A brain-computer shared control method and system for air-ground collaborative unmanned systems

Abstract

The present invention discloses a brain-computer shared control method and system for an air-ground collaborative unmanned system, including: step S1, constructing a mobile robot model; step S2, constructing an MPC shared controller based on the mobile robot model; step S3, constructing the action space, reward function and neural network model of the DDPG algorithm; step S4, training the neural network model based on DDPG to obtain an optimized model; step S5, adjusting the control authority parameters of the MPC-based shared controller according to the optimized model to control the robot to complete the specified task. The technical solution of the present invention can adjust the authority of the auxiliary controller and the driver under different working conditions and different states of the driver, thereby realizing the driver's control intention while ensuring the safety of the air-ground collaborative multi-robot system.

Description

Technical Field

[0001] The present invention belongs to the field of automatic control technology, and specifically relates to a brain-computer shared control method and system for an air-ground collaborative unmanned reinforcement system. Background Art

[0002] How to improve the search efficiency in non-open areas is an important research topic for search and rescue. Micro drones are small in size, light in weight, and have strong reconnaissance capabilities. They can play their full role in relatively dangerous reconnaissance or search missions; ground robots are stealthy, can carry a variety of equipment, and can replace personnel to enter dangerous environments. Therefore, the air-ground collaborative system can well replace staff to enter dangerous areas.

[0003] However, most of the current research on air-ground collaboration focuses on the control of mobile robots in simple environments, either completely autonomous or completely non-autonomous. There is no in-depth exploration of shared control methods for humans and intelligent auxiliary controllers in complex and changing scenarios, nor is there any exploration of the interference caused by multi-tasking on human control. When controlling an air-ground collaborative multi-robot system, the accuracy of the operation is seriously reduced due to the multi-tasking of the operator, which has a great impact on the safety level of the unmanned system. In summary, the current problems in the field of air-ground collaborative system control technology are: 1) The air-ground collaborative system requires the cooperation of multiple people, and the collaborative efficiency is low; 2) It is difficult for robots to cooperate with each other when a single driver performs multi-tasking operations; 3) When the driver operates the mobile robot in multiple tasks, his attention will be distracted, resulting in a decrease in command accuracy; 4) The traditional shared control method does not consider the impact of driver state changes on system safety.

[0004] To sum up, the difficulty in solving the above problems lies in: introducing a new operation mode of brain-computer interface to improve the driver's ability to operate complex air-ground collaborative systems; designing a shared controller to improve the safety of the driver when operating complex systems; considering changes in the environment and driver status, adaptively adjusting the control authority of the auxiliary controller to ensure the stability of the intelligence level of the complex system and safe operation. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a brain-computer shared control method and system for air-ground collaborative unmanned systems, which can adaptively adjust the control authority of the auxiliary controller and the driver in real time according to different working conditions and driver status based on reinforcement learning strategies.

[0006] To achieve the above object, the present invention adopts the following scheme:

[0007] A brain-computer shared control method for air-ground collaborative unmanned systems, comprising:

[0008] Step S1, constructing a mobile robot model;

[0009] Step S2: constructing an MPC shared controller based on the mobile robot model;

[0010] Step S3, constructing the action space, reward function and neural network model of the DDPG algorithm;

[0011] Step S4: training a DDPG-based neural network model to obtain an optimized model;

[0012] Step S5: adjusting the control authority parameters of the MPC-based shared controller according to the optimization model to control the robot to complete the specified task.

[0013] As a preferred embodiment, the mobile robot model adopts a two-wheel differential mobile robot model, and the robot state quantity ξ=[xy θ] T , state quantity u=[v ω] T , the kinematic model of the mobile robot is expressed as:

[0014]

[0015] Among them, x and y represent the positions of the mobile robot in the inertial coordinate system, and θ is the current heading angle of the mobile robot.

[0016] Preferably, the optimization function of the MPC shared controller is set to:

[0017] minJ(x(k),U k )

[0018] u min ≤u(k)≤u max ,i=0,...,H C -1

[0019] Δu(k+i)=u(k+i)-u(k+i-1), i=0,...,H C -1

[0020] Δu min ≤Δu(k)≤Δu max ,i=0,...,H C -1

[0021] y min ≤y b (k)≤y max ,i=0,...,H P -1

[0022] Δu(k+i)=0,i=H C ,H C +1,...,H P -1

[0023] The objective function can be set as follows:

[0024]

[0025] In the above cost function, k represents the current moment, U(k) represents the brain-computer interface input at the current moment, x(k) represents the system state at the current moment, and H p and H c represent the prediction time domain and the control time domain respectively, and H C ≤H P , α represents the control authority allocation, and α~clip(0,1). and are the weight matrices for the penalty output error, penalty control action, and penalty control change rate, respectively. r(·) is the controller reference input, u r (·) is the corresponding control reference input.

[0026] As a preferred method, the mobile robot is controlled by the controller model to complete the exploration task, wherein the controller model control signal corresponds to the control command generated by the brain-computer interface, command X = {x su ,x sd ,x tl ,x tr ,x ms ,x stop}, where the commands represent acceleration, deceleration, turn left, turn right, maintain speed, and stop;

[0027] Let I C ∈{1,2,3,4,5,6} represents the intention command decoded from the user's EEG signal, where 1 is acceleration, 2 is deceleration, 3 is left turn, 4 is right turn, 5 is maintain speed, and 6 is stop. The following interface model is designed as follows:

[0028]

[0029] in, Respectively represent the linear velocity and angular velocity commands output by the interface model at time k, represents the user intention output by the SSVEP-based BCI at time k, v max ,v min , and Δv represent the maximum linear velocity, minimum linear velocity, and linear velocity gain, respectively, ω min and ω max represents the limit of angular velocity, and Δω is the angular velocity gain.

[0030] The present invention provides a brain-computer sharing control system for an air-ground collaborative unmanned system, comprising:

[0031] A first building module is used to build a mobile robot model;

[0032] A second building module is used to build an MPC shared controller based on the mobile robot model;

[0033] The second building block is used to construct the action space, reward function and neural network model of the DDPG algorithm;

[0034] The training module is used to train the DDPG-based neural network model to obtain the optimized model;

[0035] The control module is used to adjust the control authority parameters of the MPC-based shared controller according to the optimization model to control the robot to complete the specified task.

[0036] As a preferred embodiment, the mobile robot model adopts a two-wheel differential mobile robot model, and the robot state quantity ξ=[xy θ] T , state quantity u=[v ω] T , the kinematic model of the mobile robot is expressed as:

[0037]

[0038] Among them, x and y represent the positions of the mobile robot in the inertial coordinate system, and θ is the current heading angle of the mobile robot.

[0039] As a preference, the optimization function of the MPC shared controller is set to:

[0040] minJ(x(k),U k )

[0041] u min ≤u(k)≤u max ,i=0,...,H C -1

[0042] Δu(k+i)=u(k+i)-u(k+i-1), i=0,...,H C -1

[0043] Δu min ≤Δu(k)≤Δu max ,i=0,...,H C -1

[0044] y min ≤y b (k)≤y max ,i=0,...,H P -1

[0045] Δu(k+i)=0,i=H C ,HC +1,...,H P -1

[0046] The objective function can be set as follows:

[0047]

[0048] In the above cost function, k represents the current moment, U(k) represents the brain-computer interface input at the current moment, x(k) represents the system state at the current moment, and H p and H c represent the prediction time domain and the control time domain respectively, and H C ≤H P , α represents the control authority allocation, and α~clip(0,1). and are the weight matrices for the penalty output error, penalty control action, and penalty control change rate, respectively. r(·) is the controller reference input, u r (·) is the corresponding control reference input.

[0049] As a preferred method, the mobile robot is controlled by the controller model to complete the exploration task, wherein the controller model control signal corresponds to the control command generated by the brain-computer interface, command X = {x su ,x sd ,x tl ,x tr ,x ms ,x stop}, where the commands represent acceleration, deceleration, turn left, turn right, maintain speed, and stop;

[0050] Let I C ∈{1,2,3,4,5,6} represents the intention command decoded from the user's EEG signal, where 1 is acceleration, 2 is deceleration, 3 is left turn, 4 is right turn, 5 is maintain speed, and 6 is stop. The following interface model is designed as follows:

[0051]

[0052] in, Respectively represent the linear velocity and angular velocity commands output by the interface model at time k, represents the user intention output by the SSVEP-based BCI at time k, v max ,v min , and Δv represent the maximum linear velocity, minimum linear velocity, and linear velocity gain, respectively, ω min and ω max represents the limit of angular velocity, and Δω is the angular velocity gain.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1) The shared control method for air-ground collaborative robot system proposed in this application can improve the reconnaissance and exploration level of unmanned systems, improve the safety of mission execution, speed up system response, expand the reconnaissance range, and reduce human resource consumption;

[0055] 2) The shared control method for air-ground collaborative robot system proposed in this application can adjust the human-machine control authority parameters in the shared controller to cope with different situations. It can not only make up for the shortcomings of artificial intelligence in solving edge problems, but also solve the safety problems of operators in situations such as fatigue and distraction. The unmanned system can meet the control intention of the controller while ensuring the safety requirements of the unmanned system;

[0056] 3) The shared control method for the air-ground collaborative robot system proposed in this application integrates deep reinforcement learning and model prediction, and establishes a shared control method that can generate different response strategies for different situations. Compared with traditional control methods, it has better robustness and richer response strategies;

[0057] 4) The layout method proposed in this application is highly versatile and can be applied to air-ground collaborative multi-robot unmanned systems, improving the intelligence level of current unmanned systems while reducing the workload of operators. It has obvious application value and prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solution of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0059] Figure 1 This is a flow chart of a brain-computer sharing control method for an air-to-ground collaborative unmanned system according to an embodiment of the present invention;

[0060] Figure 2 Schematic diagram of the training process based on the DDPG algorithm in a virtual scene in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0062] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0063] Embodiment 1:

[0064] like Figure 1 As shown, an embodiment of the present invention provides a brain-computer shared control method for an air-ground collaborative unmanned system. For an air-ground collaborative multi-robot unmanned system, the operator is often unable to operate a complex robot system at the same time, so brain-computer interface control technology is introduced. However, the operator often has distraction problems when observing the environment, thereby reducing the safety level of the unmanned system. This method is based on a deep reinforcement learning method to improve the model prediction shared controller, and uses the DDPG algorithm to adjust the human-machine control authority of the shared controller, which not only ensures that the unmanned system complies with the controller's intention, but also ensures the overall safety of the unmanned system. It includes the following steps:

[0065] Step S1, constructing a mobile robot model;

[0066] Step S2: constructing an MPC shared controller based on the mobile robot model;

[0067] Step S3, constructing the action space, reward function and neural network model of the DDPG algorithm;

[0068] Step S4: training a DDPG-based neural network model to obtain an optimized model;

[0069] Step S5: adjusting the control authority parameters of the MPC-based shared controller according to the optimization model to control the robot to complete the specified task.

[0070] As an implementation mode of the present invention, step S1 includes:

[0071] Mathematically model the task area, establish the environment model, and initialize the environment information, which is the source of environmental impact on the mobile robot. The simulation environment is based on the Gazebo software under the ROS system. Gazebo is an open source multi-robot simulator that is widely used for simulation and experiments in the field of robotics. It provides a highly customizable environment that allows users to simulate and test a variety of robots, sensors, environments, and behaviors;

[0072] The mobile robot model adopts a two-wheel differential mobile robot model, and defines the posture state of the mobile robot in the environment as:

[0073] ξ=[xy θ] T

[0074] Among them, x and y represent the positions of the mobile robot in the inertial coordinate system, and θ is the current heading angle of the mobile robot.

[0075] Define the actual running speed of the robot as:

[0076] u=[vω] T (1.1)

[0077] The speed of the input drive wheel is defined as:

[0078] u drive =[v l v r ] T (1.2)

[0079] According to the constraints of the mobile robot, the following kinematic equations can be obtained:

[0080]

[0081] From this, the kinematic model of the mobile robot can be obtained as follows:

[0082]

[0083] The kinematic equation of the mobile robot with differential drive speed as input can be obtained:

[0084]

[0085] For an actual robot system, it is necessary to determine which type of control variable to use. If the control input is the actual speed u of the robot, then (1.4) is used as the kinematic equation. If the control input is the differential drive speed u drive Then (1.5) is used as the kinematic equation. Combined with the actual situation of the mobile robot in this laboratory, the actual speed of the robot is used as the control quantity, and equation (1.4) is expressed in a more general form:

[0086]

[0087] Among them, the robot state ξ=[xy θ] T , state quantity u=[v ω] T Combining equations (1.1) and (1.4), the kinematic model of the mobile robot can be converted into the following form:

[0088]

[0089] As an implementation of the embodiment of the present invention, step S2 includes:

[0090] S21. Discretize the mobile robot model

[0091] The discrete-time model defining the nonlinear system is expressed as follows:

[0092] x(k+1)=f(x(k),u(k)) (2.1)

[0093] y c (k) = g c (x(k),u(k)) (2.2)

[0094] y b (k) = g b (x(k),u(k)) (2.3)

[0095] Among them, k is the current time of the system, is the state quantity, To control the input, To control the output, is the constraint output. Define the constraint expression as

[0096] u min ≤u(k)≤u max (2.4)

[0097] Δu min ≤Δu(k)≤Δu max (2.5)

[0098] y min ≤y b (k)≤y max (2.6)

[0099] Among them, equations (2.4) and (2.5) are the control quantity and control increment constraints, and the specific boundary values ​​are mainly set according to the limits of the actuator;

[0100] S22. Constructing the optimization function of the model predictive controller

[0101] The optimization function of the model predictive controller is set as follows:

[0102] minJ(x(k),U k )(2.7)st

[0103] u min ≤u(k)≤u max ,i=0,...,H C -1 (2.8)

[0104] Δu(k+i)=u(k+i)-u(k+i-1), i=0,...,H C -1 (2.9)

[0105] Δumin ≤Δu(k)≤Δu max ,i=0,...,H C -1 (2.10)

[0106] y min ≤y b (k)≤y max ,i=0,...,H P -1 (2.11)

[0107] Δu(k+i)=0,i=H C ,H C +1,...,H P -1 (2.12)

[0108] Among them, the objective function can be set as follows:

[0109]

[0110] In the above cost function, k represents the current moment, H p and H c represent the prediction time domain and the control time domain respectively, and H C ≤H P , α represents the control authority allocation, and α~clip(0,1). and are the weight matrices for the penalty output error, penalty control action, and penalty control change rate, respectively. r(·) is the controller reference input, u r (·) is the corresponding control reference input.

[0111] As an implementation of the embodiment of the present invention, step S3 includes:

[0112] S31. The action space of the DDPG algorithm is constructed as α, corresponding to equation (2.13), which represents the control authority allocation of the shared control algorithm, and α ~ clip (0, 1);

[0113] S32 constructs the DDPG algorithm environment state. The algorithm environment state space is represented by s=[U risk ,U conflict ,U distract i on ], where U risk represents the risk state function, U conflict represents the human-machine conflict state function, U distraction represents the driver distraction state function;

[0114] S33 constructs the reward function of the DDPG algorithm. The reward function of the algorithm is expressed as r = r risk +rconflict +r task , where r risk =-W r U risk , r conflict =-W c U conflict , r task =-W t t, penalty factor W r , W c , W t They correspond to robot risk, human-robot conflict, and task completion efficiency respectively;

[0115] S34. Constructing a neural network model

[0116] Construct two fully connected feedforward neural networks with N_Layer hidden layers and N_Num neurons in each layer, which serve as the online Actor network and the online Critic network respectively. Initialize the node parameters π(s|θ π ) and the node parameters Q(s,a|θ) of the online critic network Q ), using ReLu() as the activation function, setting the network learning rate to LR, and using the error back propagation and gradient descent method to update the network node parameters. In order to make the training more stable, two target actor networks and target critic networks with the same structure and node parameters as the online actor network and the online critic network, but with different update methods, are constructed, namely π′(s|θ π ′)←π(s|θ π ), Q′(s,a|θ′ Q )←Q(s,a|θ Q ). The online network is used as the real-time network and updated in real time through error back propagation and gradient descent during training. The target network is soft-updated by copying some node parameters of the online network.

[0117] As an implementation method of the embodiment of the present invention, Figure 2 As shown, step S4 includes:

[0118] S41. Initialize the online network target network parameters and training environment, the initial node parameters θ π ′=θ π ,θ′ Q =θ Q ;

[0119] S42. The controller model controls the mobile robot to complete the exploration task, where the controller model control signal corresponds to the control command generated by the brain-computer interface, command X = {x su ,x sd,x tl ,x tr ,x ms ,x stop}, where the commands represent acceleration, deceleration, turn left, turn right, maintain speed, and stop.

[0120] Let I C ∈{1,2,3,4,5,6} represents the intention command decoded from the user's EEG signal, where 1 is acceleration, 2 is deceleration, 3 is left turn, 4 is right turn, 5 is maintain speed, and 6 is stop. The following interface model is designed:

[0121]

[0122] in, They represent the linear velocity and angular velocity commands output by the interface model at time k respectively. represents the user intention output by the SSVEP-based BCI at time k. max ,v min , and Δv represent the maximum linear velocity, minimum linear velocity, and linear velocity gain, respectively. ω min and ω max represents the limit of angular velocity, Δω is the angular velocity gain;

[0123] S43. Experience storage and update

[0124] When using the DDPG method as a decision framework, it is necessary to design a memory bank that can store state transitions. The Actor network and the Critic network are trained by extracting experience from different time nodes in the past. Assuming that the decision cycle is δ seconds, the action remains unchanged during the decision cycle time. In the round, the sensor observes the environment and the robot state as s at time t. t , select action a according to the online Actor network t After executing the action, the environment will give the next reward r t , the state transition is s t+1 , after each decision step, a four-tuple is generated Store it in the temporary experience library, where r t * Represents the single-step reward return value of the current decision, repeat the steps until the maximum decision step time Max_Step is reached, or the controller model controls the robot to complete the task. If the task is completed within the specified time, the victory global reward r is given win , that is, r t =r t_end * +r win If it is determined that there is a collision with an obstacle or the task is not completed within the specified time, a global reward r will be given for failure lose, that is, r t =r t_end * +r lose , if it is determined to be a tie, no operation is performed, that is, t_end represents the end time of the round. The four-tuple of each decision cycle in the temporary experience library All of them are stored in the total memory bank. If the total memory bank does not overflow after adding all the experience of the current temporary experience bank, it will be stored directly. If it overflows, the old experience in the memory bank will be deleted according to the amount of overflowing experience, and then the temporary experience bank will be stored and the temporary experience bank will be cleared.

[0125] S44. Network sampling training

[0126] When the total memory in step S43 is full, the robot trains the online actor network and online critic network once after each decision. The training process is: extract N experiences from the total memory <s i ,a i ,r i ,s i+1 >, calculate Among them, γ is the discount factor, which represents the ability to foresee future rewards. The loss is calculated using the mean square error (MSE):

[0127]

[0128] Update the node parameters θ of the online critic network through error back propagation and gradient descent Q .

[0129] Update the node parameters of the online Actor network based on policy gradient:

[0130]

[0131] in, represents the policy gradient, Represents gradient calculation.

[0132] In order to make the training process more stable, it is necessary to update the online Actor network at a lower frequency and the online Critic network at a higher frequency. Therefore, it is set to update the Actor network once every m updates of the Critic network. After the node parameters of the online Actor network and the online Critic network are updated, the node parameters of the target Actor network and the target Critic network are soft updated. Each time the node parameters of the online network are weighted with the node parameters of the target network, the soft update is conducive to the stability of the target network, and its update method is θ′ Q ←τθ Q+(1-τ)θ′ Q ,θ π ′←τθ π +(1-τ)θ π ′, τ represents the soft update ratio of the target network.

[0133] If the current training round number episode is less than the maximum training round number Max_Episode, then return to step S42 to obtain new experience data. Otherwise, save the current network model node parameters and end the training.

[0134] As an implementation of the embodiment of the present invention, step 5 includes:

[0135] Use the trained network model to make a single-step decision on the robot's shared control authority, and output the action value a of the online Actor network. t Limit the actual actions performed For repeated single-step decision making.

[0136] The operator will wear an EEG cap and make judgments about the environment by observing the images sent back by the drone. He will then use SSVEP to generate EEG control signals to control the robot to complete the target task. The EEG collector will collect the operator's EEG signals in real time to analyze the operator's degree of distraction.

[0137] During the operator control process, the online Actor network will output action values ​​to adjust the control authority parameters in the shared controller to better complete the control task.

[0138] Embodiment 2:

[0139] The embodiment of the present invention provides a brain-computer sharing control system for an air-ground collaborative unmanned system, comprising:

[0140] A first building module is used to build a mobile robot model;

[0141] A second building module is used to build an MPC shared controller based on the mobile robot model;

[0142] The second building block is used to construct the action space, reward function and neural network model of the DDPG algorithm;

[0143] The training module is used to train the DDPG-based neural network model to obtain the optimized model;

[0144] The control module is used to adjust the control authority parameters of the MPC-based shared controller according to the optimization model to control the robot to complete the specified task.

[0145] As an implementation method of the present invention, the mobile robot model adopts a two-wheel differential mobile robot model, and the robot state ξ=[xy θ] T , state quantity u=[v ω] T , the kinematic model of the mobile robot is expressed as:

[0146]

[0147] Among them, x and y represent the positions of the mobile robot in the inertial coordinate system, and θ is the current heading angle of the mobile robot.

[0148] As an implementation method of the embodiment of the present invention, the optimization function of the MPC shared controller is set to:

[0149] minJ(x(k),U k )

[0150] u min ≤u(k)≤u max ,i=0,...,H C -1

[0151] Δu(k+i)=u(k+i)-u(k+i-1), i=0,...,H C -1

[0152] Δu min ≤Δu(k)≤Δu max ,i=0,...,H C -1

[0153] y min ≤y b (k)≤y max ,i=0,...,H P -1

[0154] Δu(k+i)=0,i=H C ,H C +1,...,H P -1

[0155] The objective function can be set as follows:

[0156]

[0157] In the above cost function, k represents the current moment, U(k) represents the brain-computer interface input at the current moment, x(k) represents the system state at the current moment, and H p and H c represent the prediction time domain and the control time domain respectively, and H C ≤H P, α represents the control authority allocation, and α~clip(0,1). and are the weight matrices for the penalty output error, penalty control action, and penalty control change rate, respectively. r(·) is the controller reference input, u r (·) is the corresponding control reference input.

[0158] As an implementation method of an embodiment of the present invention, the mobile robot is controlled by a controller model to complete the exploration task, wherein the controller model control signal corresponds to the control command generated by the brain-computer interface, and the command X={x su ,x sd ,x tl ,x tr ,x ms ,x stop}, where the commands represent acceleration, deceleration, turn left, turn right, maintain speed, and stop;

[0159] Let I C ∈{1,2,3,4,5,6} represents the intention command decoded from the user's EEG signal, where 1 is acceleration, 2 is deceleration, 3 is left turn, 4 is right turn, 5 is maintain speed, and 6 is stop. The following interface model is designed as follows:

[0160]

[0161] in, Respectively represent the linear velocity and angular velocity commands output by the interface model at time k, represents the user intention output by the SSVEP-based BCI at time k, v max ,v min , and Δv represent the maximum linear velocity, minimum linear velocity, and linear velocity gain, respectively, ω min and ω max represents the limit of angular velocity, and Δω is the angular velocity gain.

[0162] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A brain-computer sharing control method for air-ground collaborative unmanned systems, characterized in that: include: Step S1, constructing a mobile robot model; Step S2: constructing an MPC shared controller based on the mobile robot model; Step S3, constructing the action space, reward function and neural network model of the DDPG algorithm; Step S4: training a DDPG-based neural network model to obtain an optimized model; Step S5: adjusting the control authority parameters of the MPC-based shared controller according to the optimization model to control the robot to complete the specified task; The mobile robot model adopts a two-wheel differential mobile robot model, and the robot state ξ=[xy θ] T , state quantity u=[v ω] T , the kinematic model of the mobile robot is expressed as: Among them, x and y represent the position of the mobile robot in the inertial coordinate system, and θ is the current heading angle of the mobile robot; The optimization function of the MPC shared controller is set as: minJ(x(k),U k ) u min ≤u(k)≤u max ,i=0,...,H C -1 Δu(k+i)=u(k+i)-u(k+i-1),i=0,…,H C -1 Δu min ≤Δu(k)≤Δu max ,i=0,…,H C -1 y min ≤y b (k)≤y max ,i=0,...,H P -1 Δu(k+i)=0,i=H C ,H C +1,...,H P -1 The objective function is set as follows: In the cost function, k represents the current moment, To control the input, is the constraint output, y c (k+i+1) is the control output, U k represents the brain-computer interface input at the current moment, x(k) represents the system state at the current moment, and H p and H c represent the prediction time domain and the control time domain respectively, and H C ≤H P , α represents the control authority allocation, and α~clip(0,1); and are the weight matrices of penalty output error, penalty control action and penalty control change rate, r(·) is the controller reference input, u r (·) is the corresponding control reference input.

2. The brain-computer sharing control method for air-ground collaborative unmanned system according to claim 1 is characterized in that: The controller model is used to control the mobile robot to complete the exploration task, where the controller model control signal corresponds to the control command generated by the brain-computer interface, command X = {x su ,x sd ,x tl ,x tr ,x ms ,x stop }, where the commands represent acceleration, deceleration, turn left, turn right, maintain speed, and stop; Let I C ∈{1,2,3,4,5,6} represents the intention command decoded from the user's EEG signal, where 1 is acceleration, 2 is deceleration, 3 is left turn, 4 is right turn, 5 is maintain speed, and 6 is stop. The following interface model is designed as follows: in, Respectively represent the linear velocity and angular velocity commands output by the interface model at time k, represents the user intention output by the SSVEP-based BCI at time k, v max ,v min , and Δv represent the maximum linear velocity, minimum linear velocity, and linear velocity gain, respectively, ω min and ω max represents the limit of angular velocity, and Δω is the angular velocity gain.

3. A brain-computer sharing control system for air-ground collaborative unmanned systems, characterized in that: include: A first building module is used to build a mobile robot model; A second building module is used to build an MPC shared controller based on the mobile robot model; The second building block is used to construct the action space, reward function and neural network model of the DDPG algorithm; The training module is used to train the DDPG-based neural network model to obtain the optimized model; A control module is used to adjust the control authority parameters of the MPC-based shared controller according to the optimization model to control the robot to complete the specified task; The mobile robot model adopts a two-wheel differential mobile robot model, and the robot state ξ=[xy θ] T , state quantity u=[v ω] T , the kinematic model of the mobile robot is expressed as: Among them, x and y represent the position of the mobile robot in the inertial coordinate system, and θ is the current heading angle of the mobile robot; The optimization function of the MPC shared controller is set as: minJ(x(k),U k ) u min ≤u(k)≤u max ,i=0,...,H C -1 Δu(k+i)=u(k+i)-u(k+i-1),i=0,…,H C -1 Δu min ≤Δu(k)≤Δu max ,i=0,…,H C -1 y min ≤y b (k)≤y max ,i=0,...,H P -1 Δu(k+i)=0,i=H C ,H C +1,...,H P -1 The objective function is set as follows: In the cost function, k represents the current moment, To control the input, is the constraint output, y c (k+i+1) is the control output, U k represents the brain-computer interface input at the current moment, x(k) represents the system state at the current moment, and H p and H c represent the prediction time domain and the control time domain respectively, and H C ≤H P , α represents the control authority allocation, and α~clip(0,1), the objective function and are the weight matrices of penalty output error, penalty control action and penalty control change rate, r(·) is the controller reference input, u r (·) is the corresponding control reference input.

4. The brain-computer sharing control system for air-ground collaborative unmanned systems as claimed in claim 3 is characterized in that: The controller model is used to control the mobile robot to complete the exploration task, where the controller model control signal corresponds to the control command generated by the brain-computer interface, command X = {x su ,x sd ,x tl ,x tr ,x ms ,x stop }, where the commands represent acceleration, deceleration, turn left, turn right, maintain speed, and stop; Let I C ∈{1,2,3,4,5,6} represents the intention command decoded from the user's EEG signal, where 1 is acceleration, 2 is deceleration, 3 is left turn, 4 is right turn, 5 is maintain speed, and 6 is stop. The following interface model is designed as follows: in, Respectively represent the linear velocity and angular velocity commands output by the interface model at time k, represents the user intention output by the SSVEP-based BCI at time k, v max ,v min , and Δv represent the maximum linear velocity, minimum linear velocity, and linear velocity gain, respectively, ω min and ω max represents the limit of angular velocity, and Δω is the angular velocity gain.

Images