Multi-copter trajectory tracking optimization method based on mlp acceleration estimation and reinforcement learning

Abstract

The application discloses a kind of multi-rotor trajectory tracking optimization methods based on MLP acceleration estimation and reinforcement learning, it is related to flight control technical field, method includes: receiving the target position of unmanned aerial vehicle, the position, velocity, attitude, relative position error and current battery voltage value of man-machine are synthesized into observation vector, input reinforcement learning model and infer, output the reference acceleration planned to controller;The desired attitude and desired thrust of unmanned aerial vehicle are obtained according to kinematics formula by controller;The optimization algorithm of control parameter is adjusted in real time by thrust-throttle compensation network model, corresponding throttle value is calculated, and the optimal bottom control instruction is solved.The method of the application takes the target acceleration generated by reinforcement learning as the core control quantity, realizes high-performance flight control through hierarchical control strategy, even in the case of significant battery voltage fluctuation, the system can still maintain superior trajectory tracking performance and control stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle flight control, and particularly relates to a multi-rotor trajectory tracking optimization method based on MLP acceleration estimation and reinforcement learning. BACKGROUND

[0002] In the classic quadrotor control, the proportional-integral-derivative (PID) or linear quadratic regulator (LQR) control is a common method in the field of unmanned aerial vehicle control after linearizing the quadrotor model. However, the linear controller is inherently difficult to cope with complex aerodynamic disturbances and significant nonlinear dynamics in high-speed flight, and the improvement of the traditional controller is difficult to achieve an ideal balance among the requirements of computing resources, real-time response performance and actual control effect. In addition, the use of the complete quadrotor nonlinear model will result in a high computational load of various improved algorithms. For low-power edge computing units, deploying the algorithm not only cannot fully play the performance of the algorithm, but also will affect the operation of other positioning, identification or planning algorithms of the unmanned aerial vehicle.

[0003] The planning method of reinforcement learning performs excellently in various environments. Reinforcement learning directly maps the state to the desired acceleration through an end-to-end neural network, avoiding the complex online optimization calculation in traditional nonlinear model control. In terms of environmental adaptability, reinforcement learning can learn to cope with unmodeled dynamics and disturbances through a large number of interactions with the environment.

[0004] The application of reinforcement learning algorithm in unmanned aerial vehicle control has significantly improved the robustness and computing performance, but its application in high-maneuvering flight scenarios of small multi-rotor unmanned aerial vehicles still faces significant challenges.

[0005] When performing high-speed maneuvering actions, small multi-rotor unmanned aerial vehicles will cause rapid voltage decay of the battery due to large current discharge. Due to the strict size and weight restrictions of the platform, small unmanned aerial vehicles can usually only carry low-capacity battery packs, and will exhibit significant voltage decay after deep discharge. The nonlinear characteristics of this power system will directly lead to the degradation of the power performance of the aircraft. Voltage drop directly affects the peak thrust output of the motor, limits the maneuvering ability of the aircraft, and thus affects the accuracy of attitude control and braking performance. On the other hand, the reinforcement learning network implicitly includes the thrust-throttle mapping relationship at a specific voltage during the training process. In high-maneuvering flight conditions, the battery of the unmanned aerial vehicle will rapidly decay in voltage, resulting in an open-loop error between the target acceleration output of the controller and the actual execution capability: as the voltage continues to drop, the control instructions output by the network gradually mismatch the response capability of the power system. This unmanned aerial vehicle control challenge caused by the electromechanical system coupling effect puts higher requirements on the robustness of the reinforcement learning strategy. SUMMARY

[0006] The problem to be solved by the present application is to provide a multi-rotor trajectory tracking optimization method based on MLP acceleration estimation and reinforcement learning, which can avoid the approximation error caused by traditional model linearization, ensure the calculation efficiency, and effectively overcome the problem of thrust characteristic change caused by battery voltage attenuation.

[0007] The present application adopts the following technical scheme: a multi-rotor trajectory tracking optimization method based on MLP acceleration estimation and reinforcement learning, comprising the following steps:

[0008] Step 1, receiving the target position of the unmanned aerial vehicle, synthesizing the current position, speed, attitude, relative position error and current battery voltage value of the unmanned aerial vehicle into an observation vector, and performing flight control of the unmanned aerial vehicle through a hierarchical control strategy;

[0009] Step 2, in the top layer, constructing a reinforcement learning model and performing model training, inputting the synthesized observation vector into the trained reinforcement learning model, performing reasoning, and outputting the planned reference acceleration;

[0010] Step 3, in the middle layer, receiving the planned reference acceleration through the controller, obtaining the expected attitude and expected thrust of the unmanned aerial vehicle according to the kinematic formula, and performing bottom layer control;

[0011] Step 4, in the bottom layer, based on the thrust-throttle compensation network model, adjusting the attitude and thrust control parameters in real time, and calculating the corresponding throttle value by inputting the expected thrust into the trained thrust-throttle compensation network model, and solving the optimal bottom layer control instruction;

[0012] Step 5, using the optimal bottom layer control instruction to control the aircraft by the flight control, and optimizing the trajectory tracking under the condition of large change of the battery voltage of the unmanned aerial vehicle.

[0013] Preferably, in step 1, the observation vector is represented as: ;

[0014] wherein, is the current center of mass position of the unmanned aerial vehicle; is the current speed of the unmanned aerial vehicle; is a rotation matrix for representing the attitude of the aircraft; is the real-time voltage of the battery; is the relative position error.

[0015] Preferably, in step 2, the reinforcement learning model performs reasoning according to the input observation vector, defines a four-dimensional action vector , processes the action vector , and the first three dimensions are normalized to represent the direction of the reference acceleration , and the fourth dimension is multiplied by the preset maximum acceleration Calculate the output acceleration Output acceleration The acceleration is used as a reference for planning and is transmitted to the controller.

[0016] Preferably, the reinforcement learning model uses a proximal policy optimization algorithm, and a reward function is used during model training. , means as follows:

[0017] ;

[0018] in, Indicates distance reward. Indicates speed penalty, This represents a penalty for smoothness of motion; the greater the abrupt change in the output motion, the greater the penalty. This indicates a time penalty; the longer it takes the machine to reach the goal, the greater the penalty. This represents the acceleration boundary penalty, which is applied when the acceleration given by the reinforcement learning network... The value is close to When the acceleration boundary penalty is introduced, a large penalty is imposed when the thrust calculated by the acceleration given by the reinforcement learning network exceeds the current thrust limit of the machine.

[0019] Preferably, the body acceleration boundary The thrust-throttle compensation network model is derived through inference. This model uses a... MLP network with structure, input throttle value and voltage Output thrust .

[0020] Specifically, during the training of the reinforcement learning model, if the given acceleration direction exceeds the maximum angle... This will affect the distance reward and time reward, causing the overall reward value to decrease. Setting the reward function to automatically abandon this strategy and adaptively adjust the acceleration direction until the reinforcement learning model converges, so that the drone can correctly and quickly reach the designated target point.

[0021] Preferably, in step 3, the desired attitude and desired thrust of the UAV are obtained according to the kinematic formulas, as follows:

[0022] Step 3.1: Calculate the desired thrust :

[0023] .

[0024] Step 3.2: Describe the machine rotation using the zxy sequence. and The body coordinate system represents the x-axis and y-axis unit vectors after considering the z-axis yaw angle rotation:

[0025] ;

[0026] ;

[0027] wherein, is the reference yaw angle, and the superscript represents transposition.

[0028] Step 3.3, in the body coordinate system, , and are mutually orthogonal and are unit vectors, and the expected attitude components , and are calculated and represented as follows:

[0029] .

[0030] Step 3.4, the rotation matrix of the expected attitude is obtained:

[0031] .

[0032] Preferably, in step 4, the expected thrust is transmitted to the thrust-throttle compensation network model, and the corresponding throttle value is calculated by real-time inference of the MLP network:

[0033] ;

[0034] wherein, is the MLP network function, and the current voltage given by the sensor is represented by

[0035] The technical scheme of the present application also provides an electronic device comprising:

[0036] one or more processors;

[0037] a storage device having one or more programs stored thereon;

[0038] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned any one of the MLP-based acceleration estimation and reinforcement learning multi-rotor trajectory tracking optimization methods.

[0039] Compared with the prior art, the above technical scheme has the following technical effects:

[0040] 1. The multi-rotor trajectory tracking optimization method of the present application, the overall process shows the use of reinforcement learning planning and thrust-throttle compensation network, by combining learning algorithm and traditional control, ensures that the unmanned aerial vehicle can still maintain superior trajectory tracking performance and control stability under the condition of battery voltage change, realizes reasonable planning and efficient and safe movement.

[0041] 2. The present application proposes a new four-rotor unmanned aerial vehicle hybrid control architecture, taking the target acceleration generated by reinforcement learning as the core control quantity, realizing high-performance flight control through hierarchical control strategy; at the top layer, the reinforcement learning network directly outputs a three-dimensional target acceleration vector, fully utilizing its optimal decision-making advantage in complex environment, avoiding the stability risk brought by directly outputting the bottom layer control quantity; at the middle layer, the optimal controller designed based on the differential flatness theory converts the target acceleration into optimal control instructions, avoiding the approximation error brought by traditional model linearization, and ensuring the calculation efficiency; at the bottom layer, the lightweight thrust-throttle dynamic compensation network adjusts the control parameters in real time, effectively overcoming the problem of thrust characteristic change caused by battery voltage decay. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is the acceleration flowchart generated by the multi-rotor trajectory tracking optimization method of the present application;

[0043] Figure 2 is the control instruction calculation flowchart of the multi-rotor trajectory tracking optimization method of the present application;

[0044] Figure 3 is the coordinate system definition and propeller number definition diagram of the embodiment of the present application;

[0045] Figure 4 is the thrust-throttle dynamic compensation network structure diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the application will be further described in detail below with reference to the drawings. The described embodiments are only a part of the embodiments involved in the present application. All non-innovative embodiments of other researchers in the field on the basis of the embodiments belong to the protection scope of the present application. At the same time, the step numbers in the embodiments are only set for the convenience of description and explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0047] In one embodiment of the present application, based on pybullet and gazebo simulation platform, using the strange elephant 35 four-rotor unmanned aerial vehicle, taking the target acceleration generated by reinforcement learning as the core control quantity, realizing high-performance flight control through hierarchical control strategy.

[0048] Firstly, the control system receives the target position of the UAV, and inputs the current position, velocity, attitude, relative position error and current battery voltage value of the UAV into the trained reinforcement learning network as an observation vector. The network outputs the planned reference acceleration according to the input value, as shown in the flowchart of Figure 1 .

[0049] Then, the controller of the system receives the reference acceleration and calculates the expected attitude and thrust according to the kinematic formula to perform the bottom-layer control.

[0050] Finally, the system constructs a specially designed lightweight thrust-throttle dynamic compensation network model to adjust the optimization algorithm of the control parameters in real time, and solves the optimal bottom-layer control instruction to ensure the accuracy of the control under the condition of a large change in battery voltage, as shown in the flowchart of Figure 2 .

[0051] The multi-rotor trajectory tracking optimization method provided in this embodiment shows the use of reinforcement learning planning and thrust-throttle compensation network in the overall process. By combining reinforcement learning algorithm and traditional control, the UAV can achieve reasonable planning and efficient and safe movement under the condition of battery voltage change, as follows:

[0052] Special note: In this embodiment, vectors are represented by bold lowercase letters, matrices are represented by bold uppercase letters, and all other symbols represent scalars unless otherwise specified.

[0053] Firstly, the world coordinate system and the body coordinate system are defined as follows:

[0054] World coordinate system is defined using the orthogonal basis , and the body is defined using the orthogonal basis . Both coordinate systems are described in the world coordinate system.

[0055] The body coordinate system and the world coordinate system are defined as shown in Figure 3 , , , respectively represent the unit vectors of the three axes of the world coordinate system, , , respectively represent the unit vectors of the three axes of the body coordinate system in the world coordinate system, Figure 3 , , , , respectively represent the numbers, speeds and rotation directions of the four motors.

[0056] In addition, the symbol represents the position of the center of mass of the UAV, represents the current speed of the UAV, and the attitude of the UAV body is represented by a rotation matrix .

[0057] In this embodiment, the learning strategy of the reinforcement learning model is as follows:

[0058] The observation space vector is represented as: wherein

[0059] ;

[0060] ;

[0061] ;

[0062] wherein, , , respectively represent the three-axis position of the UAV in the world coordinate system, , , respectively represent the three-axis speed of the UAV in the world coordinate system, represents the real-time voltage of the battery.

[0063] The relative position error is obtained by subtracting the target position from the current position :

[0064] ;

[0065] The action vector is a four-dimensional vector:

[0066] ;

[0067] ;

[0068] ;

[0069] wherein, , respectively represent the direction and size of the action instruction, , , respectively represent the three-axis components of the expected acceleration direction, represents the size of the expected acceleration.

[0070] After inference by the reinforcement learning model, the action vector is processed, and the first three dimensions are normalized to represent the direction of the reference acceleration The fourth dimension is the magnitude of the acceleration multiplied by the preset maximum acceleration. Calculate the output acceleration The target value for the plan is passed to the controller:

[0071] ;

[0072] .

[0073] In this embodiment, the reinforcement learning model uses the PPO (Proximity Policy Optimization) algorithm to improve the stability and reliability of training. The reward function used during reinforcement learning model training... , means as follows:

[0074] ;

[0075] Specifically, This represents the distance bonus, calculated from the distance between the current time step's position and the target point.

[0076] This indicates a speed penalty, which is introduced only after the machine has come to a stop. This indicates the speed at the current time step.

[0077] This represents a penalty for smoothness of motion; the greater the abrupt change in the model's output motion, the greater the penalty. This represents the action vector at the current time step. This represents the action vector of the previous time step.

[0078] This represents a time penalty; the longer it takes the machine to reach the target, the greater the penalty. The time taken for the machine to reach the target point in this round is displayed.

[0079] Indicates the acceleration boundary penalty:

[0080] ;

[0081] in, This indicates the maximum acceleration (acceleration boundary) that the machine can currently achieve. This represents the proximity coefficient, which is the acceleration given by the reinforcement learning network. The value is close to Acceleration boundary penalties are only introduced when the acceleration calculated by the reinforcement learning network exceeds the upper limit of the current thrust that the machine can achieve; if the thrust is greater than the upper limit of the current thrust that the machine can achieve, a severe penalty is imposed.

[0082] Acceleration boundary The thrust-throttle compensation network is inferred by a thrust-throttle compensation network.

[0083] In this embodiment, a thrust-throttle compensation network model is constructed, and a MLP (Multi-Layer Perceptron) network of the structure as shown in Figure 4 is used.

[0084] The input throttle value and the voltage , and the output thrust , are as follows:

[0085] .

[0086] Specifically, during the process of keeping the unmanned aerial vehicle hovering, the self-weight (actual thrust) of the aircraft is changed by increasing the load, and the continuously decaying voltage, actual thrust and real-time throttle data are sampled to obtain a sample data set.

[0087] In this embodiment, about 2100 sample data are collected, and the 4s battery capacity ranges from 12.4 to 16.8V, the throttle value ranges from 0.42 to 0.79, and the thrust value ranges from 7.84 to 11.76N. The sample data set is divided into a training set, a validation set and a test set according to the proportions of 70%, 20% and 10%.

[0088] Further, the thrust-throttle compensation network model training process is performed, the optimizer selects Adam, the initial learning rate is set to 0.001, and the loss function uses MSE (Mean Squared Error). After 120 epoch iterations of training, the training and validation loss functions are significantly reduced, the validation set error is lower than the training set, indicating that the thrust-throttle compensation network model of this embodiment can correctly reflect the mapping relationship among the voltage, throttle and thrust in actual flight.

[0089] In order to obtain the maximum thrust that the aircraft can provide under the current state, the current voltage and the maximum throttle value are input into the MLP network, and the formula is as follows:

[0090] ;

[0091] .

[0092] In order to ensure that the quadcopter does not fall during flight, it is necessary to satisfy the condition that the component of the thrust on the z-axis in the world coordinate system is greater than or equal to the gravity of the aircraft:

[0093] ;

[0094] where, is the tilt angle, representing the angle between and

[0095] the maximum tilt angle can be expressed as:

[0096] ;

[0097] The acceleration boundary vector is derived from Newton's second law and is given by:

[0098] .

[0099] A special note here: if the given acceleration direction exceeds the maximum angle during the training process of the reinforcement learning model, it will affect the distance reward and time reward, resulting in a decrease in the overall reward value. The algorithm will automatically discard this strategy and adaptively adjust the direction of acceleration during the model convergence process to achieve the best results, so there is no need to specially limit the direction of acceleration in the reward function.

[0100] In particular, the reinforcement learning model of this embodiment converges significantly after training one million (1e 7 ) rounds, and can correctly and quickly reach the specified target point in simulation.

[0101] The expected acceleration output by the reinforcement learning model already contains the effects of gravity acceleration and battery attenuation and other interference factors during the training process, and can be directly used for attitude calculation or bottom-layer control execution.

[0102] Further, the expected attitude and expected thrust of the UAV are obtained according to the kinematic formula.

[0103] The expected thrust is calculated as follows:

[0104] ;

[0105] The rotation of the body is described in the order of z-x-y, and are defined as follows:

[0106] ;

[0107] ;

[0108] where, is the reference yaw angle.

[0109] In particular, since the body orientation is constant in the algorithm design of the present embodiment, Default is 0.

[0110] In the body coordinate system, , and are mutually orthogonal and are unit vectors, calculate the expected attitude components , and , as follows:

[0111] ;

[0112] The rotation matrix of the expected attitude :

[0113] .

[0114] Further, the controller solves the expected thrust and the expected attitude , and delivers the expected thrust to the thrust-throttle compensation network model to calculate the corresponding throttle value.

[0115] According to the expected thrust and the current voltage given by the sensor , the MLP network infers the throttle value of the control command in real time :

[0116] .

[0117] Further, the throttle value of the control command is used for the px4 flight control, and the flight control combines the expected attitude to directly give the motor speed, which has an actual control effect on the unmanned aerial vehicle.

[0118] In summary, based on the MLP acceleration estimation and the multi-rotor trajectory tracking optimization method of reinforcement learning, a hybrid control architecture of quad-rotor UAV is proposed, taking the target acceleration generated by reinforcement learning as the core control quantity, and realizing high-performance flight control through hierarchical control strategy. In the top layer, the reinforcement learning model directly outputs a three-dimensional target acceleration vector, fully utilizing its optimal decision-making advantage in complex environments, and avoiding the stability risk brought by directly outputting the bottom layer control quantity; in the middle layer, the optimal controller designed based on the differential flatness theory converts the target acceleration into optimal control instructions, avoiding the approximation error brought by traditional model linearization, and ensuring the calculation efficiency; in the bottom layer, the specially designed lightweight thrust-throttle dynamic compensation network adjusts the control parameters in real time, effectively overcoming the problem of thrust characteristic change caused by battery voltage attenuation, and even in the presence of significant battery voltage fluctuations, the optimization method of the application can still maintain superior trajectory tracking performance and control stability.

[0119] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for multi-copter trajectory tracking optimization based on MLP acceleration estimation and reinforcement learning, characterized in that, Comprising the following steps: Step 1, receiving the target position of the unmanned aerial vehicle, synthesizing the current position, speed, attitude, relative position error and current battery voltage value of the unmanned aerial vehicle into an observation vector, and performing flight control of the unmanned aerial vehicle through a hierarchical control strategy; Step 2, in the top layer, constructing a reinforcement learning model and performing model training, inputting the synthesized observation vector into the trained reinforcement learning model, performing reasoning, and outputting a planned reference acceleration; Step 3, in the middle layer, receiving the planned reference acceleration through the controller, obtaining the expected attitude and expected thrust of the unmanned aerial vehicle according to the kinematic formula, and performing bottom layer control; Step 4, in the bottom layer, based on the thrust-throttle compensation network model, real-time adjustment of attitude, thrust control parameters, the expected thrust is passed to the trained thrust-throttle compensation network model, the corresponding throttle value is calculated, and the optimal bottom layer control instruction is solved; Step 5, controlling the body through the flight control according to the optimal bottom layer control instruction, and realizing the optimization of trajectory tracking under the condition of large change of the battery voltage of the unmanned aerial vehicle; The observation vector is represented as: ; wherein, is the current position of the UAV; is the current velocity of the UAV; is a rotation matrix representing the body attitude, , , , represent the unit vectors of the three axes of the body coordinate system in the world coordinate system, respectively; is the real-time voltage of the battery; is the relative position error; The relative position error is calculated as follows: ； wherein, is a target position of the UAV; The reinforcement learning model reasons according to the input observation vector, and the strategy is as follows: Defining a four-dimensional motion vector : ； ； ； wherein, , respectively represent the direction and magnitude of the motion command, , , respectively represent the three-axis components of the desired acceleration direction, represents the magnitude of the desired acceleration; The action vector The first three vectors are normalized to represent the direction of the reference acceleration The fourth vector is the acceleration multiplied by a preset maximum acceleration The output acceleration is obtained : ； ； adding the output acceleration as a reference acceleration for planning, to the controller; The reinforcement learning model uses a proximal policy optimization algorithm, and a reward function is used during model training is represented as follows: ； wherein, represents a distance reward, obtained from the distance between the current time step robot position and the goal point position; represents the speed penalty introduced by the body after hovering, represents the speed at the current time step; denotes an action smoothness penalty, the larger the output action jump, the larger the penalty; wherein, denotes an action vector of the current time step, denotes an action vector of the previous time step; denotes a time penalty, the more time the body takes to reach the goal, the greater the penalty, where, denotes the time taken by the body to reach the goal point in the current turn. represents an acceleration boundary penalty: ； wherein, represents the vehicle acceleration boundary, represents the proximity coefficient, when the acceleration given by the reinforcement learning network is numerically close to the vehicle acceleration boundary, a penalty is introduced; when the acceleration given by the reinforcement learning network leads to a calculated thrust that exceeds the current upper limit of the vehicle thrust, a maximum penalty is given. The thrust-throttle compensation network model uses a MLP network, input throttle value and voltage , output thrust , formula as follows: ； wherein, is an MLP network function; The MLP network uses Structure, each of the internal nodes with weights, in the process of hovering by increasing the load changes the weight of the UAV, indicating the actual thrust. 2.The MLP acceleration estimation and reinforcement learning based multi-copter trajectory tracking optimization method of claim 1, wherein, The body acceleration boundary The method is as follows: Step 2.

1. Input the current voltage into the MLP network and the maximum throttle value to find the maximum thrust that the engine can provide in its current state , which is given by the formula: ； ； wherein, , , represents the maximum thrust components of the x, y, z axes in the world coordinate system, greater than or equal to the body gravity: ； wherein is the tilt angle, is the body mass; maximum tilt angle is represented as follows: ； Step 2.

2. Acceleration boundary vector from Newton's second law The formula is as follows: ； wherein denotes the unit vector of the z-axis in the world coordinate system, the superscript denotes the transpose, the tilt angle is the angle between and . 3.The MLP acceleration estimation and reinforcement learning based multi-copter trajectory tracking optimization method of claim 2, wherein, If the given acceleration direction exceeds the maximum angle during the training of the reinforcement learning model , the distance reward and the time reward will be affected, the overall reward value will decrease, the corresponding strategy will be automatically abandoned by setting the reward function, the acceleration direction will be adaptively adjusted, and the unmanned aerial vehicle will correctly and quickly reach the specified target point when the reinforcement learning model converges. 4.The MLP acceleration estimation and reinforcement learning based multi-copter trajectory tracking optimization method of claim 2, wherein, In step 3, the expected attitude and expected thrust of the unmanned aerial vehicle are obtained according to the kinematic formula, and the method is as follows: Step 3.1, Calculate expected thrust : ； Step 3.

2. Body rotation is described in z-x-y order, with and denote the x and y unit vectors of the body frame after accounting for the z-axis yaw rotation. ； ； wherein is the reference yaw angle; Step 3.3, in the body coordinate system, , and are mutually orthogonal and are unit vectors, the expected attitude components , and are calculated as follows: ； Step 3.4, deriving the rotation matrix for the desired pose : ； The rotation matrix as the desired pose output in the control instructions. 5.The MLP acceleration estimation and reinforcement learning based multi-copter trajectory tracking optimization method of claim 4, wherein, In step 4, the desired thrust is passed to the thrust-throttle compensation network model, which computes the corresponding throttle value in real-time through the MLP network inference : ； wherein represents the current voltage given by the sensor.

6. An electronic device, comprising: Comprise: One or more processors; A storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-rotor trajectory tracking optimization method of any one of claims 1 to 5.

Images