Self-adaptive agent decision framework training method with dynamic environment perception capability

Abstract

The invention relates to a self-adaptive agent decision framework training method with dynamic environment perception capability, which comprises the following steps of: 1, collecting multi-modal sensor data in parallel, and carrying out hardware-level preprocessing on the collected data; step 2, based on the data preprocessed in the step 1, performing dynamic security boundary driven real-time feature extraction to obtain security boundary coordinates; 3, based on the safety boundary coordinates obtained in the step 2, generating an execution action strategy in a decision period; and 4, based on the action strategy generated in the step 3, constructing a feedback closed loop of the distributed execution framework, and realizing real-time closed loop optimization. According to the method, the core requirements of scenes such as mobile robot navigation and intelligent industrial control on autonomous decision-making safety and real-time performance are met.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent decision-making and dynamic environment perception technology, and relates to an adaptive intelligent agent decision-making framework training method, especially an adaptive intelligent agent decision-making framework training method with dynamic environment perception capability. Background Technology

[0002] Traditional intelligent agent decision-making frameworks in dynamic environments typically employ a path of "software-level data processing - heavy-duty model feature extraction - static decision-making strategy": In the data acquisition phase, the synchronization of multimodal sensor data relies on software timestamp alignment, and timing deviations are eliminated through operating system-level clock calibration, with synchronization accuracy often exceeding milliseconds. In the feature extraction phase, un-lightweighted deep learning models (such as the original FPN or the unpruned YOLO series) are often used to process the data through general-purpose computing units (such as GPUs). Although rich features can be extracted, the model parameters are large and the computational complexity is high. In the decision-making phase, reinforcement learning strategies often use fixed reward functions and static time windows, and the decision logic is generated based on a pre-set environmental model, lacking a dynamic response mechanism to real-time environmental changes.

[0003] Traditional technologies rely on software-level sensor synchronization, which makes it difficult to eliminate the temporal deviation of multi-source data, resulting in perceptual misalignment of fused data and easy to cause misjudgment of obstacle positions. System feature extraction relies on heavy models and general computing units, resulting in high feature extraction latency, which cannot meet the real-time decision-making requirements of dynamic environments. At the same time, static decision-making strategies are difficult to adapt to sudden environmental changes. When obstacles suddenly appear or path delays fluctuate, the lag in strategy updates can easily lead to safety risks.

[0004] A search revealed that two existing patent documents in the prior art have significant limitations: Although the intelligent adaptive calibration method and system provided by CN202510983413.4 achieves time alignment between asynchronous event streams and RGB images through spatiotemporal clustering and optimizes calibration parameters by outputting static masks from dynamic perception networks, thereby improving the robustness of multimodal vision systems, it only focuses on the multimodal data calibration stage and does not involve hardware-level accelerated preprocessing, data compression processing, or connect to subsequent feature extraction and decision-making modules. Therefore, it cannot provide end-to-end low-latency, high-quality data support for intelligent agent decision-making. Although the state-dynamic perception-based multi-agent cooperative learning method disclosed in CN201910162280.9 achieves multi-agent cooperative training by processing the agent state set through a dynamic perception layer and generating decision actions through a Q-value network with communication units, it only optimizes the software-level decision logic for multi-agent cooperation. It lacks real-time risk modeling in dynamic environments, low-latency processing of hardware and software collaboration, and high-speed feedback closed loop, making it difficult to meet the rapid response requirements under sudden environmental changes. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing an adaptive intelligent agent decision-making framework training method with dynamic environment perception capabilities. This method can solve problems such as large temporal deviations in multimodal data, insufficient real-time feature extraction and decision-making, and delayed environmental adaptation feedback faced by intelligent agents in dynamic environments, thereby meeting the core requirements of autonomous decision-making safety and real-time performance in scenarios such as mobile robot navigation and intelligent industrial control.

[0006] The above-mentioned objective of this invention is achieved through the following technical solution: A training method for an adaptive agent decision-making framework with dynamic environment perception capabilities includes the following steps: Step 1: Parallel acquisition of multimodal sensor data, and hardware-level preprocessing of the acquired data; Step 2: Based on the data preprocessed in Step 1, perform real-time feature extraction driven by dynamic security boundary to obtain the security boundary coordinates; Step 3: Based on the safety boundary coordinates obtained in Step 2, generate an action strategy to be executed within the decision cycle; Step 4: Based on the action strategy generated in Step 3, construct the feedback loop of the distributed execution framework to achieve real-time closed-loop optimization; Furthermore, the specific steps of step 1 include: Step 1.1: Simultaneously capture multi-source sensor data, including environmental point cloud, RGB-D image, and motion state data, using LiDAR, vision sensor, and inertial measurement unit.

[0007] Step 1.2: Perform ASIC hardware-level spatiotemporal alignment and compression on the collected data; Furthermore, the specific method of step 1.2 is as follows: The first stage performs timestamp calibration of sensor data, achieving nanosecond-level synchronization via the IEEE 1588v2 precision clock protocol. Its logic is based on a local timestamp correction function. ,in For the first Initial sampling time of each sensor, and These are the sending and receiving times of the node's message, respectively. The maximum error in cross-modal data timestamp alignment achieved through this correction is no more than 0.3ms. The second-level deployment can be configured with the convolutional engine CCE, and the point cloud path adopts... The sparse convolutional kernel structure supports dynamic voxel grid downsampling rates. Compress point cloud data to its original size And preserve edge features; image path adopts and Depth-separable convolutions with variable kernels support lossless image resizing; the kernel recombination rule adjusts weight coefficients based on the current sensor data sampling period and channel latency differences, and a cross-scheduling window for convolution weights is set. Dynamic interpolation and fusion of multimodal sensing feature maps ensure that the time difference of the final output fused sensing stream is controlled within 0.5ms; The third-level integrated residual compensation unit automatically activates the lightweight GAN module to perform high-frequency texture reconstruction when motion blur or sudden changes in illumination are detected, ensuring the signal-to-noise ratio of the output perceptual stream. The overall pipeline execution cycle is fixed at 5ms.

[0008] Furthermore, the specific steps of step 2 include: Step 2.1: Based on the preprocessed data from Step 1, perform lightweight STFPN dual-branch feature extraction; Step 2.2: First, fuse the outputs of the two branches to form a dynamic semantic field. Then, use this semantic field as input to construct a risk probability model based on Lyapunov stability theory to generate the safe passage domain boundary. At the same time, construct a risk distribution function based on trajectory aggregation entropy as a supplement. Finally, output the set of safe boundaries. Furthermore, the specific method of step 2.1 is as follows: Based on the preprocessed data from step 1, a lightweight spatiotemporal feature pyramid network (STFPN) is deployed on the FPGA programmable logic unit. The first branch uses a 3D sparse convolutional network to perform sparse spatial sampling and residual encoding on the compressed point cloud, extracting the instantaneous displacement vector field of the obstacle in the local reference coordinate system. The output format is a three-channel tensor. The second branch loads the cropped and optimized YOLOv5 semantic segmentation structure, fuses RGB-D image information, and outputs a segmentation mask tensor. .

[0009] Furthermore, the specific steps of step 2.2 include: (1) Fuse the outputs of the two branches to form a dynamic semantic field .

[0010] (2) The dynamic safety boundary calculation module uses Using Lyapunov stability theory as input, a risk probability model is constructed, the core mathematical expression of which is: ; in For spatial coordinates, As a window for future prediction, This represents the semantic gradient change at the current point. The system status refresh cycle, As the semantic perturbation stabilization factor, it is set to This model is used to dynamically generate agent-centric safe passage domain boundaries, only when... Only then are they included in the low-risk zone for strategy matching.

[0011] (3) Construct a risk distribution function based on trajectory aggregation entropy as a supplement: , in, For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, This is the normalization coefficient.

[0012] The final output is a set of safe boundaries. The boundary update frequency is 50Hz.

[0013] Furthermore, the specific steps of step 3 include: Step 3.1: Generate a latency-sensitive strategy based on the security boundary coordinates obtained in Step 2; Safety boundary coordinates The agent's target path is input into an asynchronous dominant actor-commentator A3C network, which generates action policies during the decision-making cycle.

[0014] Step 3.2: Introduce a time-delay-sensitive reward function and a dynamic decision-making time window; Furthermore, the specific method of step 3.2 is as follows: If the action decision time is delayed Exceeding the environmental status update cycle 50% of Then apply a reward decay coefficient. Simultaneously set a dynamic decision-making time window. .when At that time, the reward function is modified to: ,in As a delay penalty factor, By setting the network to prioritize strategies with low computational complexity, the agent is strengthened to prioritize actions from the low-latency strategy library in unexpected scenarios.

[0015] Furthermore, the specific steps of step 4 include: Step 4.1: Based on the action strategy generated in Step 3, perform real-time monitoring and capture of environmental disturbances. While the intelligent agent is performing actions, it collects data on changes in environmental conditions, and monitors the transmission delay fluctuations of the core monitoring path and the changes in the status of new obstacles. Step 4.2: Inject disturbance assessment and priority scheduling to achieve real-time closed-loop optimization; Furthermore, the specific steps of step 4.2 include: (1) When INT detects path delay fluctuations Or, when a sudden obstacle appears, the HDC high-latency capture mechanism is triggered to copy the current critical network packets; (2) The system constructs a disturbance cognition priority mapping table to achieve real-time closed-loop optimization: the disturbance intensity is evaluated based on the feedback selection function. ,in Indicates the first The current latency change of each path node, For its reference delay, This is the latest velocity vector of the node. As the reference vector, For marking obstacle areas, This is a disturbance-weighted index. It is a stability factor. Simultaneously, a perturbation level scoring function is defined. Perform priority sorting, where To disrupt historical influence. Represents the rate of change of policy entropy. These are the weighting coefficients.

[0016] The system is based on the score As a result, the high-priority perturbation is in An internal STFPN input layer is injected to update the perceptual flow and security boundary, enabling real-time closed-loop optimization.

[0017] The advantages and beneficial effects of this invention are as follows: 1. This invention improves the safety and adaptability of decision-making in dynamic environments through feature extraction and decision optimization techniques. In step 2, the STPPN dual-branch feature extraction channel, 3D sparse convolution processing of point clouds, and cropped YOLOv5 parsing images generate obstacle motion vector fields and semantic segmentation masks in parallel. Combined with a risk probability model based on Lyapunov stability theory, it can output the safe passage domain boundary centered on the agent in real time, ensuring that the decision is always constrained by the environmental risk quantification result. In step 3, the latency-sensitive reward function and dynamic decision time window further force the network to prioritize low-latency strategies. In sudden scenarios, reward correction is used to enhance rapid response, so that the decision not only meets the safety boundary constraints but also adapts to dynamic environmental changes.

[0018] 2. This invention constructs an efficient dynamic environment adaptation mechanism. In step 4, the in-band network telemetry (INT) technology monitors path delay fluctuations and sudden obstacles in real time through packet-level timestamps and path markers. When a delay fluctuation of ≥15% is detected, the HDC mechanism can inject environmental change characteristics into the STPPN input layer within ≤1ms. At the same time, the perturbation cognition priority mapping table calculates the score Si based on the historical impact μimpact of the perturbation and the policy entropy change rate ΔH_i, realizing intelligent scheduling of feedback resources. High-priority perturbations are allocated to the STPPN high-frequency injection channel, while low-priority perturbations are processed with delay, avoiding response lag caused by resource competition.

[0019] 3. In step 1 of this invention, the synchronous acquisition of LiDAR, visual sensors, and inertial measurement units achieves multi-dimensional coverage of environmental information. The three-stage pipeline architecture of the ASIC hardware accelerator, combined with convolutional kernel recombination technology, reduces the timing deviation of multi-source sensors to timestamp error ≤0.5ms, solving the problem of perception distortion caused by timing misalignment in traditional multimodal data. At the same time, voxel mesh compression simplifies point cloud data to 30%~35% of its original size while retaining edge features. Combined with the hardware acceleration capability of ASIC, it reduces the computational load while ensuring the integrity of key environmental features, making the signal-to-noise ratio of the fused perception stream ≥35dB, providing high-quality input for subsequent feature extraction and decision-making. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the impact of different types of perturbations on the entropy value of the policy network according to the present invention; Figure 2 This is a flowchart of the disturbance level scoring function calculation method of the present invention; Figure 3 This is a time-series diagram showing the disturbance from occurrence to scoring in this invention; Figure 4 This is the disturbance priority mapping and resource allocation diagram of the present invention. Detailed Implementation

[0021] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings: A training method for an adaptive agent decision-making framework with dynamic environment perception capabilities includes the following steps: Step 1: Parallel acquisition of multimodal sensor data, and hardware-level preprocessing of the acquired data; The specific steps of step 1 include: Step 1.1: Simultaneously capture multi-source sensor data, including environmental point cloud, RGB-D image, and motion state data, using LiDAR, vision sensor, and inertial measurement unit.

[0022] Step 1.2: Perform ASIC hardware-level spatiotemporal alignment and compression on the collected data. The specific method for step 1.2 is as follows: The first stage performs timestamp calibration of sensor data, achieving nanosecond-level synchronization via the IEEE 1588v2 precision clock protocol. Its logic is based on a local timestamp correction function. ,in For the first Initial sampling time of each sensor, and These are the sending and receiving times of the node's message, respectively. The maximum error in cross-modal data timestamp alignment achieved through this correction is no more than 0.3ms. The second-level deployment can be configured with the convolutional engine CCE, and the point cloud path adopts... The sparse convolutional kernel structure supports dynamic voxel grid downsampling rates. Compress point cloud data to its original size And preserve edge features; image path adopts and Depth-separable convolutions with variable kernels support lossless image resizing; the kernel recombination rule adjusts weight coefficients based on the current sensor data sampling period and channel latency differences, and a cross-scheduling window for convolution weights is set. Dynamic interpolation and fusion of multimodal sensing feature maps ensure that the time difference of the final output fused sensing stream is controlled within 0.5ms; The third-level integrated residual compensation unit automatically activates the lightweight GAN module to perform high-frequency texture reconstruction when motion blur or sudden changes in illumination are detected, ensuring the signal-to-noise ratio of the output perceptual stream. The overall pipeline execution cycle is fixed at 5ms.

[0023] The working principle of step 1 is as follows: Environmental point cloud, RGB-D image, and motion state data are simultaneously captured by LiDAR, vision sensors, and inertial measurement units. An ASIC hardware accelerator is used to perform spatiotemporal alignment of the raw data, and convolutional kernel recombination technology is employed to eliminate temporal biases from multiple sensors, generating a fused sensing stream with a timestamp error ≤0.5ms. The aim is to eliminate temporal biases caused by differences in sampling mechanisms among multimodal sensors such as LiDAR, vision sensors, and inertial measurement units, ensuring strict temporal alignment of environmental point cloud, RGB-D image, and motion state data. The point cloud data is compressed to 30%~30% of its original size using voxel mesh compression. While preserving edge features, the ASIC hardware accelerator significantly reduces data volume to improve processing efficiency and avoids the loss of key environmental features. It adopts a three-stage pipeline architecture. The first stage performs timestamp calibration of sensor data and achieves nanosecond-level synchronization through the IEEE 1588v2 precision clock protocol. The second stage deploys a configurable convolutional engine (CCE) that supports dynamic switching of point cloud voxelization parameters and image downsampling rate. The third stage integrates a residual compensation unit. When motion blur or sudden changes in illumination are detected, a generative adversarial network (GAN) is automatically activated to repair the data, ensuring that the signal-to-noise ratio of the output sensor stream is ≥35dB. For the time allocation of each processing stage in the three-stage pipeline architecture, the key parameters of the convolution kernel reassembly technique, and the elimination of timing deviations, the overall pipeline execution cycle is fixed at 5ms. The first-stage timestamp calibration module occupies 1ms and is mainly responsible for the high-precision synchronization of LiDAR, RGB-D vision, and IMU data. By integrating the IEEE1588v2 precision clock protocol, the sensor synchronization error is controlled at the nanosecond level. Its logic is based on the local timestamp correction function.

[0024] The corrected sampling time of the i-th sensor. For the first Initial sampling time of each sensor, and These represent the sending and receiving times of the message at that node.

[0025] This correction ensures that the maximum error in cross-modal data timestamp alignment does not exceed 0.3ms. The second-level CCE module is allocated 2.5ms of execution time and deploys configurable convolutional kernel reassembly units. The point cloud path uses a 5×5×5 sparse convolutional kernel structure, supporting dynamic voxel grid downsampling rates {1 / 2, 1 / 4, 1 / 8}. The image path uses depthwise separable convolutions with 3×3 and 5×5 variable kernels to support lossless switching of image sizes from the original 640×480 to the compressed 160×120. The convolutional kernel reassembly rule adjusts the weight coefficients based on the current sensor data sampling period and channel delay differences, and a cross-scheduling window for convolutional weights is set. Dynamic interpolation and fusion of multimodal sensing feature maps ensures that the time difference of the final output fused sensing stream is controlled within 0.5ms. The third-level residual compensation unit occupies 1.5ms, and its key function is to quickly determine whether image blurring or abnormal lighting events exceed the reconstruction threshold. A decrease of more than 20% or a standard deviation in brightness distribution In the event of abnormal offsets, a lightweight GAN module is automatically activated for high-frequency texture reconstruction, with the reconstruction time controlled within 1ms, ultimately ensuring that the signal-to-noise ratio of the fused perceptual stream reaches or exceeds 35dB. Through the above architecture design, this ASIC module not only achieves accurate alignment and feature preservation of multimodal information, but also provides highly consistent and high-quality data input for subsequent STPPN feature extraction.

[0026] Step 1.1: Synchronous data acquisition from multiple sensor sources The system simultaneously captures environmental point clouds, RGB-D images, and motion state data using lidar, vision sensors, and inertial measurement units.

[0027] Step 1.2: ASIC Hardware-Level Spatiotemporal Alignment and Compression The raw data is processed using an ASIC hardware accelerator.

[0028] Its core working principle is achieved through a three-stage pipeline architecture.

[0029] The first stage performs timestamp calibration of sensor data, achieving nanosecond-level synchronization via the IEEE 1588v2 precision clock protocol. Its logic is based on a local timestamp correction function. ,in For the first Initial sampling time of each sensor, and These are the sending and receiving times of the node's message, respectively. The maximum error in cross-modal data timestamp alignment achieved through this correction is no more than 0.3ms.

[0030] The second-level deployment allows for a configurable convolutional engine (CCE), and the point cloud path uses... The sparse convolutional kernel structure supports dynamic voxel grid downsampling rates. Compress point cloud data to its original size And preserve edge features; image path adopts and Depth-separable convolutions with variable kernels support lossless image resizing. The kernel recombination rule adjusts weight coefficients based on the current sensor data sampling period and channel latency differences, using a cross-scheduling window for the convolution weights. Dynamic interpolation and fusion of multimodal sensing feature maps ensure that the time difference of the final output fused sensing stream is controlled within 0.5ms.

[0031] The third-level integrated residual compensation unit detects motion blur or sudden changes in illumination (judgment condition is average gradient entropy). A decrease of more than 20% or a standard deviation in brightness distribution (Abnormal offset), automatically activates the lightweight GAN module for high-frequency texture reconstruction to ensure the signal-to-noise ratio of the output perceptual stream. The overall pipeline execution cycle is fixed at 5ms.

[0032] Step 1, through synchronous acquisition by multimodal sensors and ASIC hardware-level preprocessing, not only eliminates the temporal deviation of multi-source data and achieves efficient compression of point cloud data, but also generates a high-quality fused perception stream with a timestamp error ≤0.5ms and a signal-to-noise ratio ≥35dB. The compressed point cloud data that retains edge features, the temporally aligned RGB-D images, and the motion state data together provide an accurate and suitable input foundation for the STPPN dual-branch feature extraction channel in Step 2, ensuring that subsequent feature extraction can be carried out based on environmental data with consistent time dimensions and complete information, thus avoiding feature parsing deviations caused by data asynchrony or insufficient quality from the source.

[0033] Step 2: Based on the data preprocessed in Step 1, perform real-time feature extraction driven by dynamic security boundary to obtain the security boundary coordinates; The specific steps of step 2 include: Step 2.1: Based on the preprocessed data from Step 1, perform lightweight STFPN dual-branch feature extraction; The specific method for step 2.1 is as follows: Based on the preprocessed data from step 1, a lightweight spatiotemporal feature pyramid network (STFPN) is deployed on the FPGA programmable logic unit. The first branch uses a 3D sparse convolutional network to perform sparse spatial sampling and residual encoding on the compressed point cloud, extracting the instantaneous displacement vector field of the obstacle in the local reference coordinate system. The output format is a three-channel tensor. The second branch loads the cropped and optimized YOLOv5 semantic segmentation structure, fuses RGB-D image information, and outputs a segmentation mask tensor. .

[0034] Step 2.2: First, fuse the outputs of the two branches to form a dynamic semantic field. Then, use this semantic field as input to construct a risk probability model based on Lyapunov stability theory to generate the safe passage domain boundary. At the same time, construct a risk distribution function based on trajectory aggregation entropy as a supplement. Finally, output the set of safe boundaries. The specific steps of step 2.2 include: (1) Fuse the outputs of the two branches to form a dynamic semantic field .

[0035] (2) The dynamic safety boundary calculation module uses Using Lyapunov stability theory as input, a risk probability model is constructed, the core mathematical expression of which is: ; in For spatial coordinates, As a window for future prediction, This represents the semantic gradient change at the current point. The system status refresh cycle, As the semantic perturbation stabilization factor, it is set to This model is used to dynamically generate agent-centric safe passage domain boundaries, only when... Only then are they included in the low-risk zone for strategy matching.

[0036] (3) Construct a risk distribution function based on trajectory aggregation entropy as a supplement: , in, For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, This is the normalization coefficient.

[0037] The final output is a set of safe boundaries. The boundary update frequency is 50Hz.

[0038] The working principle of step 2 is as follows: Step 2, Real-time Feature Extraction Driven by Dynamic Safety Boundaries: Construct a lightweight spatiotemporal feature pyramid network (STFPN) and deploy a dual-branch feature extraction channel in the FPGA programmable logic unit: The first branch processes point cloud data through 3D sparse convolution and outputs the obstacle motion vector field; the second branch uses a cropped YOLOv5 model to parse the semantic segmentation mask in the image; the dual-channel output is input to the dynamic safety boundary calculation module, which constructs a real-time risk probability model based on Lyapunov stability theory and outputs the boundary coordinates of the safe passage domain centered on the agent; Step 2.1: Lightweight STFPN Dual-Branch Feature Extraction A lightweight spatiotemporal feature pyramid network (STFPN) is deployed in the FPGA programmable logic unit. The first branch uses a 3D sparse convolutional network to perform sparse spatial sampling and residual encoding on the compressed point cloud to extract the instantaneous displacement vector field of obstacles in the local reference coordinate system. The output format is a three-channel tensor. The second branch loads the cropped and optimized YOLOv5 semantic segmentation structure, fuses RGB-D image information, and outputs a segmentation mask tensor. Its working principle involves channel pruning and 8-bit fixed-point quantization. This process reduces the parameter size and performs redundant channel filtering and numerical precision mapping in the feature extraction process, thereby reducing the overall computational complexity of the model to a fraction of that of the original model. .

[0039] Step 2.2: Dynamic Semantic Field Fusion and Security Boundary Calculation The dual-branch outputs are fused to form a dynamic semantic field. Subsequently, the dynamic security boundary calculation module... Using Lyapunov stability theory as input, a risk probability model is constructed, the core mathematical expression of which is: ,in For spatial coordinates, As a window for future prediction, This represents the semantic gradient change at the current point. The system status refresh cycle, As the semantic perturbation stabilization factor, it is set to This model is used to dynamically generate agent-centric safe passage domain boundaries, only when... Only when these areas are identified are they included in the low-risk region for policy matching. To further improve the discriminative power of risk region identification, a risk distribution function based on trajectory aggregation entropy is also constructed as a supplement: ,in For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, These are the normalization coefficients. The final output is the set of safety boundaries. The boundary update frequency is 50Hz.

[0040] Step 2 relies on the obstacle motion vector field and semantic segmentation mask extracted in parallel by the dual-branch structure of STPPN, combined with the risk probability model built based on Lyapunov stability theory, to finally output the boundary coordinates of the safe passage domain centered on the agent. These boundary coordinates clearly define the low-risk area range in the environment, and together with the target path preset by the agent, they constitute the core input of the A3C reinforcement learning decision-maker in Step 3. This provides the decision-maker with clear environmental risk constraints and path planning references, enabling the subsequent action strategy generation to closely revolve around the goal of "adapting to the dynamic environment within the safe boundary".

[0041] In step 2, the Lightweight Spatiotemporal Feature Pyramid Network (STFPN) possesses efficient spatial semantic decoupling and temporal continuous modeling capabilities. This network operates in parallel on an FPGA using a dual-branch structure. The first branch utilizes a 3D sparse convolutional network to perform sparse spatial sampling and residual encoding on the compressed point cloud, extracting the instantaneous displacement vector field of obstacles in the local reference coordinate system. The output format is a three-channel tensor. Each vector represents the movement trend of the obstacle along the x, y, and z axes per unit time; the second branch loads the cropped and optimized YOLOv5 semantic segmentation structure, fuses RGB-D image information, and outputs a segmentation mask tensor. It is used to identify the boundaries between potential dynamic regions and static scenes. The outputs of the two branches are fused to form a dynamic semantic field. This tensor provides motion semantic information for each valid pixel in the space. Subsequently, the dynamic safety boundary calculation module uses... Using Lyapunov stability theory as input, a risk probability function is constructed to model the future behavior of obstacles without introducing sudden changes in system energy. Its core mathematical expression is as follows:

[0042] Spatial coordinates In the future The probability of risk, It represents the semantic gradient change of spatial points x and y at time t. It is a semantic perturbation stabilizing factor. This is the system status refresh cycle.

[0043] in For spatial coordinates, As a window for future prediction, This represents the semantic gradient change at the current point. The system status refresh cycle, This is a semantic perturbation stabilization factor, whose value is related to the effective period of the environmental state in step 3. There is a coupling relationship, specifically set as follows: ; This risk probability model is used to dynamically generate a safe passage domain boundary centered on the agent. The passage domain is represented by equiprobability lines, indicating a feasible navigation area. A safe passage domain is defined only when... Only when the obstacle changes speed or direction drastically is the path included in the low-risk area for policy matching. This probabilistic model ensures rapid convergence in high-risk areas and provides real-time feedback on path change trends when obstacles change speed or direction drastically. This provides the reinforcement learning decision-maker with stable convergence and safety redundancy as a boundary input, directly affecting the effectiveness of the action policy and the selection of the reward function's time window, further improving the policy's response efficiency in sudden scenarios.

[0044] Furthermore, in step 2, in the lightweight STFPN, channel pruning and 8-bit fixed-point quantization not only reduce the parameter scale but also perform redundant channel filtering and numerical precision mapping in the feature extraction process, reducing the overall computational complexity of the model to 22% of the original model while maintaining a balance between structural tensor flux and information density. The pruning strategy is based on global sensitivity ranking, eliminating low-contribution paths in the perception channel dimension and reconstructing weights in conjunction with the quantization mapping table, thereby ensuring stable operation on the FPGA structure with minimal resource consumption. The dynamic safety boundary calculation module connected after the STFPN output embeds a spatiotemporal uncertainty propagation algorithm. This algorithm takes the tensor field after fusing the point cloud motion field and image semantic mask as input, uses temporal conditional sampling to construct a prediction sequence, and generates a probability distribution map of obstacles in the local grid space within the next 2 seconds in each cycle. To improve the discriminativeness of risk area identification, a risk distribution function based on trajectory aggregation entropy is constructed:

[0045] The spatial locations x and y are predicted in the future time domain. The probability of risk within the range, Z is the normalization coefficient, and N is the total number of sampled trajectories. It is the semantic field mapping weight of the i-th sampled trajectory. It is the spatial offset distance of the i-th sampling trajectory across x and y. It is the risk diffusion radius.

[0046] in, For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, This is the normalization coefficient. The purpose of this formula is to aggregate and model a large number of sampled trajectories through spatial Gaussian weighting, and to assign higher risk responses to locations closer to the agent's path or with high semantic gradients, ensuring coverage and accuracy in identifying sudden dynamic targets. Through continuous updating... With Update The system can obtain a high-risk area distribution map centered on the agent within 2 seconds in real time. Based on this, a shortest path vector generation algorithm is introduced to construct a set of safe boundaries with obstacle avoidance strategies. These boundaries consist of continuous low-risk areas and incorporate obstacle escape vector fields. The output format is as follows:

[0047] It is the set of safety boundaries at time t. It is the x-coordinate of the k-th boundary point. It is the y-coordinate of the k-th boundary point. It is the local optimal navigation direction of the k-th boundary point, and M is the total number of boundary points.

[0048] Each boundary point carries a locally optimal navigation direction. The boundary update frequency is 50Hz, ensuring a stable and continuous low-latency data input channel between the perception and decision-making interfaces. As one of the main inputs of the A3C decision-maker, it provides constraint inputs and environmental feedback benchmarks for the next stage of reward function correction and low-latency strategy optimization.

[0049] Step 3: Based on the safety boundary coordinates obtained in Step 2, generate an action strategy to be executed within the decision cycle; The specific steps of step 3 include: Step 3.1: Generate a latency-sensitive strategy based on the security boundary coordinates obtained in Step 2; Safety boundary coordinates The agent's target path is input into the asynchronous advantage actor-commenter (A3C) network, which generates action policies during the decision cycle.

[0050] Step 3.2: Introduce a time-delay-sensitive reward function and a dynamic decision-making time window; The specific method for step 3.2 is as follows: If the action decision time is delayed Exceeding the environmental status update cycle 50% of Then apply a reward decay coefficient. Simultaneously set a dynamic decision-making time window. .when At that time, the reward function is modified to: ,in As a delay penalty factor, By setting the network to prioritize strategies with low computational complexity, the agent is strengthened to prioritize actions from the low-latency strategy library in unexpected scenarios.

[0051] The working principle of step 3 is as follows: Step 3: Online Incremental Optimization of the Reinforcement Learning Decision Maker: Input the safety boundary coordinates and the agent's target path from Step 2 into the Asynchronous Advantage Actor-Commentator (A3C) network, and generate action policies within the decision cycle; the key improvement lies in introducing a time-delay-sensitive reward function: if the action decision has a time delay... Exceeding 50% of the environmental status update cycle, i.e. Then apply a reward decay coefficient. ;

[0052] It is the reward decay coefficient. It is the delay in action decision-making. It is the maximum decision delay This forces the network to prioritize strategies with low computational complexity, while simultaneously setting a dynamic decision-making time window:

[0053] T d It is a dynamic decision-making time window. It is the time window coefficient, T e It is the environmental status update cycle. when At that time, the reward function is modified to:

[0054] Here, R is the modified reward function, k is the original reward function, and t is the time delay penalty factor. d It is the action decision delay, T d It is a dynamic decision-making time window in, The above settings enhance the agent's ability to prioritize actions from the low-latency policy library in emergency scenarios.

[0055] Step 3.1: Generation of Delay-Sensitive Policies Safety boundary coordinates The agent's target path is input into the asynchronous advantage actor-commenter (A3C) network, which generates action policies during the decision cycle.

[0056] Step 3.2: Delay-sensitive rewards and decision-making time window constraints.

[0057] The key improvement lies in the introduction of a time-delay-sensitive reward function and a dynamic decision-making time window.

[0058] Its working principle is: if the action decision delay Exceeding the environmental status update cycle 50% of Then apply a reward decay coefficient. Simultaneously set a dynamic decision-making time window. .when At that time, the reward function is modified to: ,in As a delay penalty factor, By setting the network to prioritize strategies with low computational complexity, the agent is strengthened to prioritize actions from the low-latency strategy library in unexpected scenarios.

[0059] Step 3, through the design of a latency-sensitive reward function and a dynamic decision-making time window, forces the A3C network to generate a low-latency action policy that conforms to safety boundary constraints. In actual execution, this policy needs to be continuously optimized based on real-time environmental feedback. This provides a demand-driven approach for the construction of the feedback loop of the distributed execution framework in Step 4: When the action policy output in Step 3 is executed, Step 4 can monitor changes in the environmental state through INT technology, and quickly feed back disturbances such as path latency fluctuations and sudden obstacles to the perception layer, forming a complete loop of "perception-decision-execution-feedback", ensuring that the decision-making policy can be dynamically adjusted with the environment.

[0060] Step 4: Based on the action strategy generated in Step 3, construct the feedback loop of the distributed execution framework to achieve real-time closed-loop optimization; The specific steps of step 4 include: Step 4.1: Based on the action strategy generated in Step 3, perform real-time monitoring and capture of environmental disturbances. While the intelligent agent is performing actions, it collects data on changes in environmental conditions, and monitors the transmission delay fluctuations of the core monitoring path and the changes in the status of new obstacles. Step 4.2: Inject disturbance assessment and priority scheduling to achieve real-time closed-loop optimization; The specific steps of step 4.2 include: (1) When INT detects path delay fluctuations Or, when a sudden obstacle appears, the HDC high-latency capture mechanism is triggered to copy the current critical network packets; (2) The system constructs a disturbance cognition priority mapping table to achieve real-time closed-loop optimization: the disturbance intensity is evaluated based on the feedback selection function. ,in Indicates the first The current latency change of each path node, For its reference delay, This is the latest velocity vector of the node. As the reference vector, For marking obstacle areas, This is a disturbance-weighted index. It is a stability factor. Simultaneously, a perturbation level scoring function is defined. Perform priority sorting, where To disrupt historical influence. Represents the rate of change of policy entropy. These are the weighting coefficients.

[0061] The system is based on the score As a result, the high-priority perturbation is in An internal STFPN input layer is injected to update the perceptual flow and security boundary, enabling real-time closed-loop optimization.

[0062] The working principle of step 4 is as follows: Step 4: Constructing the feedback loop of the distributed execution framework: While the agent is executing actions, environmental state change data is collected through in-band network telemetry (INT) technology. When path delay fluctuations ≥15% or sudden obstacles are detected, the HDC high-latency capture mechanism is triggered to copy the current network packets. The environmental change characteristics are injected into the STPPN input layer of Step 2 within ≤1ms, so that the system can continuously adjust perception and decision-making according to real-time environmental changes, and realize dynamic adaptation and cyclic optimization of the entire process.

[0063] In the distributed execution framework, the real-time feedback of agent actions is designed as a strongly coupled state monitoring path. The In-Band Telemetry (INT) technology uses packet-level timestamp injection and path marking to monitor state disturbances in the execution link with high frequency and low latency. The core system indicators are path transmission delay fluctuations and changes in the status of new obstacles. Within the INT detection cycle, once a path delay fluctuation reaches or exceeds 15%, or a newly emerging high-probability obstacle area overlaps with the current passage boundary, the High Latency Capture (HDC) mechanism is immediately triggered to copy critical network data packets. This mechanism is deployed at the core switching node and, based on a hardware-level offloading channel, completes the extraction of state features of the sensing link within 1 millisecond. To effectively compress and focus disturbances before data injection into the STPPN, a feedback selection function is constructed.

[0064] This is the feedback selection function value, where t is the current time and N is the total number of path nodes. It is the obstacle region identifier of the i-th path node. It represents the current delay change of the i-th path node. It is the baseline delay of the i-th path node. It is a disturbance-weighted index. It is the latest velocity vector of the i-th path node. It is the reference velocity vector of the i-th path node output in the previous cycle of STPPN. It is a stabilizing factor.

[0065] in Indicates the first The current latency change of each path node, For its reference delay, This is the latest velocity vector of the node. This is the reference vector output by STPPN in the previous cycle. For marking obstacle areas, This is a disturbance-weighted index. This is a stability factor to prevent zero denominator. This function unifies the measurement of path delay perturbation and motion vector perturbation, mapping dynamic changes to a feedback intensity score through exponential weighting. Only when... Only then does it enter the STPPN feedback injection channel. The threshold is dynamically adjusted based on the system's current total computing power and the A3C decision load floating window, avoiding interference from low-priority disturbances on the sensing link bandwidth. Injected data is inserted into the previous round's residual fusion layer in the STPPN network via skip connections, updating... Tensor and trigger boundary re-estimation mechanism to ensure that the spatial coupling between the current safe passage domain and the policy output converges stably within two frame periods.

[0066] In the distributed execution framework of step 4, the system must respond to path disturbances and dynamic changes in obstacles within a millisecond-level time window, and rapidly feed information back to the STFPN network to update the perception boundary. To avoid feedback link congestion or computing resource contention when large-scale disturbances occur concurrently, a disturbance cognition priority mapping table is further introduced as the core scheduling unit of the rule logic, based on INT technology and HDC capture mechanism. The disturbance cognition priority mapping table dynamically calculates the disturbance level score based on two key indicators: the historical impact of the disturbance and the fluctuation response degree of the policy objective function. The score result determines the bandwidth allocation of the INT channel and the write priority of the STFPN residual injection channel. The historical impact of perturbations is modeled by the average effect of the perturbation on the agent's action adjustment amplitude over the past N feedback cycles, denoted as […]. The objective function fluctuation response is based on the rate of change of confidence entropy of the A3C strategy output after the current disturbance occurs, denoted as... ,like Figure 1 The figure shown illustrates the impact of different types of disturbances on the entropy value of the policy network.

[0067] It is the rate of change of policy entropy of perturbation i at time t. It is the contribution of perturbation i to the policy entropy of the policy network at time t. It is the contribution of perturbation i to the policy entropy of the policy network at time t-1.

[0068] in Indicates disturbance The contribution of the policy entropy to the policy network at the current moment. Based on this, the perturbation level scoring function is defined as follows:

[0069] like Figure 2 The diagram shown is a flowchart for calculating the disturbance level scoring function.

[0070] It is the rating of the i-th perturbation. It is the historical influence weighting coefficient. It is the historical influence of the i-th perturbation. It is the real-time response coefficient. It is the rate of change of policy entropy of disturbance i at time t.

[0071] in and These are the historical impact weighting coefficient and the real-time response coefficient, respectively, and their sum is 1. The default value is... This aims to balance long-term intervention trends with the value of immediate feedback. Disturbance Score Within each period, they are sorted, and the system maps high-priority perturbations to bandwidth limit allocation ranges based on the scoring results, such as... Figure 3 The diagram shows the timeline from disturbance occurrence to scoring. Simultaneously, the high-frequency injection path of the STPPN residual tensor channel is activated, while low-priority disturbances are handled using delayed injection or degraded to the state cache channel, ensuring the system maintains the core path's decision-making response speed even in resource-constrained or disturbance-intensive scenarios. Figure 4 The diagram shown is a mapping of disturbance priority and resource allocation.

[0072] By setting up this perturbation perception priority mapping, the system has the ability to allocate perception resources in a task-oriented manner when facing complex external dynamic changes, while maintaining a balanced load on the INT link bandwidth under different perturbation intensity scenarios. This enables the perception-decision-execution closed loop to achieve stability and high-throughput feedback coordination in a high-concurrency environment, and also provides a standardized measurement reference for sharing perturbation levels among collaborative agents in the future.

[0073] Step 4.1: Real-time monitoring and capture of environmental disturbances While the agent is executing actions, in-band network telemetry (INT) technology is used to collect data on changes in environmental state, focusing on monitoring path transmission latency fluctuations and changes in the status of new obstacles. Its working principle involves high-frequency, low-latency monitoring of state disturbances in the execution chain through packet-level timestamp injection and path marking.

[0074] Step 4.2: Disturbance Assessment and Priority Scheduling Injection When INT detects path delay fluctuations In the event of a sudden obstacle, the HDC high-latency capture mechanism is triggered to copy the current critical network packets. To intelligently schedule feedback resources, the system constructs a disturbance recognition priority mapping table.

[0075] Its working principle is based on evaluating the disturbance strength using a feedback selection function: ,in Indicates the first The current latency change of each path node, For its reference delay, This is the latest velocity vector of the node. As the reference vector, For marking obstacle areas, This is a disturbance-weighted index. It is a stability factor. Simultaneously, a perturbation level scoring function is defined. Perform priority sorting, where To disrupt historical influence. Represents the rate of change of policy entropy. These are weighting coefficients. The system bases the score on... As a result, the high-priority perturbation is in An internal STFPN input layer is injected to update the perceptual flow and security boundary, enabling real-time closed-loop optimization.

[0076] This solution constructs an integrated intelligent agent framework with dynamic environmental perception, adaptive decision-making, and high-speed feedback capabilities. It synchronously collects environmental information through multimodal sensors and performs spatiotemporal alignment and compression on an ASIC hardware accelerator. A lightweight STPPN network extracts point cloud and image features in parallel on an FPGA. Combining Lyapunov stability and spatiotemporal uncertainty propagation, a dynamic safe passage boundary is constructed. Subsequently, the boundary information and target path are input into a latency-sensitive A3C decision-maker to generate low-latency action strategies. At the same time, during the action execution phase, environmental disturbances are monitored in real time through INT technology, and feedback channels are scheduled based on a disturbance cognition priority mapping table. High-impact disturbances are injected into the perception network within 1ms to form a real-time closed loop, ensuring stable, fast, and controllable behavioral response capabilities even under sudden environmental changes.

[0077] Example 1 Step 1: Parallel Acquisition and Hardware-Level Preprocessing of Multimodal Sensor Data: Environmental point cloud, RGB-D image, and motion state data are simultaneously captured by LiDAR, vision sensor, and inertial measurement unit. An ASIC hardware accelerator is used to perform spatiotemporal alignment of the raw data. Convolutional kernel recombination technology is used to eliminate temporal biases from multiple sensors, generating a fused sensing stream with a timestamp error ≤0.5ms. The aim is to eliminate temporal biases caused by differences in sampling mechanisms among multimodal sensors such as LiDAR, vision sensor, and inertial measurement unit, ensuring strict temporal alignment of environmental point cloud, RGB-D image, and motion state data. The point cloud data is processed using a voxel grid. Compressing data to 30%~35% of its original size while preserving edge features significantly reduces data volume to improve processing efficiency while avoiding the loss of key environmental features. The ASIC hardware accelerator adopts a three-stage pipeline architecture. The first stage performs timestamp calibration of sensor data, achieving nanosecond-level synchronization through the IEEE 1588v2 precision clock protocol. The second stage deploys a configurable convolutional engine (CCE) that supports dynamic switching of point cloud voxelization parameters and image downsampling rate. The third stage integrates a residual compensation unit, which automatically activates a generative adversarial network (GAN) to repair data when motion blur or sudden changes in illumination are detected, ensuring that the signal-to-noise ratio of the output sensor stream is ≥35dB. For the time allocation of each processing stage in the three-stage pipeline architecture, the key parameters of the convolution kernel reassembly technique, and the elimination of timing deviations, the overall pipeline execution cycle is fixed at 5ms. The first-stage timestamp calibration module occupies 1ms and is mainly responsible for the high-precision synchronization of LiDAR, RGB-D vision, and IMU data. By integrating the IEEE1588v2 precision clock protocol, the sensor synchronization error is controlled at the nanosecond level. Its logic is based on the local timestamp correction function.

[0078] The corrected sampling time of the i-th sensor. For the first Initial sampling time of each sensor, and These represent the sending and receiving times of the message at that node.

[0079] This correction ensures that the maximum error in cross-modal data timestamp alignment does not exceed 0.3ms. The second-level CCE module is allocated 2.5ms of execution time and deploys configurable convolutional kernel reassembly units. The point cloud path uses a 5×5×5 sparse convolutional kernel structure, supporting dynamic voxel grid downsampling rates {1 / 2, 1 / 4, 1 / 8}. The image path uses depthwise separable convolutions with 3×3 and 5×5 variable kernels to support lossless switching of image sizes from the original 640×480 to the compressed 160×120. The convolutional kernel reassembly rule adjusts the weight coefficients based on the current sensor data sampling period and channel delay differences, and a cross-scheduling window for convolutional weights is set. Dynamic interpolation and fusion of multimodal sensing feature maps ensures that the time difference of the final output fused sensing stream is controlled within 0.5ms. The third-level residual compensation unit occupies 1.5ms, and its key function is to quickly determine whether image blurring or abnormal lighting events exceed the reconstruction threshold. A decrease of more than 20% or a standard deviation in brightness distribution In the event of abnormal offsets, a lightweight GAN module is automatically activated for high-frequency texture reconstruction, with the reconstruction time controlled within 1ms, ultimately ensuring that the signal-to-noise ratio of the fused perceptual stream reaches or exceeds 35dB. Through the above architecture design, this ASIC module not only achieves accurate alignment and feature preservation of multimodal information, but also provides highly consistent and high-quality data input for subsequent STPPN feature extraction.

[0080] Step 1.1: Synchronous data acquisition from multiple sensor sources The system simultaneously captures environmental point clouds, RGB-D images, and motion state data using lidar, vision sensors, and inertial measurement units.

[0081] Step 1.2: ASIC hardware-level spatiotemporal alignment and compression.

[0082] The raw data is processed using an ASIC hardware accelerator.

[0083] Its core working principle is achieved through a three-stage pipeline architecture.

[0084] The first stage performs timestamp calibration of sensor data, achieving nanosecond-level synchronization via the IEEE 1588v2 precision clock protocol. Its logic is based on a local timestamp correction function. ,in For the first Initial sampling time of each sensor, and These are the sending and receiving times of the node's message, respectively. The maximum error in cross-modal data timestamp alignment achieved through this correction is no more than 0.3ms.

[0085] The second-level deployment allows for a configurable convolutional engine (CCE), and the point cloud path uses... The sparse convolutional kernel structure supports dynamic voxel grid downsampling rates. Compress point cloud data to its original size And preserve edge features; image path adopts and Depth-separable convolutions with variable kernels support lossless image resizing. The kernel recombination rule adjusts weight coefficients based on the current sensor data sampling period and channel latency differences, using a cross-scheduling window for the convolution weights. Dynamic interpolation and fusion of multimodal sensing feature maps ensure that the time difference of the final output fused sensing stream is controlled within 0.5ms.

[0086] The third-level integrated residual compensation unit detects motion blur or sudden changes in illumination (judgment condition is average gradient entropy). A decrease of more than 20% or a standard deviation in brightness distribution (Abnormal offset), automatically activates the lightweight GAN module for high-frequency texture reconstruction to ensure the signal-to-noise ratio of the output perceptual stream. The overall pipeline execution cycle is fixed at 5ms.

[0087] Step 1, through synchronous acquisition by multimodal sensors and ASIC hardware-level preprocessing, not only eliminates the temporal deviation of multi-source data and achieves efficient compression of point cloud data, but also generates a high-quality fused perception stream with a timestamp error ≤0.5ms and a signal-to-noise ratio ≥35dB. The compressed point cloud data that retains edge features, the temporally aligned RGB-D images, and the motion state data together provide an accurate and suitable input foundation for the STPPN dual-branch feature extraction channel in Step 2, ensuring that subsequent feature extraction can be carried out based on environmental data with consistent time dimensions and complete information, thus avoiding feature parsing deviations caused by data asynchrony or insufficient quality from the source.

[0088] Step 2, Real-time Feature Extraction Driven by Dynamic Safety Boundaries: Construct a lightweight spatiotemporal feature pyramid network (STFPN) and deploy a dual-branch feature extraction channel in the FPGA programmable logic unit: The first branch processes point cloud data through 3D sparse convolution and outputs the obstacle motion vector field; the second branch uses a cropped YOLOv5 model to parse the semantic segmentation mask in the image; the dual-channel output is input to the dynamic safety boundary calculation module, which constructs a real-time risk probability model based on Lyapunov stability theory and outputs the boundary coordinates of the safe passage domain centered on the agent; Step 2.1: Lightweight STFPN Dual-Branch Feature Extraction A lightweight spatiotemporal feature pyramid network (STFPN) is deployed in the FPGA programmable logic unit. The first branch uses a 3D sparse convolutional network to perform sparse spatial sampling and residual encoding on the compressed point cloud to extract the instantaneous displacement vector field of obstacles in the local reference coordinate system. The output format is a three-channel tensor. The second branch loads the cropped and optimized YOLOv5 semantic segmentation structure, fuses RGB-D image information, and outputs a segmentation mask tensor. Its working principle involves channel pruning and 8-bit fixed-point quantization. This process reduces the parameter size and performs redundant channel filtering and numerical precision mapping in the feature extraction process, thereby reducing the overall computational complexity of the model to a fraction of that of the original model. .

[0089] Step 2.2: Dynamic Semantic Field Fusion and Security Boundary Calculation The dual-branch outputs are fused to form a dynamic semantic field. Subsequently, the dynamic security boundary calculation module... Using Lyapunov stability theory as input, a risk probability model is constructed, the core mathematical expression of which is: ,in For spatial coordinates, As a window for future prediction, This represents the semantic gradient change at the current point. The system status refresh cycle, As the semantic perturbation stabilization factor, it is set to This model is used to dynamically generate agent-centric safe passage domain boundaries, only when... Only when these areas are identified are they included in the low-risk region for policy matching. To further improve the discriminative power of risk region identification, a risk distribution function based on trajectory aggregation entropy is also constructed as a supplement: ,in For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, These are the normalization coefficients. The final output is the set of safety boundaries. The boundary update frequency is 50Hz.

[0090] Step 2 relies on the obstacle motion vector field and semantic segmentation mask extracted in parallel by the dual-branch structure of STPPN, combined with the risk probability model built based on Lyapunov stability theory, to finally output the boundary coordinates of the safe passage domain centered on the agent. These boundary coordinates clearly define the low-risk area range in the environment, and together with the target path preset by the agent, they constitute the core input of the A3C reinforcement learning decision-maker in Step 3. This provides the decision-maker with clear environmental risk constraints and path planning references, enabling the subsequent action strategy generation to closely revolve around the goal of "adapting to the dynamic environment within the safe boundary".

[0091] Step 3: Online Incremental Optimization of the Reinforcement Learning Decision Maker: Input the safety boundary coordinates and the agent's target path from Step 2 into the Asynchronous Advantage Actor-Commentator (A3C) network, and generate action policies within the decision cycle; the key improvement lies in introducing a time-delay-sensitive reward function: if the action decision has a time delay... Exceeding 50% of the environmental status update cycle, i.e. Then apply a reward decay coefficient. ;

[0092] It is the reward decay coefficient. It is the delay in action decision-making. It is the maximum decision delay This forces the network to prioritize strategies with low computational complexity, while simultaneously setting a dynamic decision-making time window:

[0093] T d It is a dynamic decision-making time window. It is the time window coefficient, T e It is the environmental status update cycle. when At that time, the reward function is modified to:

[0094] Here, R is the modified reward function, k is the original reward function, and t is the time delay penalty factor. d It is the action decision delay, T d It is a dynamic decision-making time window in, As a delay penalty factor, The above settings enhance the agent's ability to prioritize actions from the low-latency policy library in emergency scenarios.

[0095] In step 2, the Lightweight Spatiotemporal Feature Pyramid Network (STFPN) possesses efficient spatial semantic decoupling and temporal continuous modeling capabilities. This network operates in parallel on an FPGA using a dual-branch structure. The first branch utilizes a 3D sparse convolutional network to perform sparse spatial sampling and residual encoding on the compressed point cloud, extracting the instantaneous displacement vector field of obstacles in the local reference coordinate system. The output format is a three-channel tensor. Each vector represents the movement trend of the obstacle along the x, y, and z axes per unit time; the second branch loads the cropped and optimized YOLOv5 semantic segmentation structure, fuses RGB-D image information, and outputs a segmentation mask tensor. It is used to identify the boundaries between potential dynamic regions and static scenes. The outputs of the two branches are fused to form a dynamic semantic field. This tensor provides motion semantic information for each valid pixel in the space. Subsequently, the dynamic safety boundary calculation module uses... Using Lyapunov stability theory as input, a risk probability function is constructed to model the future behavior of obstacles without introducing sudden changes in system energy. Its core mathematical expression is as follows:

[0096] Spatial coordinates In the future The probability of risk, It represents the semantic gradient change of spatial points x and y at time t. It is a semantic perturbation stabilizing factor. This is the system status refresh cycle.

[0097] in For spatial coordinates, As a window for future prediction, This represents the semantic gradient change at the current point. The system status refresh cycle, This is a semantic perturbation stabilization factor, whose value is related to the effective period of the environmental state in step 3. There is a coupling relationship, specifically set as follows: ; This risk probability model is used to dynamically generate a safe passage domain boundary centered on the agent. The passage domain is represented by equiprobability lines, indicating a feasible navigation area. A safe passage domain is defined only when... Only when the obstacle changes speed or direction drastically is the path included in the low-risk area for policy matching. This probabilistic model ensures rapid convergence in high-risk areas and provides real-time feedback on path change trends when obstacles change speed or direction drastically. This provides the reinforcement learning decision-maker with stable convergence and safety redundancy as a boundary input, directly affecting the effectiveness of the action policy and the selection of the reward function's time window, further improving the policy's response efficiency in sudden scenarios.

[0098] In step 2, in the lightweight STFPN, channel pruning and 8-bit fixed-point quantization not only reduce the parameter size but also perform redundant channel filtering and numerical precision mapping in the feature extraction process, reducing the overall computational complexity of the model to 22% of the original model while maintaining a balance between structural tensor flux and information density. The pruning strategy is based on global sensitivity ranking, eliminating low-contribution paths in the perception channel dimension and reconstructing weights in conjunction with the quantization mapping table, thereby ensuring stable operation on the FPGA structure with minimal resource consumption. The dynamic safety boundary calculation module connected after the STFPN output embeds a spatiotemporal uncertainty propagation algorithm. This algorithm takes the tensor field after fusing the point cloud motion field and image semantic mask as input, uses temporal conditional sampling to construct a prediction sequence, and generates a probability distribution map of obstacles in the local grid space within the next 2 seconds in each cycle. To improve the discriminativeness of risk area identification, a risk distribution function based on trajectory aggregation entropy is constructed:

[0099] The spatial locations x and y are predicted in the future time domain. The probability of risk within the range, Z is the normalization coefficient, and N is the total number of sampled trajectories. It is the semantic field mapping weight of the i-th sampled trajectory. It is the spatial offset distance of the i-th sampling trajectory across x and y. It is the risk diffusion radius.

[0100] in, For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, This is the normalization coefficient. The purpose of this formula is to aggregate and model a large number of sampled trajectories through spatial Gaussian weighting, and to assign higher risk responses to locations closer to the agent's path or with high semantic gradients, ensuring coverage and accuracy in identifying sudden dynamic targets. Through continuous updating... With Update The system can obtain a high-risk area distribution map centered on the agent within 2 seconds in real time. Based on this, a shortest path vector generation algorithm is introduced to construct a set of safe boundaries with obstacle avoidance strategies. These boundaries consist of continuous low-risk areas and incorporate obstacle escape vector fields. The output format is as follows:

[0101] It is the set of safety boundaries at time t. It is the x-coordinate of the k-th boundary point. It is the y-coordinate of the k-th boundary point. It is the local optimal navigation direction of the k-th boundary point, and M is the total number of boundary points.

[0102] Each boundary point carries a locally optimal navigation direction. The boundary update frequency is 50Hz, ensuring a stable and continuous low-latency data input channel between the perception and decision-making interfaces. As one of the main inputs of the A3C decision-maker, it provides constraint inputs and environmental feedback benchmarks for the next stage of reward function correction and low-latency strategy optimization.

[0103] Step 3.1: Generation of Delay-Sensitive Policies Safety boundary coordinates The agent's target path is input into the asynchronous advantage actor-commenter (A3C) network, which generates action policies during the decision cycle.

[0104] Step 3.2: Delay-sensitive rewards and decision-making time window constraints.

[0105] The key improvement lies in the introduction of a time-delay-sensitive reward function and a dynamic decision-making time window.

[0106] Its working principle is: if the action decision delay Exceeding the environmental status update cycle 50% of Then apply a reward decay coefficient. Simultaneously set a dynamic decision-making time window. .when At that time, the reward function is modified to: ,in As a delay penalty factor, By setting the network to prioritize strategies with low computational complexity, the agent is strengthened to prioritize actions from the low-latency strategy library in unexpected scenarios.

[0107] Step 3, through the design of a latency-sensitive reward function and a dynamic decision-making time window, forces the A3C network to generate a low-latency action policy that conforms to safety boundary constraints. In actual execution, this policy needs to be continuously optimized based on real-time environmental feedback. This provides a demand-driven approach for the construction of the feedback loop of the distributed execution framework in Step 4: When the action policy output in Step 3 is executed, Step 4 can monitor changes in the environmental state through INT technology, and quickly feed back disturbances such as path latency fluctuations and sudden obstacles to the perception layer, forming a complete loop of "perception-decision-execution-feedback", ensuring that the decision-making policy can be dynamically adjusted with the environment.

[0108] Step 4: Constructing the feedback loop of the distributed execution framework: While the agent is executing actions, environmental state change data is collected through in-band network telemetry (INT) technology. When path delay fluctuations ≥15% or sudden obstacles are detected, the HDC high-latency capture mechanism is triggered to copy the current network packets. The environmental change characteristics are injected into the STPPN input layer of Step 2 within ≤1ms, so that the system can continuously adjust perception and decision-making according to real-time environmental changes, and realize dynamic adaptation and cyclic optimization of the entire process.

[0109] In the distributed execution framework, the real-time feedback of agent actions is designed as a strongly coupled state monitoring path. The In-Band Telemetry (INT) technology uses packet-level timestamp injection and path marking to monitor state disturbances in the execution link with high frequency and low latency. The core system indicators are path transmission delay fluctuations and changes in the status of new obstacles. Within the INT detection cycle, once a path delay fluctuation reaches or exceeds 15%, or a newly emerging high-probability obstacle area overlaps with the current passage boundary, the High Latency Capture (HDC) mechanism is immediately triggered to copy critical network data packets. This mechanism is deployed at the core switching node and, based on a hardware-level offloading channel, completes the extraction of state features of the sensing link within 1 millisecond. To effectively compress and focus disturbances before data injection into the STPPN, a feedback selection function is constructed.

[0110] This is the feedback selection function value, where t is the current time and N is the total number of path nodes. It is the obstacle region identifier of the i-th path node. It represents the current delay change of the i-th path node. It is the baseline delay of the i-th path node. It is a disturbance-weighted index. It is the latest velocity vector of the i-th path node. It is the reference velocity vector of the i-th path node output in the previous cycle of STPPN. It is a stabilizing factor.

[0111] in Indicates the first The current latency change of each path node, For its reference delay, This is the latest velocity vector of the node. This is the reference vector output by STPPN in the previous cycle. For marking obstacle areas, This is a disturbance-weighted index. This is a stability factor to prevent zero denominator. This function unifies the measurement of path delay perturbation and motion vector perturbation, mapping dynamic changes to a feedback intensity score through exponential weighting. Only when... Only then does it enter the STPPN feedback injection channel. The threshold is dynamically adjusted based on the system's current total computing power and the A3C decision load floating window, avoiding interference from low-priority disturbances on the sensing link bandwidth. Injected data is inserted into the previous round's residual fusion layer in the STPPN network via skip connections, updating... Tensor and trigger boundary re-estimation mechanism to ensure that the spatial coupling between the current safe passage domain and the policy output converges stably within two frame periods.

[0112] In the distributed execution framework of step 4, the system must respond to path disturbances and dynamic changes in obstacles within a millisecond-level time window, and rapidly feed information back to the STFPN network to update the perception boundary. To avoid feedback link congestion or computing resource contention when large-scale disturbances occur concurrently, a disturbance cognition priority mapping table is further introduced as the core scheduling unit of the rule logic, based on INT technology and HDC capture mechanism. The disturbance cognition priority mapping table dynamically calculates the disturbance level score based on two key indicators: the historical impact of the disturbance and the fluctuation response degree of the policy objective function. The score result determines the bandwidth allocation of the INT channel and the write priority of the STFPN residual injection channel. The historical impact of perturbations is modeled by the average effect of the perturbation on the agent's action adjustment amplitude over the past N feedback cycles, denoted as […]. The objective function fluctuation response is based on the rate of change of confidence entropy of the A3C strategy output after the current disturbance occurs, denoted as... ,like Figure 1 The figure shown illustrates the impact of different types of disturbances on the entropy value of the policy network.

[0113] It is the rate of change of policy entropy of perturbation i at time t. It is the contribution of perturbation i to the policy entropy of the policy network at time t. It is the contribution of perturbation i to the policy entropy of the policy network at time t-1.

[0114] in Indicates disturbance The contribution of the policy entropy to the policy network at the current moment. Based on this, the perturbation level scoring function is defined as follows:

[0115] like Figure 2 The diagram shown is a flowchart for calculating the disturbance level scoring function.

[0116] It is the rating of the i-th perturbation. It is the historical influence weighting coefficient. It is the historical influence of the i-th perturbation. It is the real-time response coefficient. It is the rate of change of policy entropy of disturbance i at time t.

[0117] in and These are the historical impact weighting coefficient and the real-time response coefficient, respectively, and their sum is 1. The default value is... This aims to balance long-term intervention trends with the value of immediate feedback. Disturbance Score Within each period, they are sorted, and the system maps high-priority perturbations to bandwidth limit allocation ranges based on the scoring results, such as... Figure 3 The diagram shows the timeline from disturbance occurrence to scoring. Simultaneously, the high-frequency injection path of the STPPN residual tensor channel is activated, while low-priority disturbances are handled using delayed injection or degraded to the state cache channel, ensuring the system maintains the core path's decision-making response speed even in resource-constrained or disturbance-intensive scenarios. Figure 4 The diagram shown is a mapping of disturbance priority and resource allocation.

[0118] By setting up this perturbation perception priority mapping, the system has the ability to allocate perception resources in a task-oriented manner when facing complex external dynamic changes, while maintaining a balanced load on the INT link bandwidth under different perturbation intensity scenarios. This enables the perception-decision-execution closed loop to achieve stability and high-throughput feedback coordination in a high-concurrency environment, and also provides a standardized measurement reference for sharing perturbation levels among collaborative agents in the future.

[0119] Step 4.1: Real-time monitoring and capture of environmental disturbances While the agent is executing actions, in-band network telemetry (INT) technology is used to collect data on changes in environmental state, focusing on monitoring path transmission latency fluctuations and changes in the status of new obstacles. Its working principle involves high-frequency, low-latency monitoring of state disturbances in the execution chain through packet-level timestamp injection and path marking.

[0120] Step 4.2: Disturbance Assessment and Priority Scheduling Injection When INT detects path delay fluctuations In the event of a sudden obstacle, the HDC high-latency capture mechanism is triggered to copy the current critical network packets. To intelligently schedule feedback resources, the system constructs a disturbance recognition priority mapping table.

[0121] Its working principle is based on evaluating the disturbance strength using a feedback selection function: ,in Indicates the first The current latency change of each path node, For its reference delay, This is the latest velocity vector of the node. As the reference vector, For marking obstacle areas, This is a disturbance-weighted index. It is a stability factor. Simultaneously, a perturbation level scoring function is defined. Perform priority sorting, where To disrupt historical influence. Represents the rate of change of policy entropy. These are weighting coefficients. The system bases the score on... As a result, the high-priority perturbation is in An internal STFPN input layer is injected to update the perceptual flow and security boundary, enabling real-time closed-loop optimization.

[0122] This solution constructs an integrated intelligent agent framework with dynamic environmental perception, adaptive decision-making, and high-speed feedback capabilities. It synchronously collects environmental information through multimodal sensors and performs spatiotemporal alignment and compression on an ASIC hardware accelerator. A lightweight STPPN network extracts point cloud and image features in parallel on an FPGA. Combining Lyapunov stability and spatiotemporal uncertainty propagation, a dynamic safe passage boundary is constructed. Subsequently, the boundary information and target path are input into a latency-sensitive A3C decision-maker to generate low-latency action strategies. At the same time, during the action execution phase, environmental disturbances are monitored in real time through INT technology, and feedback channels are scheduled based on a disturbance cognition priority mapping table. High-impact disturbances are injected into the perception network within 1ms to form a real-time closed loop, ensuring stable, fast, and controllable behavioral response capabilities even under sudden environmental changes.

[0123] The innovation of this invention lies in: Step 1 involves simultaneous acquisition of data from multiple modal sensors and ASIC hardware-accelerated preprocessing: environmental point clouds, RGB-D images, and motion state data (such as acceleration and angular velocity) from the inertial measurement unit are simultaneously captured by three types of devices: LiDAR, vision sensors, and inertial measurement units, to achieve comprehensive perception of multi-dimensional information about the dynamic environment; at the same time, an ASIC hardware accelerator is used to perform hardware-level preprocessing on the raw data, with the core purpose of quickly completing the initial integration of multi-source data through the parallel processing capabilities at the hardware level.

[0124] Step 1 uses convolutional kernel reassembly to eliminate temporal discrepancies: To address the temporal asynchrony problem caused by differences in sampling frequency and transmission delay in multimodal sensors, convolutional kernel reassembly is used to optimize the convolutional processing logic of the original data. By dynamically adjusting the timing and range of the convolutional kernel, timestamp discrepancies of multi-source data are eliminated, ultimately generating a fused perception stream with a timestamp error ≤0.5ms. This ensures that various types of data are strictly matched in the time dimension during subsequent feature extraction, avoiding environmental perception distortion caused by temporal misalignment.

[0125] Step 1: Point cloud voxel mesh compression. The massive point cloud data collected by LiDAR is compressed using a voxel mesh algorithm, reducing the data volume to 30%~35% of the original size. This significantly reduces storage and computational load while preserving edge features to ensure that the compressed data still reflects key environmental structural information. This process is compatible with the computing power limitations of ASIC accelerators and provides high-quality input for subsequent 3D sparse convolution extraction of obstacle motion vector fields.

[0126] Step 1: ASIC Three-Stage Pipeline Architecture Design: The ASIC hardware accelerator adopts a three-stage collaborative processing pipeline architecture: The first stage implements nanosecond-level timestamp calibration based on the IEEE 1588v2 precision clock protocol to ensure the consistency of the time base of multi-source sensor data; the second stage deploys a configurable convolutional engine CCE, which supports dynamic switching of point cloud voxelization parameters and image downsampling rate according to environmental complexity, balancing processing efficiency and accuracy; the third stage integrates a residual compensation unit, which automatically activates a generative adversarial network (GAN) to repair data when motion blur or sudden changes in illumination are detected, ensuring that the signal-to-noise ratio of the output sensing stream is ≥35dB and improving data reliability.

[0127] In step 2, the STPPN dual-branch feature extraction channel deploys a lightweight spatiotemporal feature pyramid network (STFPN) on the FPGA programmable logic unit, setting up a dual-branch parallel processing structure: the first branch processes the compressed point cloud data through 3D sparse convolution to extract the motion vector field of obstacles in three-dimensional space; the second branch uses a cropped YOLOv5 model to parse the RGB-D image and generate a semantic segmentation mask; the dual-channel outputs are fused and input into the dynamic safety boundary calculation module to provide multi-dimensional feature support for risk assessment.

[0128] Step 2 involves a risk probability model based on Lyapunov stability theory: A real-time risk probability model is constructed based on Lyapunov stability theory (ensuring that the system state converges to the safe domain). The model takes the feature tensor output by STFPN as input and calculates the risk probability of each spatial location in the environment. The model outputs the boundary coordinates of the safe passage domain centered on the agent. Only when the location risk probability Pr(x,y,t+τ) < 0.3 is it determined to be a low-risk area and used for policy matching, ensuring that the agent's decision is always based on a stable and safe environmental boundary.

[0129] Step 3: Latency-sensitive reward function and dynamic decision window. A latency-sensitive reward function is introduced into the A3C reinforcement learning network. When the action decision latency exceeds 50% of the environment state update cycle, a reward decay coefficient is applied, forcing the network to prioritize strategies with low computational complexity. At the same time, a dynamic decision window is set. When td > T_d, the penalty for timeout decisions is strengthened through exponential decay, ensuring that the agent prioritizes low-latency actions in sudden scenarios.

[0130] Step 4 involves the construction of the INT+HDC feedback loop: using in-band network telemetry (INT) technology, environmental state changes during action execution are collected in real time through packet-level timestamp injection and path marking; when path latency fluctuations ≥15% or sudden obstacles are detected, the HDC high-latency capture mechanism is triggered, and key network packets are copied at the core switching node through a hardware-level split channel, extracting environmental change features within ≤1ms and injecting them into the STPPN input layer to update the perception flow and security boundaries.

[0131] This invention integrates four core modules: hardware-level synchronous preprocessing of multimodal sensors, lightweight spatiotemporal feature extraction, time-delay-sensitive reinforcement learning decision-making, and distributed high-speed feedback closed loop, forming an end-to-end adaptive decision-making framework. Among them, key technologies such as ASIC three-level pipeline architecture combined with convolutional kernel reorganization to achieve temporal deviation control, point cloud voxel mesh compression to preserve edge features, and GAN data repair to ensure signal-to-noise ratio are not disclosed in the prior art. This invention breaks through the limitations of existing technologies that only optimize a single aspect. It solves data timing and quality issues through hardware-level preprocessing, constructs safe decision constraints by combining Lyapunov stability theory, and achieves dynamic adaptation through high-speed feedback. This significantly improves decision security, real-time performance, and environmental adaptability, achieving a qualitative improvement compared to existing technologies. It can effectively fill the gap in intelligent agent decision-making technology in dynamic and complex scenarios, hence this patent application.

[0132] Literature search revealed patent CN202510983413.4, which discloses a technology for optimizing the temporal alignment and calibration parameters of multimodal visual data. This technology is related to the multimodal data preprocessing stage of this solution, but it lacks hardware acceleration processing and integration with the decision-making module. Patent CN201910162280.9 was also found, which discloses a software-level decision-making training technology for multi-agent systems. This technology is related to the decision-making stage of this solution, but it lacks hardware collaborative processing and a high-speed feedback mechanism. No other existing technologies with a technical solution or core innovation points more closely related to this solution were found.

[0133] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A method for training an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The method comprises the following steps: Step 1, collecting multi-modal sensor data in parallel, and performing hardware-level preprocessing on the collected data; Step 2, based on the preprocessed data in step 1, performing real-time feature extraction driven by dynamic safety boundary to obtain safety boundary coordinates; Step 3, based on the safety boundary coordinates obtained in step 2, generating an action strategy for execution within a decision cycle; Step 4, based on the action strategy generated in step 3, constructing a feedback loop of a distributed execution framework to realize real-time closed-loop optimization.

2. The method of claim 1, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific steps include: Step 1.1: Synchronously capturing multi-source sensor data of environment point cloud, RGB-D image and motion state data through laser radar, visual sensor and inertial measurement unit; Step 1.2: Performing spatio-temporal alignment and compression of the collected data at the ASIC hardware level.

3. The method of claim 2, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific method of step 1.2 is: The first stage performs timestamp calibration of sensor data, with nanosecond-level synchronization achieved through the IEEE 1588v2 precision clock protocol, the logic of which is based on a local timestamp correction function: wherein is the initial sampling time of the th sensor, and are the sending and receiving times of the message of the node, respectively, and the maximum error of cross-modal data timestamp alignment is not more than 0.3 ms through the correction. The second level deployment configurable convolution engine CCE, the point cloud path adopts The sparse convolution kernel structure, supports the dynamic voxel grid under the sampling rate , compresses the point cloud data to the original size And retains the edge features;The image path adopts With Variable kernel depth separable convolution to support image size lossless switching;The convolution kernel reorganization rule adjusts the weight coefficient according to the current sensor data sampling period and the channel time delay difference, and through setting the convolution weight cross scheduling window , dynamically interpolates and fuses multi-modal perception feature maps, so that the time difference of the final output fusion perception flow is controlled within 0.5ms. The third-level integrated residual compensation unit automatically activates the light-weight GAN module for high-frequency texture reconstruction when motion blur or illumination mutation is detected, ensuring the signal-to-noise ratio of the output perceptual stream ; the overall pipeline execution period is fixed at 5 ms.

4. The method of claim 1, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific steps of step 2 include: Step 2.1: Based on the preprocessed data in step 1, performing lightweight STFPN double-branch feature extraction; Step 2.2: First, fuse the double-branch output to form a dynamic semantic field, then use the semantic field as input to construct a risk probability model based on Lyapunov stability theory to generate a safe passage domain boundary, and at the same time construct a risk distribution function based on trajectory aggregation entropy as a supplement, finally output a safety boundary set.

5. The method of claim 4, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific method of step 2.1 is: Based on the data pretreated in step 1, a lightweight spatio-temporal feature pyramid network STFPN is deployed in the FPGA programmable logic unit, a first branch uses a 3D sparse convolutional network to perform sparse spatial sampling and residual coding on the compressed point cloud, extracts an instantaneous displacement vector field of the obstacle in a local reference coordinate system, and outputs a three-channel tensor ; a second branch loads a YOLOv5 semantic segmentation structure optimized by clipping, fuses RGB-D image information, and outputs a segmentation mask tensor .

6. The method of claim 4, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific steps of step 2.2 include: (1) The double-branch output is fused to form a dynamic semantic field ; (2) The dynamic security boundary calculation module takes as input, constructs a risk probability model based on Lyapunov stability theory, and the core mathematical expression is: ​ wherein is the spatial coordinate, is the future prediction window, represents the semantic gradient change of the current point, is the system state refresh period, is the semantic disturbance stability factor, set as ; the model is used for dynamically generating the safety passable domain boundary centered on the agent, and is only included in the low-risk area for policy matching when . (3) Constructing a risk distribution function based on trajectory aggregated entropy as a supplement: , in, For spatial location, For future time-domain prediction, Indicates the first Each sampling trajectory at time Spatial offset distance through that point For semantic field The sampling weights of the mapping For the risk diffusion radius, These are the normalization coefficients; Final output safety boundary set The boundary update frequency is 50 Hz.

7. The method of claim 1, wherein the method is a method of training a dynamic environment-aware adaptive agent decision framework. The specific steps of step 3 include: Step 3.1: Based on the safety boundary coordinates obtained in step 2, generate a time delay sensitive strategy; obtaining security boundary coordinates The agent target path input into the asynchronous advantage actor-critic A3C network, and the action policy is generated in the decision cycle. Step 3.2: Introduce a time delay sensitive reward function and a dynamic decision time window.

8. The method of claim 7, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific method of step 3.2 is: If the action decision time is delayed Exceeding the environmental status update cycle 50% of Then apply a reward decay coefficient. Simultaneously set a dynamic decision-making time window. ;when At that time, the reward function is modified to: ,in As a delay penalty factor, By setting the above parameters, the network is forced to prioritize strategies with low computational complexity, thus strengthening the agent's ability to prioritize actions from the low-latency strategy library in emergency scenarios.

9. A training method for an adaptive agent decision-making framework with dynamic environment perception capability according to step 1 of claim 1, characterized in that: The specific steps of step 4 include: Step 4.1: Based on the action strategy generated in step 3, perform real-time monitoring and capture of environmental disturbance While the agent is performing actions, collect environmental state change data, and monitor the transmission delay fluctuation of the core path and the change of new obstacle situation; Step 4.2: Inject disturbance evaluation and priority scheduling to realize real-time closed-loop optimization.

10. The method of claim 9, wherein the step 1 is a training method of an adaptive agent decision framework with dynamic environment perception capability, characterized in that: The specific steps of step 4.2 include: (1) When the INT detects path latency fluctuations or the emergence of a sudden obstacle, trigger the HDC high latency capture mechanism to copy the current critical network packets; (2) System constructs disturbance cognition priority mapping table, realizes real-time closed-loop optimization: based on feedback selection function to evaluate disturbance intensity: wherein represents the current delay change of the th path node, is its reference delay, is the latest speed vector of the node, is the reference vector, is the obstacle region identifier, is the disturbance weighting index, is the stability factor; at the same time, define the disturbance level scoring function for priority ranking, wherein is the disturbance history influence, represents the strategy entropy change rate, is the weight coefficient; The system updates the score based on the score As a result, high-priority perturbations are injected The STFPN input layer is injected into the inner injection, the perception flow is updated, and the security boundary is realized.

Images