A robot trajectory tracking control method based on deep learning
By combining the control strategies of LSTM and SMC, the LSTM network is used to compensate for the motion state error, which solves the jitter problem in robot trajectory tracking, improves control accuracy and stability, and achieves faster trajectory tracking.
Patent Information
- Application Number
- CN202210238283.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-10
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-03-10
AI Technical Summary
The existing robot trajectory tracking control methods are prone to vibration when facing complex environments and external interference, which affects control accuracy and stability.
Combining long and short-term memory neural network (LSTM) and sliding mode control (SMC), a new control strategy is established to compensate for motion state errors through the LSTM network, and a compensation control amount is generated to combat external interference and improve trajectory tracking. Accuracy and stability.
It effectively suppresses jitter in sliding mode control, improves the convergence speed and control accuracy of trajectory tracking, while maintaining robustness.
Smart Images

Figure CN114625136B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of robot control, and in particular relates to a robot trajectory tracking control method based on deep learning. Background Art
[0002] A mobile robot is a comprehensive robot system that can achieve autonomous movement towards a target and complete corresponding control tasks. This system integrates professional technologies such as artificial intelligence, information processing, intelligent control, and image processing, spanning multiple disciplines such as mechanical engineering, computer engineering, and electronic engineering. It is one of the most active areas of scientific and technological development. In the past few years, the interest in tracking mobile robots has grown significantly due to the wide application of mobile robots in unstructured and complex environments, such as forestry, mining, agriculture, military applications, search and rescue, hospital missions, and space exploration. However, all these applications require that mobile robots can achieve navigation tasks efficiently and accurately. Trajectory tracking control is one of the most explored navigation problems in robot motion control. The purpose of trajectory tracking is to enable the controlled system to actively move to the target path and stabilize according to the output of the controller.
[0003] Tracked mobile robots have the advantages of large traction, low slippage, and good off-road performance. They are one of the key research objects of mobile robots. As a MIMO system, domestic and foreign scholars have proposed many advanced control methods in the field of tracked mobile robot trajectory tracking system, such as model predictive control, neural network control, adaptive control, PID control, sliding mode control, etc. Amidi et al. proposed three relatively practical tracking control methods, including a pure tracking trajectory tracking control method, but the steady-state error is large. Jia Heming et al. designed a backstepping controller based on feedback gain, used neural networks to compensate for model uncertainties, and used adaptive robust controllers to compensate for the estimation errors of neural networks online, so as to achieve accurate tracking of arbitrary curved paths by mobile robots. Kuang Wenlong et al. designed a trajectory tracking control law that combines the Lyapunov method and the inverse sliding mode technology, which improves the convergence speed, but the control law design is too complicated. Based on the fuzzy sliding mode control theory, Jiao Jun et al. established a cascade system consisting of motor drive equations and motion equations to effectively achieve tracking control. Deep learning is a machine learning method based on learning data representation. DL architectures such as deep neural networks (DNNs) and recurrent neural networks (RNNs) have shown excellent performance in discovering complex structures in high-dimensional data. With the development and application of deep learning algorithms, the combination of traditional control methods and deep learning algorithms can achieve efficient and stable control goals, which has great potential.
[0004] Kunming Zheng et al. designed a dual fuzzy sliding mode controller and developed a deep learning algorithm to get as close to the dynamic model as possible. Xuan Yao et al. proposed a deep convolutional neural network-sliding mode control algorithm based on deep learning methods, designed a two-layer convolutional neural network, and achieved high-precision trajectory tracking. Minghao Zhou et al. combined the advantages of DCNN and the new FOTSM to reduce the chatter phenomenon and improve the control performance of the rigid robot manipulator.
[0005] SMC (sliding mode control, SMC) as an improved nonlinear control method, can design switching control laws for high-speed switching so that the trajectory of the system always remains on a predetermined sliding surface. The sliding mode method is widely used because of its insensitivity to parameter changes, high robustness to external interference, simple physical implementation and fast dynamic response. Its defect is that it is prone to chattering, which makes the actual control effect not very ideal; LSTM (Long Short Term-Memory, LSTM) has memory ability and can control the influence of past information on current information. Using this feature, the trajectory tracking error of the past moment is input to predict the control output of the current moment and compensate the control system.
[0006] To this end, the present invention combines the sliding mode control method with the deep learning algorithm and proposes a novel LSTM-SMC control strategy to track and control the robot trajectory. Summary of the invention
[0007] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a robot trajectory tracking control method based on deep learning.
[0008] In order to achieve the above object, the present invention provides the following technical solutions:
[0009] A robot trajectory tracking control method based on deep learning includes the following steps:
[0010] Establish the kinematic model, tracking error model, and dynamic model of the tracked robot;
[0011] Establishing a controller based on a long short-term memory neural network (LSTM) and a sliding mode control (SMC), wherein the controller includes a kinematic controller, a dynamic controller based on SMC, and an LSTM network;
[0012] Based on the kinematic model, tracking error model and dynamic model of the crawler robot, the tracking trajectory of the crawler robot is controlled by the controller based on the long short-term memory neural network LSTM and sliding mode control SMC;
[0013] The control steps of the controller are:
[0014] The input of the kinematic controller is the expected trajectory of the tracked robot and the expected motion velocity v b , angular velocity ω b and the error e between the actual motion state and the output q of the kinematic controller p , the LSTM network output u and the control output q generated by the first integral operation are used as the input of the dynamic controller;
[0015] The output control quantity τ of the SMC-based dynamic controller is calculated by the dynamic model. After integration, we get the actual running speed v and angular velocity ω of the tracked robot. The real motion state p of the tracked robot is generated by calculating the kinematic model. p ;
[0016] Real motion state p p and q generated by the kinematic controller p As the input of the error calculation module, the motion state error e is output;
[0017] The motion state error e is input into the LSTM network, and a compensation control bias u is output. u is used to control and compensate for the system with external interference.
[0018] Preferably, the kinematic model is:
[0019]
[0020] In the formula, v p is the speed of the tracked robot at its center of mass C; θ p is the angle between the tracked robot's driving direction and the positive direction of the x-axis; ω p is the steering angular velocity of the tracked robot.
[0021] Preferably, trajectory tracking means that the actual crawler robot starts from a given initial position in the reference coordinate system, reaches and tracks a given desired trajectory, and calculates the trajectory error (e x ,e y ,e θ ) T and control output (v,ω) T , so that e x →0,e y →0,e θ →0;
[0022] The trajectory tracking error formula is as follows:
[0023]
[0024] In the formula, e x and e y They are the x-axis and y-axis errors based on the actual position of the crawler robot and the reference position, namely the longitudinal displacement error and the lateral displacement error; θ is the heading angle error; θ b is the reference heading angle; v b and ω b are the reference driving speed and steering angular velocity of the crawler robot respectively. The error dynamic model is obtained by differentiating the tracking error formula. The error dynamic model, i.e. the tracking error model, is:
[0025]
[0026] In the formula, v p is the speed of the tracked robot at the center of mass C, ω p is the steering angular velocity of the tracked robot, ω b is the reference steering angular velocity of the tracked robot.
[0027] Preferably, the kinetic model is:
[0028]
[0029] In the formula, τ 1 and τ 2 They are the driving force and torque of the crawler robot respectively; m is the mass; I is the moment of inertia; η=(η 1 ,η 2 ) T is external interference, η 1 , η 2 All of them are unknown external interference.
[0030] Preferably, the control steps of the LSTM network are:
[0031] The parameters of the LSTM network are determined by the input of the SMC controller (e x ,e y ,e θ ) and the switching control output τ sw The state error is obtained through data training. After being processed by the trained LSTM network, a compensation control bias u is output. Its nonlinear function relationship is u=η(e). Therefore, the control law after compensation is as follows:
[0032]
[0033] Where k and δ are gains, k>0, δ>|ε|; S is the trajectory on the sliding surface, and u is the output of the LSTM network, which is used to control and compensate for the system with external interference;
[0034]
[0035] Where a 1 、a 2 is a coefficient greater than 0, ω p is the steering angular velocity of the tracked robot, e v is the running speed state error, e w is the angular velocity state error, m is the mass, and I is the moment of inertia.
[0036] Preferably, the control law of the kinematic controller is:
[0037]
[0038] In the formula, v p is the speed of the tracked robot at the center of mass C, ω p is the steering angular velocity of the tracked robot, v b and ω b are the reference driving speed and steering angular velocity of the tracked robot respectively; K 1 , K 2 , K 3 are all coefficients greater than 0 and are bounded; e θ is the heading angle error, e x and e y They are the x-axis and y-axis errors based on the actual position and reference position of the tracked robot, respectively.
[0039] The robot trajectory tracking control method based on deep learning provided by the present invention has the following beneficial effects:
[0040] (1) The control method combining LSTM and SMC proposed in the present invention realizes the compensation of LSTM network for control, improves the convergence speed and the accuracy and stability of control;
[0041] (2) Without reducing the robustness, the combination of LSTM and SMC can effectively suppress the jitter phenomenon of sliding mode control. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0043] Figure 1 This is a schematic diagram of the kinematic model of the tracked robot;
[0044] Figure 2 It is a schematic diagram of trajectory tracking error model;
[0045] Figure 3 This is the LSTM network structure diagram;
[0046] Figure 4 This is the internal structure diagram of a single layer of LSTM;
[0047] Figure 5 It is the schematic diagram of sliding mode;
[0048] Figure 6 This is the controller structure diagram combining LSTM and SMC;
[0049] Figure 7 It is the control structure diagram based on the kinematic model;
[0050] Figure 8 It is the control structure diagram of the SMC-based dynamic controller;
[0051] Fig. 9 The linear trajectory simulation results of the controller based on LSTM and SMC; (a) is the trajectory tracking result; (b) is the tracking error result; (c) is the control output result;
[0052] Fig.10 The figure is the linear trajectory simulation result of the SMC controller; (a) is the trajectory tracking result; (b) is the tracking error result; (c) is the control output result;
[0053] Fig.11 The circular trajectory simulation results of the controller based on LSTM and SMC; (a) is the trajectory tracking result; (b) is the tracking error result; (c) is the control output result;
[0054] Fig.12 The circular trajectory simulation results of the SMC controller; (a) is the trajectory tracking result; (b) is the tracking error result; (c) is the control output result. DETAILED DESCRIPTION
[0055] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.
[0056] Example 1
[0057] The present invention provides a robot trajectory tracking control method based on deep learning, specifically taking a tracked mobile robot (referred to as a tracked robot) as an example for control. Figures 6 to 8 As shown, the following steps are included:
[0058] Step 1: Establish the model of tracked mobile robot.
[0059] Step 1.1: Establish kinematic model
[0060] In order to facilitate the establishment of the kinematic model, reasonable assumptions are made about the tracked mobile robot and its movement. The mobile robot can be represented by a rigid body with two independent tracks. It only moves in the plane, the tracks on both sides are exactly the same, the center of mass of the whole machine coincides with the geometric center at point M, and the tracks on both sides do not slide sideways.
[0061] Figure 1 The schematic diagram of the kinematic model of the tracked mobile robot is shown. Oxy is the inertial coordinate system, and the robot's motion state is determined by its position in the inertial coordinate system F 1 The position (x) in (x,y) p ,y p ) and the direction of motion θ p The robot position is given by the coordinates of its geometric center C, which is also the local coordinate system F 2 (x m ,y m ). The distance between the two tracks is L, and the width of the tracks is d.
[0062] x p ,y p are the x-axis and y-axis coordinates of the tracked robot in the reference coordinate system; v p is the speed of the tracked robot at its center of mass C; θ p is the angle between the tracked robot's driving direction and the positive direction of the x-axis; ω p is the steering angular velocity of the tracked robot.
[0063] Depend on Figure 1 The geometric relationship in the reference coordinate system can be obtained, and the kinematic model of the tracked mobile robot is shown in formula (1);
[0064]
[0065] “.” means the derivative with respect to time. For example, if V is the velocity, a “.” on V means the derivative of the velocity with respect to time, which is the acceleration.
[0066] Step 1.2: Establish tracking error model
[0067] Trajectory tracking means that the actual robot starts from a given initial position in the reference coordinate system, reaches and tracks a given desired trajectory, and uses the trajectory error (e x ,e y ,e θ ) T and control output (v,ω) T , so that e x →0,ey →0,e θ →0.
[0068] Figure 2 A schematic diagram of the error tracking model for a tracked mobile robot is shown.
[0069] The trajectory tracking error formula is shown in formula (2):
[0070]
[0071] In the formula, e x and e y They are the x-axis and y-axis errors based on the actual position of the crawler robot and the reference position, namely the longitudinal displacement error and the lateral displacement error; θ is the heading angle error; x b and b are the coordinates of the tracked robot’s reference position on the x-axis and y-axis respectively; θ b is the reference heading angle. Differentiating the tracking error formula, we get the error dynamic model, as shown in formula (3):
[0072]
[0073] In the formula, v p is the speed of the tracked robot at the center of mass C, ω p is the steering angular velocity of the tracked robot, ω b is the reference steering angular velocity of the tracked robot.
[0074] Step 1.3: Establish a dynamic model
[0075] The simplified dynamic model of a tracked mobile robot can generally be expressed as:
[0076]
[0077] τ 1 and τ 2 are the driving force and torque of the robot respectively; m is the mass; I is the moment of inertia; η=(η 1 ,η 2 ) T is external interference, η 1 , η 2 All of them are unknown external interference.
[0078] Step 2: Build a controller based on long short-term memory neural network LSTM and sliding mode control SMC.
[0079] The controller combining LSTM and SMC designed in the present invention mainly consists of three parts: a kinematic controller, a dynamic controller based on SMC and an LSTM network. The dynamic controller outputs a control quantity τ, and then obtains the robot motion speed v and angular velocity ω through subsequent calculations.
[0080] The following is a brief introduction to the long short-term memory neural network LSTM and sliding mode control SMC.
[0081] (1) Introduction to LSTM network:
[0082] LSTM is an improvement on the RNN algorithm, which controls the influence of previous information on current information through three different gates, namely:
[0083] (1) Which cell states should be forgotten.
[0084] (2) Which new states should be added.
[0085] (3) Based on the current state and current input, what should the output be?
[0086] LSTM network such as Figure 3 As shown in the figure, t represents the time; X is the network input at the current time; C is the memory cell unit; h is the temporary memory unit. The specific internal structure of a single LSTM layer is as follows Figure 4 shown.
[0087] LSTM has the ability to remove or add information to the cell state through carefully designed structures called "gates". Gates are a way to selectively let information pass through, and they consist of a sigmoid neural network layer and a bitwise multiplication operation. The sigmoid layer outputs a value between 0 and 1, describing how much of each part can pass through. 0 means "no amount is allowed to pass through", and 1 means "allow any amount to pass through".
[0088] LSTM has three gates to protect and control the cell state.
[0089] (1) The first step in an LSTM is to decide what information to discard from the cell state. This decision is made by a layer called a forget gate. This gate reads h t-1 and X t , output a value between 0 and 1 for each cell in state C t-1 A number in . 1 means "totally keep" and 0 means "totally discard".
[0090] Forget Gate:
[0091] f t =σ(w fh h t-1 +w fxx t +b f ) (5)
[0092] (2) The second step is to determine what new information is stored in the cell state. This involves two parts. First, a sigmoid function called the "input gate" determines what value will be updated. Then, a tanh function creates a new temporary memory cell. Second, use the memory cell C from the previous moment. t-1 and temporary memory cells of the present moment These two pieces of information are used to update memory cells.
[0093] Input Gate:
[0094] i t =σ(w ih h t-1 +w ix x t +b i ) (6)
[0095]
[0096] Update memory cells:
[0097]
[0098] (3) The third step is to determine what value to output. This output will be based on the memory cell C at the current moment. t First, a sigmoid layer is used to determine which part of the cell state will be output. Next, the cell state is processed through a tanh function and the result is multiplied by the output of the sigmoid layer, and only the part that determines the output is output.
[0099] Output Gate:
[0100] o t =σ(w oh h t-1 +w ox x t +b o ) (9)
[0101]
[0102] (2) Sliding mode control
[0103] Sliding mode variable structure control is a control strategy for variable structure control systems. This control characteristic can force the system to make small-amplitude, high-frequency up and down movements along a specified state trajectory under certain characteristics, which is the so-called sliding mode or "sliding mode" movement. Variable structure control has the advantages of fast response, insensitivity to parameter changes and disturbances, and no need for online system identification. The disadvantage of this method is that when the state trajectory reaches the sliding surface, it crosses back and forth on both sides of the sliding surface, resulting in vibration. Figure 5 Schematic diagram of sliding mode.
[0104] Step 3: Based on the robot's kinematic model, tracking error model, and dynamic model, the robot's tracking trajectory is controlled by the controller based on the long short-term memory neural network LSTM and sliding mode control SMC.
[0105] Sliding mode control is used as the main control, and the LSTM network is used to jointly perform online identification and compensation control. The universal approximation characteristics of the neural network are used to approximate external interference and uncertainty, which can effectively reduce the jitter phenomenon of the sliding mode control. First, the control law based on the dynamic equation is obtained, and then the sliding mode controller is designed. The PI type sliding surface is selected, and the quasi-sliding mode method is used to reduce the chatter. From this, the control law equation of the SMC can be obtained, and the interference term in the equation is replaced with the output of the LSTM network to achieve its control compensation function.
[0106] For robots with simple mechanical structures, it is easy to obtain analytical solutions to the dynamics model and achieve precise and stable control. However, in the presence of uncertain disturbances, it is difficult or even impossible to obtain an accurate dynamics model for robots with complex mechanical structures in practical applications. Although scholars have studied some numerical methods, this problem has not been perfectly solved. Past control methods have many disadvantages, such as complex control laws and high requirements for model accuracy. The present invention proposes an algorithm that combines LSTM and SMC to compensate for system control as accurately as possible and reduce the jitter phenomenon of SMC.
[0107] The LSTM network performs well when processing time series data, and the control quantity is time series data. It is theoretically feasible to predict the control at the current moment by the control at the past moment.
[0108] Due to the existence of external interference, the system is difficult to accurately model. In order to achieve more accurate and efficient control, the LSTM network is used to compensate the system. Figure 3 and Figure 4 As shown in Figure 2, the network input is a 1×3 matrix, which consists of three error quantities: longitudinal error, lateral error, and driving angle error. The output is a 1×2 matrix, which is the compensation control quantity.
[0109] The tracking control problem can be solved by keeping the trajectory s(t) = 0 on the sliding surface. Define the error e v =v p -v,e ω =ω p -ω, and select the PI type synovial surface as shown in formula (12):
[0110]
[0111] Among them: a 1 , a 2 >0.
[0112] Differentiating formula (12) yields:
[0113]
[0114] make but for:
[0115]
[0116] In the formula, a 1 、a 2 is a coefficient greater than 0, ω p is the steering angular velocity of the tracked robot, e v is the running speed state error, e w is the angular velocity state error, m is the mass, and I is the moment of inertia.
[0117] The ideal sliding mode does not exist. In reality, sliding mode control is accompanied by chattering. Therefore, the quasi-sliding mode method is used to reduce chattering.
[0118]
[0119] Where k, δ are gains, k>0, δ>|ε|.
[0120] The control law obtained from equations (14) and (15) is:
[0121]
[0122] Where η is an unknown external disturbance. The LSTM network has good performance in processing data with time series. Therefore, the state error of the past moment is input to the LSTM network to obtain the control amount of the next moment, and the system is compensated by the control amount. The parameters of the LSTM network are determined by the input of the SMC controller (e x ,e y ,e θ ) and switching control output (τ sw ) data training, set the initial learning rate γ0 =0.005, epochs=200, multiply by 0.2 after 100 rounds of training to reduce the learning rate. Due to the existence of external interference, there will be a deviation between the actual control amount and the expected control amount, and the state error (e x ,e y ,e θ ) After the training LSTM network is processed, a compensation control bias u is output, and its nonlinear function relationship is u = η(e). Therefore, the control law after compensation is as shown in formula (17):
[0123]
[0124] Among them, u is the output of the LSTM network, which is used to control and compensate the system with external interference.
[0125] Specifically, Figure 7 The control structure diagram based on the kinematic model is shown below. The working process of the controller is as follows Figure 6 The input of the kinematic controller is the expected trajectory of the tracked robot and the expected motion speed (v b ), angular velocity (ω b ) and the error between the actual motion state (e), the output of the kinematic controller (q p ), LSTM network output (u) and the control output q generated by the first integral operation are used as the input of the dynamic controller. The output control quantity (τ) of the dynamic controller based on SMC is calculated by the dynamic model. After integration, the actual running speed (v) and angular velocity (ω) of the tracked robot are obtained, and the real motion state (P p ). Real motion state (P p ) and the kinematic controller generated (q p ) is used as the input of the error calculation module and outputs the motion state error (e). The LSTM network plays a role in control compensation. The motion state error e is input into the LSTM network and a compensation control bias u is output. u is used to control compensation for systems with external interference.
[0126] Juliano G. Iossaqui et al. derived the control input based on the kinematic equations of the mobile tracked robot, ignoring sliding, and gave a proof of stability. The control law is shown in formula (11);
[0127]
[0128] In the formula, v p is the speed of the tracked robot at the center of mass C, ω p is the steering angular velocity of the tracked robot, v band ω b are the reference driving speed and steering angular velocity of the tracked robot respectively; K 1 , K 2 , K 3 are all coefficients greater than 0 and are bounded; e θ is the heading angle error, e x and e y They are the x-axis and y-axis errors based on the actual position and reference position of the tracked robot, respectively.
[0129] The stability of the controller structure based on LSTM and SMC is proved by Lyapunov stability theory. The unknown disturbance is approximated by the neural network:
[0130]
[0131] is the ideal weight, is the ideal deviation, and ε is the ideal approximation error.
[0132] The network output is:
[0133] u=W out T O t +b out T (19)
[0134] W out is the output layer weight, b out is the output layer bias.
[0135] Therefore, the network approximation error can be expressed as:
[0136]
[0137] in,
[0138] Substituting equation (17) into equation (14), we can obtain As shown below:
[0139]
[0140] Define the Lyapunov function as:
[0141]
[0142] where γ 0 is the learning rate, and differentiating V yields:
[0143]
[0144] Pick
[0145]
[0146] but
[0147]
[0148] Assume δ>|ε|, then when S>Δ, we can get:
[0149] -S T (ε+δsat(S))<0 (26)
[0150] Right now Equation (26) shows that the system will converge to a small neighborhood of the sliding surface. Therefore, the system is Lyapunov stable.
[0151] In order to evaluate the performance of the control law proposed in the present invention, linear trajectory tracking and circular trajectory tracking simulation experiments were carried out on the tracked mobile robot in the Matlab / Simulink environment.
[0152] A dynamic controller based on SMC is set as a comparative experiment. The dynamic controller converts the output q of the kinematic controller into p =(v p ,ω p ) T As the reference input, the output is τ=(τ 1 ,τ 2 ) T , and then the control output q=(v,ω) is obtained through the dynamic model and integral operation T .
[0153] like Figure 8 Shown is the control structure diagram of the SMC-based dynamic model.
[0154] The control law of the SMC-based dynamic controller is:
[0155]
[0156] In order to more intuitively demonstrate the superiority of the controller combining SMC and LSTM, simulation tests were carried out on the two controllers to obtain trajectory tracking results, error results and output results, and comparative analysis was performed from various aspects.
[0157] (1) Straight line trajectory simulation
[0158] The initial position and attitude angle of the reference robot are set to (3, -1, 0.25π), the driving speed is 0.5m / s, and the steering angular velocity is 0rad / s; the initial position and attitude angle of the actual robot are set to (1, 1, 0.5π), the driving speed is 0m / s, and the steering angular velocity is 0rad / s; the parameters of the trajectory tracking controller are k1=0.93, k2=0.95, k3=1, a1=1.8, a2=1.9, the mass is m=10kg, and the moment of inertia is I=2.5kg·m 2 ,
[0159] Tracked by trajectory Fig. 9 (a) It can be seen that the expected trajectory of the reference robot is a straight line with an angle of 45 degrees to the positive direction of the X-axis. At the beginning of the simulation, the actual robot has a certain error between the actual trajectory and the expected trajectory, but after a period of tracking control, the actual trajectory coincides with the expected trajectory. Comparison of error results Fig. 9 (b) and Fig.10 (b) It can be seen that the control method combining LSTM and SMC takes about 12 seconds for the error to converge to 0, while the SMC control method completes the convergence of the X-axis coordinate error and the angular error in about 15 seconds, and the Y-axis coordinate error in about 20 seconds, and the convergence process has large fluctuations. From the control output diagram, it can be seen that the control method combining LSTM and SMC completes the convergence of the speed and angular velocity output between 8 and 10 seconds, while the SMC control method completes the convergence in about 15 seconds. The actual robot gradually accelerates at the beginning and uses a large steering angular velocity to quickly adjust its position. After a certain period of time, the speed and angular velocity reach the reference value and remain consistent for a period of time thereafter.
[0160] (2) Circular trajectory simulation
[0161] The initial position and attitude angle of the reference robot are set to (4, 0, 0.5π), the driving speed is 1m / s, and the steering angular velocity is 0.25rad / s; the initial position and attitude angle of the actual robot are set to (1, 1, 0.5π), the driving speed is 0m / s, and the steering angular velocity is 0rad / s; the parameters of the trajectory tracking controller are k1=0.825, k2=0.855, k3=0.885, a1=1.7, a2=1.9, m=10kg, I=2.5kg / m 2 .
[0162] From the trajectory tracking results Fig.11 (a) It can be seen that the expected trajectory of the reference robot is a circular curve. The actual robot has a certain trajectory tracking error at the beginning of the simulation, but after trajectory tracking control, the trajectory of the actual robot coincides with that of the reference robot. Fig.11(b) It can be seen that the trajectory tracking error of the actual robot changes dramatically at the beginning of the simulation, and the longitudinal displacement error e x It reaches a maximum value of about 0.4m at 5 seconds, then converges and gradually becomes stable. The horizontal displacement error e y The size gradually decreases steadily over time and tends to 0. The attitude angle error is the largest at around 2 seconds, and then gradually decreases and tends to be stable over time. Fig.11 (c) It can be seen that the actual robot changes its speed more dramatically at the beginning, and uses a larger steering angle velocity to quickly adjust its position. It reaches the driving speed of the reference robot after about 6 to 7 seconds, reaches the expected steering angle velocity after 10 seconds, and remains consistent in the following time. Fig.11 (b) and Fig.12 (b) It can be seen that the control method combining LSTM and SMC converges to 0 in 9 seconds, and the SMC control method converges to 0 in 15 seconds, and the convergence speed is greatly improved. In addition, the maximum heading angle error and the maximum longitudinal displacement error of the former are smaller than those of the latter, and the horizontal displacement error converges more smoothly and fluctuates less.
[0163] The present invention first constructs the kinematic model, trajectory tracking dynamic error model and dynamic model of the crawler mobile robot. Then, based on the kinematic model and the dynamic model, the kinematic control law and the SMC control law are proposed as reference control laws, and a comparative test is carried out on them; the subsequent proposed method of combining LSTM and SMC utilizes the advantages of sliding mode control and LSTM algorithm to construct a model and successfully conduct Matlab / Simulink simulation tests. The simulation experimental results show that under the conditions of given initial state and initial error, the designed control law can quickly realize linear trajectory tracking and circular trajectory tracking, proving the effectiveness and feasibility of the proposed method.
[0164] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.
Claims
1. A robot trajectory tracking control method based on deep learning, characterized in that: The following steps are involved: Establish the kinematic model, tracking error model, and dynamic model of the tracked robot; Establishing a controller based on a long short-term memory neural network (LSTM) and a sliding mode control (SMC), wherein the controller includes a kinematic controller, a dynamic controller based on SMC, and an LSTM network; Based on the kinematic model, tracking error model and dynamic model of the crawler robot, the tracking trajectory of the crawler robot is controlled by the controller based on the long short-term memory neural network LSTM and sliding mode control SMC; The control steps of the controller are: The input of the kinematic controller is the expected trajectory of the tracked robot and the expected motion velocity v b , angular velocity ω b and the error e between the actual motion state and the output q of the kinematic controller p , the LSTM network output u and the control output q generated by the first integral operation are used as the input of the dynamic controller; The output control quantity τ of the SMC-based dynamic controller is calculated by the dynamic model. After integration, we get the actual running speed v and angular velocity ω of the tracked robot. The real motion state p of the tracked robot is generated by calculating the kinematic model. p ; Real motion state p p and q generated by the kinematic controller p As the input of the error calculation module, the motion state error e is output; The motion state error e is input into the LSTM network, and a compensation control bias u is output. u is used to control and compensate for the system with external interference.
2. The robot trajectory tracking control method based on deep learning according to claim 1 is characterized in that: The kinematic model is: In the formula, v p is the speed of the tracked robot at its center of mass C; θ p is the angle between the tracked robot's driving direction and the positive direction of the x-axis; ω p is the steering angular velocity of the tracked robot.
3. The robot trajectory tracking control method based on deep learning according to claim 2 is characterized in that: Trajectory tracking means that the actual tracked robot starts from a given initial position in the reference coordinate system, reaches and tracks a given desired trajectory, and uses the trajectory error (e x ,e y ,e θ ) T and control output (v,ω) T , so that e x →0,e y →0,e θ →0; The trajectory tracking error formula is: In the formula, e x and e y They are the x-axis and y-axis errors based on the actual position of the crawler robot and the reference position, namely the longitudinal displacement error and the lateral displacement error; θ is the heading angle error; θ b is the reference heading angle; v b and ω b are the reference driving speed and steering angular velocity of the crawler robot respectively. The error dynamic model is obtained by differentiating the tracking error formula. The error dynamic model, i.e. the tracking error model, is: In the formula, v p is the speed of the tracked robot at the center of mass C, ω p is the steering angular velocity of the tracked robot, ω b is the reference steering angular velocity of the tracked robot.
4. The robot trajectory tracking control method based on deep learning according to claim 3 is characterized in that: The kinetic model is: Where τ1 and τ2 are the driving force and torque of the crawler robot respectively; m is the mass; I is the moment of inertia; η=(η1,η2) T is external interference, η1 and η2 are both unknown external interference.
5. The robot trajectory tracking control method based on deep learning according to claim 4 is characterized in that: The control steps of the LSTM network are: The parameters of the LSTM network are determined by the input of the SMC controller (e x ,e y ,e θ ) and the switching control output τ sw The state error is processed by the trained LSTM network after data training, and a compensation control bias u is output, whose nonlinear function relationship is u = η (e). Therefore, the control law after compensation is: Where k and δ are gains, k>0, δ>|ε|; S is the trajectory on the sliding surface, and u is the output of the LSTM network, which is used to control and compensate for the system with external interference; In the formula, a1 and a2 are coefficients greater than 0, ω p is the steering angular velocity of the tracked robot, e v is the running speed state error, e w is the angular velocity state error, m is the mass, and I is the moment of inertia.
6. The robot trajectory tracking control method based on deep learning according to claim 5 is characterized in that: The control law of the kinematic controller is: In the formula, v p is the speed of the tracked robot at the center of mass C, ω p is the steering angular velocity of the tracked robot, v b and ω b are the reference driving speed and steering angular velocity of the crawler robot respectively; K1, K2, K3 are all coefficients greater than 0 and bounded; e θ is the heading angle error, e x and e y They are the x-axis and y-axis errors based on the actual position and reference position of the tracked robot, respectively.
