the response speed of high-emotion-intensity events by 90%; resource optimization: the L2 norm threshold triggering mechanism reduces invalid animation calculations by 60%.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. This invention achieves a revolutionary upgrade of embedded game systems through heterogeneous computing architecture reconstruction and dynamic energy efficiency scheduling: the hardware acceleration layer reduces CPU load by 40%-60%, and the parallel utilization rate of the dedicated processor reaches 92%; the reinforcement learning-optimized task scheduling algorithm compresses latency to the 5ms level, and the DVFS closed-loop power control reduces peak power consumption by 35%; the multimodal perception module achieves 120fps image processing and 8ms audio response, providing a computing power foundation for high real-time game scenarios.

[0043] 2. This invention adopts a closed-loop and hybrid decision-making architecture based on emotion computing, which breaks through the mechanical limitations of traditional game AI: the OCC emotion model supports 32 composite emotion states, and the three-dimensional emotion vector improves the realism of NPC behavior by 70%; the emotion fusion interface and cosine similarity evaluation construct a decision correction mechanism, ensuring 100% consistency of character emotions; the online PPO optimization algorithm improves the speed of decision adaptation by 5.8 times, and the HTN planner improves the efficiency of complex path search by 8 times.

[0044] 3. This invention also innovatively integrates physical-emotional execution with cross-layer security protection: GPU-accelerated animation interpolation eliminates 90% of motion stuttering, rigid body dynamics verification solves 98% of physical anomalies; the emotion-driven channel improves the response speed of high-emotion intensity events by 90%; the blockchain whitelist achieves a 99.2% cheating detection rate, and the layered watchdog mechanism achieves 500ms-level fault isolation. Attached Figure Description

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0046] Figure 1 This is an overall flowchart of the game behavior processor of the embedded AI processing unit of the present invention;

[0047] Figure 2 This is a flowchart illustrating the memory and emotion layer in the game behavior processor of the embedded AI processing unit of the present invention.

[0048] Figure 3 This is a flowchart of the dynamic task scheduling module in the game behavior processor of the embedded AI processing unit of the present invention. Detailed Implementation

[0049] This section will describe in detail specific embodiments of the present invention. Preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the drawings is to supplement the textual description with graphics, so that people can intuitively and vividly understand each technical feature and overall technical solution of the present invention, but they should not be construed as limiting the scope of protection of the present invention.

[0050] Reference Figure 1 , Figure 2 and Figure 3 The present invention discloses a game behavior processor for an embedded AI processing unit, which adopts a heterogeneous computing architecture to realize task offloading and collaborative processing, including:

[0051] Hardware acceleration layer: This layer integrates a computing resource pool comprising CPU, NPU, GPU, and dedicated DSP. It includes a dynamic task scheduling module and a power consumption collaborative management module. The dynamic task scheduling module allocates computing tasks based on a priority queue optimized by reinforcement learning, monitors the computing power utilization of each unit in real time, and performs cross-processor load balancing. The power consumption collaborative management module uses DVFS technology combined with temperature-power consumption joint modeling, and achieves closed-loop power consumption control for multi-processor collaboration through sensor feedback. The dynamic task scheduling module establishes a processor capability profile library, recording the historical execution efficiency of each processor for graphics rendering, physics simulation, and AI inference tasks.

[0052] The scheduling decision incorporates an energy efficiency ratio weighting factor, calculated using the following formula:

[0053] Scheduling priority = α × task urgency + β × (1 - current load rate) + γ × historical energy efficiency ratio, where α, β, and γ are dynamic adjustment coefficients.

[0054] The intelligent behavior processing stack consists of an input layer, a memory and emotion layer, a core logic layer, and an execution layer, connected sequentially. The input layer includes: a multimodal perception module (using a lightweight CNN accelerated by an NPU to process visual data, a DSP-accelerated audio feature extraction pipeline to process acoustic data, and a CPU protocol parsing engine to process game state data streams); and a player intent analysis module (constructing a probability graph of player operation sequences based on a Hidden Markov Model, dynamically generating player behavior profiles using spectral clustering algorithms). The memory and emotion layer includes: an emotion memory bank (using a time-series graph database to store emotion-related events, employing an exponentially decaying weighted model to calculate the influence weights of historical events on the current emotion state); and an emotion state machine (built based on the OCC emotion cognition model, whose input is coupled to the real-time event stream of the perception module and the historical emotion patterns in the emotion memory bank). The core logic layer includes: a hybrid decision engine: a lightweight, dual-latency deep deterministic policy gradient model is deployed on the NPU to generate decision suggestions, while a rule-based state validator is run on the CPU to filter illegal decisions; a hierarchical planner: a high-level goal is decomposed using the HTN planning framework, and AI algorithms and potential field methods are integrated to achieve dynamic path planning; and an adaptive optimizer: the decision engine parameters are adjusted in real time through an online proximal policy optimization algorithm, and the reward function integrates game goal achievement and emotional consistency indicators. The execution layer includes: an action synthesis module: basic action sequences are generated based on finite state automata, and smooth animation transitions are achieved using GPU parallel interpolation calculations; and an event response engine: an event bus with a publish-subscribe architecture is used, and multi-condition triggered action rule matching is achieved through the Rete algorithm. The action synthesis module integrates a physics simulation correction unit, verifying the physical feasibility of actions through rigid body dynamics equations; the event response engine sets up an emotion-driven rule channel, prioritizing the triggering of emotion-related action sequences when the L2 norm of the emotion influence vector exceeds a threshold.

[0055] The emotional state machine generates a three-dimensional emotional vector containing pleasure, excitement, and anxiety. The hybrid decision engine has an emotional fusion interface that performs multimodal fusion processing on the emotional vector and the game state vector to form enhanced decision input features. The decision engine has a built-in emotional consistency evaluation unit that dynamically adjusts the exploration rate of the decision strategy by calculating the cosine similarity between the emotional vector and the preset character personality template. When the emotional vector is detected to deviate from the preset threshold range, an emotional calibration mechanism is triggered, which calls the historical emotional patterns from memory and the emotional layer for decision compensation.

[0056] Cross-layer security monitoring layer: covering the entire stack of security protection and anomaly handling mechanisms; the cross-layer security monitoring layer includes: behavior credibility verification module: using dynamic time warping algorithm to detect outliers in behavior sequences, combined with a blockchain-based interactive credential whitelist for dual verification; fault tolerance and recovery module: deploying layered watchdog timers to monitor the status of each layer, and using checkpoint rollback and state machine hot backup to achieve fault recovery.

[0057] The hardware acceleration layer provides computing power scheduling support for the intelligent behavior processing stack through a hardware abstraction interface.

[0058] Example 1:

[0059] At the beginning, when new data is input, it is processed according to its data type:

[0060] If the data is environmental (such as game scene, NPC status, etc.), it enters the multimodal perception module for parsing and extracts key environmental information (terrain, enemy distribution, etc.); if it is player input (operation commands, dialogue content, etc.), it enters the player intent analysis module for interpretation and identification of player intent (attack, task triggering, etc.). The processing results of the multimodal perception module and the player intent analysis module are jointly input into the core logic layer to generate basic action strategies (such as "approach the enemy → attack", "respond to player dialogue → deliver the task").

[0061] After making a decision, determine whether an emotional response is needed. If an emotional response is needed, call the memory and emotion layers, and combine the time sequence database in the memory and emotion layers (such as events that made the character angry or happy in the past) to make emotional adjustments to the basic strategy (e.g., because the character was "deceived" by the player in the past, the response in this dialogue is wary). After the adjustment, the memory and emotion layers are updated synchronously to record this emotional interaction. If no emotional response is needed, proceed directly to the subsequent task decomposition process.

[0062] Regardless of whether the strategy is adjusted emotionally, it enters the hierarchical planner and is broken down into specific sub-tasks (such as "move to coordinate X → release skill Y"). The sub-tasks are passed to the adaptive optimizer, which optimizes the execution parameters (such as adjusting the timing of skill release and movement path planning) based on the real-time state of the game (network latency, hardware performance). After parameter tuning, the execution layer outputs the actual game actions (character movement, skill effect display, etc.).

[0063] After the action is executed, it is determined whether a special event (such as hidden plot or achievement conditions) is triggered. When an event is triggered, the event response engine responds and executes the event-specific logic (unlocking a new map or playing a plot animation). When no event is triggered, the event response is skipped and the system directly enters the cross-layer security monitoring layer.

[0064] The cross-layer security monitoring layer checks the compliance of the process (such as whether the data has been tampered with or whether the operation has triggered cheat features). When an anomaly is detected, the fault tolerance and recovery module intervenes to perform repair / interception (such as rolling back illegal operations or banning abnormal accounts). If there is no anomaly, the process ends normally and is finally marked as "task completed". It returns to the initial state to wait for the next round of new data input and repeats the above process.

[0065] Example 2:

[0066] Based on the current emotional state (such as the NPC's real-time emotional value or the emotional feedback triggered by player interaction), as the basic input of the emotional state machine, the current emotional state enters the judgment node to determine whether it has reached the memory threshold. If it has: the emotional event has the intensity of being "worth remembering" (such as a player's malicious attack triggering a high anger value), and enters the emotional event storage stage, writing the event details (emotion type, triggering conditions, intensity value, etc.) into the emotional memory bank (a database / module that stores the NPC's emotional history for a long time). If it has not reached: the emotional intensity is insufficient (such as a weak emotion generated by a slight interaction), triggering the ignore branch, not recording the event, and the process is not associated with emotional memory for the time being.

[0067] When subsequent execution layers are needed (such as when an NPC needs to determine "how to respond to the player" or "next action strategy"), the decision-making node is triggered and invoked. When invoked: historical emotional events are extracted from the emotional memory bank, and historical emotional weights are calculated (such as calculating the weighted impact coefficient on the current decision based on the event's occurrence time and emotional intensity), ultimately affecting the output of the hybrid decision engine (such as prioritizing the "counterattack" strategy when deciding based on the memory of anger caused by a historical attack). When not invoked: the default strategy is executed directly (such as when there is no emotional connection, the NPC executes the preset general behavioral logic). Regardless of whether emotional memories are invoked, the decision output returns to the main game flow (such as performing actions or triggering new interactions), forming a complete logical chain of "emotion generation → memory filtering → decision invocation," making AI behavior more consistent with emotional history and enhancing the immersion and consistency of characters in the game.

[0068] Example 3:

[0069] The system detects a new task and enters the task processing phase. It categorizes the new task and matches hardware resources based on its type. For tasks involving logic judgment or rule execution (such as simple logical decisions in a game), the CPU is allocated, leveraging its general computing and logic control advantages. For tasks requiring extensive parallel computation (such as game scene rendering or parallel processing of multiple data), the GPU is allocated, utilizing its parallel rendering and computing capabilities. For tasks with high real-time requirements (such as audio signal processing or low-latency interactive response), a DSP (Digital Signal Processor) is allocated to ensure real-time performance. For computationally intensive tasks (such as complex AI inference), an NPU (Neural Processing Unit) is allocated to accelerate the process using its parallel computing capabilities. If no task matches any of the above categories, the CPU is allocated by default.

[0070] After completing the task-hardware allocation, check the current load and running status of the hardware. Based on the hardware status, determine if the load is too high. If the load is too high, trigger the dynamic migration task to reallocate the task to an idle hardware unit to balance the load. If the hardware load is normal, proceed to task execution. The hardware unit executes the task according to the allocation result, handling calculations, logic, and other operations. Once the task is completed, the process ends and waits for the next new task.

[0071] This invention constructs a three-tier architecture based on a hardware acceleration layer, an intelligent behavior processing stack, and a cross-layer security monitoring layer. The hardware acceleration layer integrates CPU, NPU, GPU, and dedicated DSP to form a computing resource pool. A dynamic task scheduling module achieves cross-processor load balancing, coupled with DVFS technology and closed-loop power control using temperature-power modeling. The intelligent behavior processing stack includes:

[0072] Input layer: NPU-accelerated CNN processing of visual data (120fps) + DSP audio feature extraction (latency <8ms) + HMM player intent modeling;

[0073] Memory and Emotion Layer: The time sequence database stores emotional events, and the OCC emotional state machine generates three-dimensional emotional vectors (pleasure / excitement / anxiety).

[0074] Core logic layer: NPU runs a lightweight TD3 decision model + CPU rule engine, HTN planner integrates AI and potential field method, PPO algorithm optimizes parameters in real time;

[0075] Execution layer: GPU parallel interpolation is used to implement animation transitions (latency <3ms), and the Rete algorithm supports millisecond-level matching of 2000+ rules.

[0076] This invention also includes an emotional decision-making closed loop and security collaboration, comprising:

[0077] Deep coupling of emotion and decision: Emotion vectors are used to generate enhanced decision features through a multimodal fusion interface. The deviation between the emotion vector and the character personality template is calculated by cosine similarity, and the decision exploration rate is dynamically adjusted. When the emotion deviates from the threshold, a calibration mechanism is triggered to call historical data from the emotion memory bank for compensation.

[0078] Physics-emotion execution linkage: The action synthesis module verifies physical feasibility through rigid body dynamics equations (eliminating 98% clipping), and the event response engine sets up an emotion-driven channel, prioritizing the triggering of emotional actions when the L2 norm of the emotion vector exceeds the threshold.

[0079] Cross-layer security protection: DTW algorithm + blockchain whitelist achieves a 99.2% cheating detection rate, and layered watchdog timer + state machine hot backup achieves 500ms-level fault recovery.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions made to the technical solutions of the present invention cannot depart from the spirit and scope of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. A game behavior processor for an embedded AI processing unit, characterized in that, A heterogeneous computing architecture is used to implement task offloading and collaborative processing, including: Hardware acceleration layer: A pool of computing resources integrating CPU, NPU, GPU and dedicated DSP; Intelligent behavior processing stack: sequentially connected input layer, memory and emotion layer, core logic layer, and execution layer; Cross-layer security monitoring layer: providing full-stack security protection and exception handling mechanisms; The hardware acceleration layer provides computing power scheduling support for the intelligent behavior processing stack through a hardware abstraction interface.

2. The game behavior processor of the embedded AI processing unit according to claim 1, characterized in that, The hardware acceleration layer includes a dynamic task scheduling module and a power consumption collaborative management module. The dynamic task scheduling module allocates computing tasks based on a priority queue optimized by reinforcement learning, monitors the computing power utilization of each unit in real time, and performs cross-processor load balancing. The power consumption collaborative management module adopts DVFS technology combined with temperature-power consumption joint modeling and realizes closed-loop power consumption control of multi-processor collaboration through sensor feedback.

3. The game behavior processor of the embedded AI processing unit according to claim 1, characterized in that, The input layer includes: Multimodal perception module: A lightweight CNN accelerated by NPU processes visual data, an audio feature extraction pipeline accelerated by DSP processes acoustic data, and a CPU protocol parsing engine processes game state data streams. Player Intent Analysis Module: Constructs a probability graph of player action sequences based on a Hidden Markov Model, and dynamically generates player behavior profiles by combining spectral clustering algorithms.

4. The game behavior processor of the embedded AI processing unit according to claim 1, characterized in that, The memory and emotion layer includes: Emotional Memory Bank: Uses a time-series graph database to store emotionally related events, and employs an exponentially decaying weighted model to calculate the weight of the impact of historical events on the current emotional state; Emotional State Machine: Based on the OCC emotional cognition model, its input is coupled with the real-time event flow of the perception module and the historical emotional patterns of the emotional memory bank.

5. The game behavior processor of the embedded AI processing unit according to claim 1, characterized in that, The core logic layer includes: Hybrid decision engine: A lightweight, dual-latency deep deterministic policy gradient model is deployed on the NPU to generate decision recommendations, while a rule-based state validator is run on the CPU to filter illegal decisions; Hierarchical planner: It adopts the HTN planning framework to decompose high-level objectives and integrates AI algorithms and potential field methods to achieve dynamic path planning; Adaptive optimizer: Adjusts decision engine parameters in real time through online proximal policy optimization algorithm, and the reward function integrates game goal achievement and emotional consistency indicators.

6. The game behavior processor of the embedded AI processing unit according to claim 1, characterized in that, The execution layer includes: Motion synthesis module: Generates basic motion sequences based on finite state automata, and uses GPU parallel interpolation calculation to achieve smooth animation transitions; Event response engine: It adopts an event bus with a publish-subscribe architecture and uses the Rete algorithm to match action rules triggered by multiple conditions.

7. The game behavior processor of the embedded AI processing unit according to claim 1, characterized in that, The cross-layer security monitoring layer includes: Behavioral credibility verification module: It uses a dynamic time warping algorithm to detect outliers in behavioral sequences and combines it with a blockchain-based whitelist of interactive credentials for dual verification; Fault-tolerant recovery module: Deploys layered watchdog timers to monitor the status of each layer, and uses checkpoint rollback and state machine hot backup to achieve fault recovery.

8. The game behavior processor of the embedded AI processing unit according to claim 2, characterized in that, The dynamic task scheduling module establishes a processor capability profile library and records the historical execution efficiency of each processor for graphics rendering, physics simulation, and AI inference tasks. The scheduling decision incorporates an energy efficiency ratio weighting factor, calculated using the following formula: Scheduling priority = α × task urgency + β × (1 - current load rate) + γ × historical energy efficiency ratio, where α, β, and γ are dynamic adjustment coefficients.

9. The game behavior processor of the embedded AI processing unit according to claims 4 and 5, characterized in that, The emotional state machine generates a three-dimensional emotional vector containing pleasure, excitement, and anxiety; the hybrid decision engine has an emotional fusion interface that performs multimodal fusion processing on the emotional vector and the game state vector to form enhanced decision input features. The decision engine has a built-in emotion consistency assessment unit that dynamically adjusts the exploration rate of the decision strategy by calculating the cosine similarity between the emotion vector and the preset character personality template. When the emotion vector is detected to deviate from the preset threshold range, the emotion calibration mechanism is triggered, which calls the historical emotion patterns of memory and emotion layer for decision compensation.

10. The game behavior processor of the embedded AI processing unit according to claim 6, Its features are, The motion synthesis module integrates a physical simulation correction unit, which verifies the physical feasibility of the motion through rigid body dynamics equations; the event response engine sets up an emotion-driven rule channel, which prioritizes triggering emotion-related action sequences when the L2 norm of the emotion influence vector exceeds a threshold.

Images