Method and system for improving distributed cooperative control and simulation performance of unmanned system

Abstract

The invention belongs to the technical field of unmanned system cluster control and semi-physical simulation, discloses a method and a system for improving distributed cooperative control and simulation performance of an unmanned system, and aims to solve the problems of large communication delay, low synchronization precision, calculation bottleneck, model updating lagging and the like of a traditional simulation system. The method comprises the following steps: constructing a distributed system consisting of a service master station and at least one simulation middle station, and sharing a uniform address space by using an optical fiber reflection memory network (RFM); microsecond-level accurate synchronization of each node is realized through a hardware broadcast and interrupt mechanism of the RFM; all the simulation middle tables execute unmanned system dynamics resolving and cooperation tasks in parallel in a synchronization period; and the service main station is responsible for global instruction generation, semi-physical simulation closed-loop control and interaction with the large model to realize synchronous iteration of model training.

Description

Technical Field

[0001] This invention belongs to the field of unmanned system cluster control and hardware-in-the-loop simulation technology, specifically relating to a method and system for improving the distributed cooperative control and simulation performance of unmanned systems. Background Technology

[0002] With the increasing application of unmanned systems, establishing high-precision simulation and control systems to test collaborative tasks involving multiple unmanned vehicles has become increasingly important. Traditional unmanned system simulation or control system architectures typically take the following forms: First, multiple simulation modules exchange data via networks such as Ethernet; second, the system employs centralized control, with a single master node responsible for scheduling and controlling all unmanned units; third, simulation may utilize a single computer for whole-machine simulation. These traditional architectures have inherent problems, such as significant communication latency, low synchronization accuracy, excessively concentrated computational load leading to bottlenecks, and untimely model updates.

[0003] First, in terms of communication, conventional Ethernet, due to protocol overhead and jitter, struggles to meet stringent real-time requirements. Even in a local area network (LAN) environment, Ethernet communication latency is typically in the millisecond range, and this latency is unpredictable and jittery, making it difficult for simulation nodes to synchronize precisely. This communication latency and uncertainty introduce problems such as data asynchrony and command lag, severely impacting the accuracy of multi-unmanned unit collaborative simulations.

[0004] Secondly, in a centralized control architecture, all computational tasks are performed by a single server or host. This model is prone to creating computational bottlenecks when dealing with complex simulations. The processing power of a single simulation host is limited. When the simulation scenario includes high-fidelity environment models, dynamic models of multiple unmanned devices, and complex control algorithms, a single machine often struggles to complete all computational tasks in a timely manner. This leads to slow or non-real-time simulation operation, limiting the accuracy and scale of the simulation.

[0005] Furthermore, traditional simulation systems primarily rely on software clocks or network protocols (such as NTP and PTP) for time synchronization, which often lacks sufficient accuracy in multi-node collaboration. If there are deviations or delays in the clocks of different nodes, the state updates of the unmanned systems during the simulation cannot be strictly synchronized, leading to a decline in the performance of the collaborative control algorithm. Poor synchronization accuracy also causes sensor / actuator data to not correspond simultaneously, reducing the reliability of the simulation.

[0006] Finally, untimely model updates are also a problem in traditional systems. For example, in the field of unmanned autonomous control, the training and deployment of new algorithms or models takes a long time, often resulting in a disconnect between simulation and the real environment. Centralized or stand-alone simulation platforms often cannot efficiently train and update models while simulating, causing the models used in the simulation to lag behind the latest data and reducing the realism of the simulation.

[0007] In summary, existing technologies suffer from drawbacks such as large communication delays, low synchronization accuracy, significant computational bottlenecks, and poor model training and update efficiency. A new approach is urgently needed to improve the performance of distributed collaborative simulation and control of unmanned systems. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for improving the distributed cooperative control and simulation performance of unmanned systems, aiming to achieve cooperative operation and synchronous control of large-scale unmanned equipment (such as drones, unmanned ships, unmanned vehicles, underwater vehicles, etc.), and to solve the problems of large communication delay, low synchronization accuracy, prominent computational bottlenecks, and poor model training and update efficiency in existing technologies.

[0009] This invention provides a method for improving the distributed cooperative control and simulation performance of unmanned systems, comprising the following steps:

[0010] S1. System Construction and Initialization;

[0011] The distributed system includes a central service console and at least one simulation middleware console, wherein each node is equipped with a reflective memory card and shares a unified address space through a reflective memory network;

[0012] S2. The service desk and the central office are synchronized;

[0013] The service console periodically writes a synchronization flag into the preset synchronization register, triggering the reflective memory card to generate a soft interrupt to each simulation platform. After receiving the soft interrupt, each simulation platform calibrates its local clock to achieve microsecond-level synchronization.

[0014] S3. Distributed collaborative computing;

[0015] Each simulation platform executes unmanned system dynamics calculations, sensor simulations, and control algorithms within a synchronization cycle, writes the generated state variables and control commands into its corresponding shared address segment, and reads the data written by other nodes to complete multi-unmanned system collaboration.

[0016] S4. Global command generation and issuance;

[0017] The central service console generates or updates global control instructions based on the data aggregated in the shared memory and writes them into the shared address segment for each simulation platform to subscribe to and execute.

[0018] S5. Simulation closed loop;

[0019] The service console sends the simulated attitude commands to the five-axis turntable and satellite signal simulator via the hardware-in-the-loop interface to drive the real hardware movement and signal transmission, and collects its feedback data and writes it into the shared address segment to form a closed loop;

[0020] S6. Simulation and model training are iterated synchronously;

[0021] The service console interacts with the large model interface to perform federated learning or hybrid parallel training on the simulation data, obtain updated model parameters, and periodically write them to the shared address segment to refresh the local models of each simulation platform, thereby achieving synchronous iteration of simulation and model training.

[0022] S7. End and data recording;

[0023] When the preset simulation duration or task completion conditions are reached, the service center records the flight logs, model weights, and performance indicators of each unmanned system, sends a simulation end synchronization flag, and the system is safely shut down.

[0024] Furthermore, in step S1, the reflective memory network is a fiber optic ring topology with a single-node bandwidth of not less than 1Gb / s and a single-hop data transmission delay between nodes of not more than 5μs.

[0025] Furthermore, in step S2, the specific method for generating a soft interrupt by writing the synchronization flag to the preset address of the synchronization register is as follows: the service console writes a synchronization byte to the shared address, and the reflection memory card of each simulation console detects that the preset address has been written and triggers a high-priority interrupt on the local host.

[0026] Furthermore, the shared address space in step S1 is divided into environmental data segment, unmanned system status segment, control instruction segment and model parameter segment according to data type.

[0027] Furthermore, in step S5, the five-axis turntable has three rotational degrees of freedom and two translational degrees of freedom. The maximum angular velocity of the five-axis turntable is not less than 300° / s, and the maximum angular acceleration is not less than 2000° / s. 2 .

[0028] Furthermore, in step S5, the satellite signal simulator can generate GPS, BeiDou, or pseudo-satellite radio frequency signals, and the update frequency of the satellite signal simulator is consistent with the simulation synchronization cycle.

[0029] Furthermore, in step S6, each simulation platform independently performs several training rounds based on local simulation data and then reports the gradient or model difference to the service platform. The service platform uses weighted average aggregation to obtain the global model and broadcasts it back to each simulation platform.

[0030] The aggregation adopts a federated averaging algorithm, and the aggregation period can be adaptively adjusted between 1 and 10 simulation synchronization periods according to the communication load.

[0031] Furthermore, in step S6, the hybrid parallel training adopts a combination of data parallelism and model parallelism. Different simulation platforms calculate the gradients of different data batches in parallel, and the model is divided into layers and assigned to different computing nodes.

[0032] Furthermore, in step S6, the service console monitors the computing load and communication latency of each simulation platform. When a load imbalance is detected, the synchronization cycle is adjusted or training tasks are redistributed to maintain system real-time performance.

[0033] This invention also provides a system for improving the distributed cooperative control and simulation performance of unmanned systems, comprising:

[0034] At least one unmanned device constitutes the equipment and perception layer;

[0035] At least one middleware server constitutes the management and execution layer;

[0036] A central service desk constitutes the decision-making and service layers;

[0037] A wide area control network is used to establish a remote communication link between the central server and the unmanned equipment;

[0038] A reflective memory network is used to establish a high-speed synchronous network between the various middleware servers within the management and execution layer, and between the middleware servers and a hardware-in-the-loop simulation system;

[0039] The middleware server is configured to synchronize key attitude data collected from the unmanned device to the hardware-in-the-loop simulation system in real time via the reflective memory network.

[0040] Beneficial effects:

[0041] This invention integrates distributed computing, network communication, RFM, a five-axis rotary table simulation system, and hardware-in-the-loop simulation technology, and is applicable to task planning, attitude control, real-time status feedback, and improved simulation training efficiency for intelligent unmanned systems. Through a multi-layered architecture and efficient distributed methods, this invention further optimizes the training and simulation of large-scale unmanned equipment, achieving accurate reproduction of equipment states and efficient collaboration in complex tasks.

[0042] This invention introduces a reflective memory network, a technical solution for achieving high-speed, real-time data synchronization among multiple computing nodes. Reflective memory distributes data through shared memory; when one node writes to its local RFM memory, the data is broadcast to all other nodes at the hardware level with extremely low latency, achieving consistent data updates across all nodes. This hardware-level data synchronization features microsecond-level transmission latency and deterministic timing characteristics (in contrast, Ethernet communication latency is typically millisecond-level with significant jitter). Due to its advantages of high speed, low latency, and deterministic transmission time, reflective memory technology has been widely used in aerospace, industrial control, military simulation, and other fields with extremely high real-time requirements. However, most previous unmanned system simulation / control platforms have not fully utilized advanced technologies such as reflective memory. In traditional HLA (High-Level Architecture) distributed simulation frameworks, the advantages of high-performance computing clusters and reflective memory networks cannot be fully utilized because the underlying communication is limited by ordinary networks. Therefore, introducing reflective memory networks into unmanned system collaborative simulation is expected to significantly improve communication synchronization and real-time performance.

[0043] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall distributed collaborative control of a multi-unmanned system.

[0045] Figure 2 This is a timing diagram for clock synchronization based on RFM;

[0046] Figure 3 This is a schematic diagram of a hardware-in-the-loop closed-loop simulation.

[0047] Figure 4 This provides an example of hardware-in-the-loop deployment in multiple scenarios. Detailed Implementation

[0048] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0049] like Figure 1As shown, this invention provides a method for improving the distributed cooperative control and simulation performance of unmanned systems, comprising the following steps:

[0050] S1. System Construction and Initialization;

[0051] Build as Figure 1 The distributed system shown has a central service console and N simulation substations, each equipped with a high-performance computer and reflective memory cards. All nodes' reflective memory cards are connected via fiber optic cables in a ring or star topology, establishing a globally shared address space. This address space is logically divided into multiple segments, such as environmental data segments, unmanned system status segments, global control command segments, and model parameter segments, to optimize access efficiency and reduce bus contention.

[0052] The reflective memory network is a fiber optic ring topology with a single-node bandwidth of no less than 1Gb / s and a single-hop data transmission delay between nodes of no more than 5μs.

[0053] S2. The service desk and the central office are synchronized;

[0054] like Figure 2 As shown, the service console, acting as the time master, writes a synchronization count value to a predefined synchronization register address in the RFM network at a fixed high-frequency period (e.g., Δt = 1ms). This write operation is broadcast to the reflection memory cards of all simulation stations via the RFM network layer at hardware speed (latency typically <5μs). Each simulation station's reflection memory card hardware detects the write to this address and immediately triggers a high-priority interrupt to its host CPU. The simulation station's interrupt service routine is triggered, immediately reads the synchronization count value, and records the local timestamp t. i The error ε between the time t0 written by the service center and the actual time t0 is calculated. i =t i -t0 calibrates the local clock. This process is repeated cyclically to ensure that the computing clocks of all nodes are synchronized with the central server at the microsecond level.

[0055] S3. Distributed collaborative computing;

[0056] Within each synchronization cycle, each simulation platform executes computational tasks in parallel. Taking a simulation platform responsible for a UAV as an example, it solves the UAV's dynamic equations and simulates sensor data (such as GPS and IMU) based on control commands read from shared memory and its own state. The newly generated state variables and sensor data are written to its dedicated address segment in shared memory in real time. Simultaneously, it can also read state data written by other simulation platforms (such as nodes responsible for unmanned vehicles) in real time to achieve complex collaborative tasks such as air-to-ground coordination and obstacle avoidance.

[0057] S4. Global command generation and issuance;

[0058] During each synchronization cycle, the service console aggregates status data uploaded by all simulation platforms from shared memory. Based on this global information, the service console executes top-level task planning or global optimization algorithms to generate or update global control commands (such as target waypoints, formation configurations, etc.). These commands are written into the global control command segment in shared memory for all simulation platforms to subscribe to and execute.

[0059] S5. Hardware-in-the-loop simulation;

[0060] Single closed loop (reference) Figure 3 The service console sends simulated attitude commands for a specific UAV to the corresponding simulation platform via shared memory. The platform then transmits these commands to a five-axis turntable via a hardware interface, driving the real IMU hardware installed on it to accurately reproduce the motion. The five-axis turntable has three rotational degrees of freedom and two translational degrees of freedom, with a maximum angular velocity of at least 300° / s² and a maximum angular acceleration of at least 2000° / s². 2 Simultaneously, the service console, based on the simulated location, drives the satellite simulator to generate corresponding radio frequency signals. The satellite signal simulator can generate GPS, BeiDou, or pseudo-satellite radio frequency signals, and its update frequency is consistent with the simulation synchronization cycle. The simulation platform collects the output of the real IMU and the data calculated by the navigation receiver, writing it back to shared memory as "semi-physical device output," forming a physical closed loop for attitude and navigation. (Extended architecture - see reference) Figure 4 This system supports multiple parallel, heterogeneous hardware-in-the-loop simulation platforms. The central service console can simultaneously coordinate multiple different simulation platforms, such as an unmanned vehicle, an unmanned drone, and an unmanned boat. Each platform is managed by its corresponding simulation middleware (lower-level machine) and executes its own closed-loop simulation tasks, while the central service console is responsible for upper-level collaborative control and task allocation.

[0061] S6. Simulation and model training are iterated synchronously;

[0062] While executing simulation tasks, each simulation platform accumulates a large amount of simulation interaction data (states, actions, rewards, etc.). Within a preset training period, each platform independently calculates the gradient or model difference of the model parameters using local data. The central service platform aggregates gradients through the "large model interface" using algorithms such as federated averaging and updates the global model. The model parameters are then broadcast back to all simulation platforms via a reflective memory network. The simulation platforms load the new model parameters for subsequent simulation and control.

[0063] S7. End and data recording;

[0064] When the simulation reaches the preset duration or the task completion condition, the service center sends an end signal. Before shutdown, it is responsible for archiving and saving key data in shared memory, such as complete flight logs of each unmanned system, final model weights, and performance evaluation metrics. Subsequently, all nodes safely exit.

[0065] Specifically, a central service platform S0 and N simulation middleware platforms S are introduced. i (i = 1, ..., N). Each node is equipped with a reflective memory card and forms a unified shared memory through a fiber optic ring topology, forming an address set A = {a0, ..., a...}. M-1 The service console writes to the synchronization register address a periodically at intervals Δt. sync A 32-bit counter value c(t). Let node S be... i At local time t i If a synchronization interrupt is generated and the count value is read, then the synchronization error is:

[0066] ε i =t i -t0 (if |ε i |≤ε sync (1)

[0067] Where t0 is the write time of the service console; ε i Represents node S i Synchronization error; ε sync This indicates the allowable range of synchronization error. Each node executes t. i ←t i -ε i Complete clock calibration. Define the memory partitioning table:

[0068]

[0069] In the formula, A env A represents the memory area where environmental situation data is stored; uav,k This represents the data area storing the status of unmanned systems; U represents the number of unmanned systems, and k corresponds to different unmanned systems; A cmd This represents the memory area where control instruction data is stored; A model This represents the memory area where model parameter data is stored.

[0070] The system stores environmental conditions, unmanned system status, control commands, and model parameters respectively. The write-broadcast-refresh process is entirely handled by RFM hardware, requiring no software protocol stack.

[0071] Each middle platform S j Solve the dynamic equations of the unmanned system under the responsibility of this node based on the distance step length Δt:

[0072]

[0073] And generate sensor observations:

[0074]

[0075] Where x is the state vector, u is the control vector, and v is the noise. This represents the state vector of the i-th unmanned system at time n+1 in the k-th simulation platform; This represents the control vector of the i-th unmanned system at time n in the k-th simulation platform; This represents the sensor observation value of the i-th unmanned system at time n in the k-th simulation platform; This represents the sensor observation function, with the input being the state vector at time n. The output is noise-free observations; v k [n+1] represents the sensor observation noise.

[0076] The result is written to A. uav,k Other nodes read data in real time to achieve collaboration. The service center starts from A. uav,k Read attitude R k With angular velocity ω k Through motion mapping Drive the five-axis rotary table to execute attitude commands and acquire the real IMU output z k Write back to A uav,k Feedback loop. The service center calculates the pseudorange of navigation satellites:

[0077] ρ s,k =‖p s (t)-p k (t)‖+c(dT s -dT k (5)

[0078] Where, p s (t) represents the satellite's position vector at time t; p k (t) represents the position vector of the unmanned system at time t; c represents the speed of light; dT s dT represents satellite clock bias; k Indicates the clock bias of unmanned systems;

[0079] The satellite simulator generates RF signals in real time; the drone receiver measures the pseudorange and writes it into shared memory for state estimation.

[0080] During the online model training phase, each simulation platform completes its local gradient within a period τ. Front Desk Executive:

[0081]

[0082] Among them, w (t+1) represents the model parameters for the (t+1)th iteration; η represents the learning rate; This represents the local gradient calculated by the i-th simulation platform in the t-th iteration;

[0083] The headquarters will also w (t+1) Write A model When the model dimension is greater than the single-node memory capacity, it is divided into layers {L1,...,L...}. P Distribute to multiple nodes and use data parallel batching B:

[0084]

[0085] Where B represents the batch size during parallel training of data;

[0086] All node gradients are synchronized via RFM, parameters are updated, and broadcast. The service center monitors node S. i Processing delay T i Communication queuing delay δ i If detected:

[0087]

[0088] Among them, T i Represents node S i Processing latency; δ i Represents node S i Communication queuing delay; Represents the (T) of all nodes (N nodes in total) j +δ j The average value of )

[0089] This triggers a migration strategy, remapping some simulated entities or training subtasks to nodes with low load in order to maintain a stable frame rate.

[0090] This invention also provides a system for improving the distributed cooperative control and simulation performance of unmanned systems, comprising:

[0091] At least one unmanned device constitutes the equipment and perception layer;

[0092] At least one middleware server constitutes the management and execution layer;

[0093] A central service desk constitutes the decision-making and service layers;

[0094] A wide area control network is used to establish a remote communication link between the middleware server and the unmanned equipment;

[0095] A reflective memory network is used to establish a high-speed synchronous network between the various middleware servers within the management and execution layers, as well as between the middleware servers and a hardware-in-the-loop simulation system.

[0096] The middleware server is configured to synchronize key attitude data collected from unmanned equipment to the hardware-in-the-loop simulation system in real time via a reflective memory network.

[0097] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for improving the distributed cooperative control and simulation performance of unmanned systems, characterized in that, The following steps are involved: S1. System Construction and Initialization; The distributed system includes a central service console and at least one simulation middleware console, wherein each node is equipped with a reflective memory card and shares a unified address space through a reflective memory network; S2. The service desk and the central office are synchronized; The service console periodically writes a synchronization flag into the preset synchronization register, triggering the reflective memory card to generate a soft interrupt to each simulation platform. After receiving the soft interrupt, each simulation platform calibrates its local clock to achieve microsecond-level synchronization. S3. Distributed collaborative computing; Each simulation platform executes unmanned system dynamics calculations, sensor simulations, and control algorithms within a synchronization cycle, writes the generated state variables and control commands into its corresponding shared address segment, and reads the data written by other nodes to complete multi-unmanned system collaboration. S4. Global command generation and issuance; The central service console generates or updates global control instructions based on the data aggregated in the shared memory and writes them into the shared address segment for each simulation platform to subscribe to and execute. S5. Simulation closed loop; The service console sends the simulated attitude commands to the five-axis turntable and satellite signal simulator via the hardware-in-the-loop interface to drive the real hardware movement and signal transmission, and collects its feedback data and writes it into the shared address segment to form a closed loop; S6. Simulation and model training are iterated synchronously; The service console interacts with the large model interface to perform federated learning or hybrid parallel training on the simulation data, obtain updated model parameters, and periodically write them to the shared address segment to refresh the local models of each simulation platform, thereby achieving synchronous iteration of simulation and model training. S7. End and data recording.

2. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S1, the reflective memory network is a fiber optic ring topology with a single node bandwidth of not less than 1Gb / s and a single-hop data transmission delay between nodes of not more than 5μs.

3. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that, In step S2, the specific method for generating a soft interrupt by writing the synchronization flag to the preset address of the synchronization register is as follows: the service center writes a synchronization byte to the shared address, and the reflection memory card of each simulation center detects that the preset address has been written and triggers a high-priority interrupt on the local host.

4. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S1, the shared address space is divided into environmental data segment, unmanned system status segment, control command segment, and model parameter segment according to data type.

5. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S5, the five-axis turntable has three rotational degrees of freedom and two translational degrees of freedom. The maximum angular velocity of the five-axis turntable is not less than 300° / s, and the maximum angular acceleration is not less than 2000° / s. 2 .

6. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S5, the satellite signal simulator can generate GPS, BeiDou or pseudo-satellite radio frequency signals, and the update frequency of the satellite signal simulator is consistent with the simulation synchronization cycle.

7. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S6, each simulation platform independently performs several training rounds based on local simulation data and then reports the gradient or model difference to the service center. The service center uses weighted average aggregation to obtain the global model and broadcasts it back to each simulation platform.

8. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S6, the hybrid parallel training adopts a combination of data parallelism and model parallelism. Different simulation platforms calculate the gradients of different data batches in parallel, and the model is divided into layers and assigned to different computing nodes.

9. The method for identifying hierarchical influence nodes in multi-agent networks based on triangular structure recursive compression according to claim 1, characterized in that: In step S6, the service console monitors the computing load and communication latency of each simulation platform. When a load imbalance is detected, the synchronization cycle is adjusted or training tasks are redistributed to maintain system real-time performance.

10. A system for improving the distributed cooperative control and simulation performance of unmanned systems, characterized in that, include: At least one unmanned device constitutes the equipment and perception layer; At least one middleware server constitutes the management and execution layer; A central service desk constitutes the decision-making and service layers; A wide area control network is used to establish a remote communication link between the central server and the unmanned equipment; A reflective memory network is used to establish a high-speed synchronous network between the various middleware servers within the management and execution layer, and between the middleware servers and a hardware-in-the-loop simulation system; The middleware server is configured to synchronize key attitude data collected from the unmanned device to the hardware-in-the-loop simulation system in real time via the reflective memory network.

Images