• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Learning Rat-Like Behavior for a Small-Scale Biomimetic Robot

    2022-02-13 09:54:30ZihangGaoGuangluJiaHongzhaoXieQiangHuangToshioFukudaQingShi
    Engineering 2022年10期

    Zihang Gao, Guanglu Jia, Hongzhao Xie, Qiang Huang, Toshio Fukuda,Qing Shi,*

    a Key Laboratory of Biomimetic Robots and Systems (Beijing Institute of Technology), Ministry of Education, Beijing 100081, China

    b Intelligent Robotics Institute, School of Mechatronical Engineering, Beijing Institute of Technology, Beijing 100081, China

    c Beijing Advanced Innovation Center for Intelligent Robots and Systems, Beijing Institute of Technology, Beijing 100081, China

    Keywords:Biomimetic Bio-inspired robot Neural network learning system Behavior generation

    A B S T R A C T Existing biomimetic robots can perform some basic rat-like movement primitives (MPs) and simple behavior with stiff combinations of these MPs.To mimic typical rat behavior with high similarity,we propose parameterizing the behavior using a probabilistic model and movement characteristics. First, an analysis of fifteen 10 min video sequences revealed that an actual rat has six typical behaviors in the open field, and each kind of behavior contains different bio-inspired combinations of eight MPs. We used the softmax classifier to obtain the behavior-movement hierarchical probability model. Secondly, we specified the MPs using movement parameters that are static and dynamic.We obtained the predominant values of the static and dynamic movement parameters using hierarchical clustering and fuzzy C-means clustering, respectively. These predominant parameters were used for fitting the rat spinal joint trajectory using a second-order Fourier series, and the joint trajectory was generalized using a back propagation neural network with two hidden layers. Finally, the hierarchical probability model and the generalized joint trajectory were mapped to the robot as control policy and commands, respectively.We implemented the six typical behaviors on the robot,and the results show high similarity when compared with the behaviors of actual rats.

    1. Introduction

    In recent years, the use of biomimetic robots has successfully promoted the understanding of a wide range of animal behaviors[1,2]. Using effective control methods and active guidance, biomimetic robots can interact with animals to observe and record their responses [3–7]. Sometimes, biomimetic robots are placed in a group of animals to explore their behavior patterns or verify scientific hypotheses in the process of interaction [8–11]. For biomimetic robots, lifelike behaviors can improve the effectiveness of their interactions with animals.For example,when robots simulate body movements related to foraging behavior, finches will spend more time on foraging[12].The tail swinging behavior of a robotic fish will have a substantial impact on the movement of fish [3].Moreover, biomimetic robots with similar behaviors can greatly improve the modeling of complex animal behaviors [13,14].

    The social activities of rats have aroused the interest of many researchers, and these studies have led to remarkable achievements [15,16]. Because of the advantages of biomimetic robots in animal interaction, various kinds of robots have been designed for the behavioral study of rats [17–21]. However, these robots have fewer degrees of freedom (DOFs) and lack flexibility in local movements,which makes it difficult to realize a natural interaction process. In our latest work, we simplified the DOFs of rats,extracted four DOFs in the pitch direction and three DOFs in the yaw direction, and used them to design the bionic spine mechanism of a robotic rat [22]. Although the robot can perform some basic rat-like movement primitives (MPs) with high flexibility, it is difficult for them to simulate the behavior characteristics of rats.Moreover, it is difficult for the robot to adapt to a variety of rat movement parameters and it does not have the ability to generalize movement.Therefore,we hope to both build a behavior model for rats and learn and generalize the spinal joint trajectory to control the robot so that it produces rat-like behaviors.

    Many modeling methods for animal behaviors have been proposed. Ding et al. [23] used a central pattern generator (CPG)-based method to model the behavior of amphibians and realize a multi-modal behavior design.Ren et al.[24]used a general internal model (GIM)-based biomimetic learning method to model the behavior of fish and achieved similar swimming patterns on a multi-joint robotic fish. These modeling methods are suitable for animals with rhythmic movement,whereas the movements of rats,especially the trajectory of spinal joints,are diverse and not rhythmic; hence, the above modeling methods are unsuitable.Leos-Barajas et al. [25] used hidden Markov models (HMMs) with hierarchical structures to model the behaviors of harbor porpoises and garter snakes.Cullen et al.[26]used non-parametric Bayesian methods to estimate the behavior of rostrhamus sociabilis more accurately than an HMM. HMM and Bayesian methods are based on probability distributions and transitions, which are more suitable for the analysis of rat behavior. However, the observed data in an open-field test revealed that the movement of rats is usually generated by a combination of MPs. It is challenging to deal with the combination and timing of these MPs using the abovementioned modeling methods. If the probability statistics of MPs are analyzed directly,it is difficult to capture the behavior characteristics of rats and will disrupt the inherent combinations of MPs.

    To better reflect the behavior characteristics of rats, we first combined the MPs in a time sequence, and the features of these combinations are reflected by a set of attributes.Each combination represents a rat movement. Using these attributes, we associated different combinations with different behaviors using the softmax classifier to learn the relationship between them.Using the classification results,we established the behavior–movement hierarchical probability model. In the hierarchical model, we not only represent the corresponding relationship between behaviors and movements and the transitions between different behaviors, but also the introduced states. This is because we hope to use states to further distinguish different behaviors and deepen our understanding of animal behaviors. Moreover, these different states may reflect different emotions in rats, which is an important consideration in future interaction experiments. The convergence and control of the model ensures its accuracy.

    In addition to behavior modeling,we need to learn the rules of rat movements.The MPs of rats can be described by static and dynamic parameters.The static parameters determine the movement amplitudes, average speeds, frequencies, or durations, whereas the dynamic parameters determine the angle and angular velocity of each spinal joint with respect to time.The MPs of rats usually consist of a variety of static parameters. Even for the same set of static parameters, different dynamic parameters may be present. To reduce computational complexity, we carried out cluster analysis on these parameters.We used hierarchical clustering and fuzzy Cmeans(FCM)clustering to extract the predominant values of the static and dynamic parameters,respectively[27,28].The predominant parameters were then used to fit the spinal joint trajectory of rats.Obviously, the fitting accuracy of joint trajectory depends on the results of cluster analysis.However,each clustering method has its own shortcomings.Hierarchical clustering does not need a predetermined number of clusters, but once a merge has been executed, it cannot be modified, and the quality of clustering is limited; the FCM clustering needs a predetermined number of clusters, and it often falls into locally optimal solutions[29,30].

    Additionally, we hope that learning rat movements will not remain at the level of copying,but lead to the ability to generalize.That is,we can infer a rat’s spinal joint trajectory for certain movement parameters through the supervised learning of the test samples. To improve the accuracy of movement generalization, we need a variety of movement parameters as the input of a neural network. This runs counter to our aim to reduce computation. To solve the uncertainty in clustering quality and balance the amount of computation with generalization accuracy,we used the Pearson correlation coefficient to measure the similarity between the robotic rat and rats.The results of clustering were modified to control the extraction of movement parameters.

    Although we have developed a similar quadruped robotic rat, it moves slowly and finds it difficult to achieve a high body pitch angle,which is not conducive to interaction with actual rats.Therefore,the hind limb of the robotic rat was transformed into a base that includes a hip servo motor and four wheels(two front driving wheels and two rearuniversalwheels).Whentheratmovesinastraightline,theangular speed of the driving wheel is determined by the average forward speed and radius ofthewheel.Whenthe rat turns,becausethe driving wheel only rotates around the axis and lacks a DOF for lateral movement, this will inevitably cause differences between the robotic rat and actual rats. To maximize the similarity of the robotic rat when turning, we used the learned rat spinal joint trajectory to drive the yaw joints of the robotic rat and control the difference in the rotation speed of the two driving wheels to generate the turning movement.

    The main contributions of this paper are as follows:

    (1) We regarded a rat movement as a temporal combination of MPs and defined the attributes of these combinations.Rat behavior is classified by different combinations of attributes. The hierarchical probability model better reflects the behavior characteristics of rats. This lays a foundation for future robot–rat interactions.

    (2) By training a neural network, we can predict the trajectory of each joint of the robot using only a set of static parameters,which greatly simplifies the movement planning process and facilitates the real-time planning of the robot.

    (3) We controlled the extraction process of the rat movement parameters by modifying the results of cluster analysis.This allows us to reduce the amount of calculation and ensure a high similarity between the robot and rats.

    2. Material and methods

    Our approach is summarized in Fig.1.In the process of behavior modeling,we first obtained the time series of MPs by tracking the coordinates of marked points. Then, we formed the movement of rats using MPs and combination attributes. A softmax classifier was used to classify the behaviors and states of the rats. Through the sequence of MPs and the results of classification, we established a hierarchical model of rat state, behavior, and movement in a probabilistic way.

    In the process of movement generalization, we extracted the static and dynamic movement parameters of each combination of MPs. These parameters were used to fit the trajectory parameters of the rat spinal joints (in the form of a Fourier series). To realize the generalization of movement, we established the mapping relationship between the static movement parameters and joint parameters using a back propagation (BP) neural network.

    In the deployment on the robot, we deduced the relationship between the static movement parameters and angular velocity of the driving wheel using the wheel model of the robot. The hierarchical model was provided to the robot as a control policy,so that the robot understood what kind of behavior and movement should be planned in a more natural and biomimetic way. Through the movement planning, we identified the probability distribution of the static movement parameters under each movement and obtained the trajectory of each joint,so that the robot can perform these behaviors and movements. The correlation coefficient was used to measure the similarity between the robot and rats and control the process of behavior classification and extraction of movement parameters in an iterative manner. In the following text,we describe each component in detail.

    2.1. Behavior modeling

    2.1.1. MPs, behaviors, and states of rats

    Fig.1. Generation process of rat-like behavior for the robotic rat.In the process of behavior modeling (light red box),we proposed a classification method for rat behaviors(B1,B2,???,B6)and states(S1 and S2),and obtained the hierarchical model of the states,behaviors,and movements.In the process of movement generalization(light blue box),we extracted the movement parameters of the rats and realized the mapping from static movement parameters to joint trajectory parameters.In the deployment on the robot(light green box),we controlled the robot by the policy and commands.The hierarchical model was used as the policy,and the generalized spinal joint angular displacement(θ1, θ2, ???, θ7) and wheel angular velocity (ω1, ω2) were used as the commands for the simulation model. The correlation coefficient between the robotic rat and rats(ρrat–robot) was used to measure the similarity and to regulate the process of behavior classification and movement parameter extraction. ui(xi, yi, zi): coordinates of the movement joints; λ: initial and transition probabilities between different states and behaviors as well as the observation probability of each combination in each kind of behavior; C1, C2, and C3: combination attributes of MPs; tanh: hyperbolic tangent.

    An open-field test is used to study the spontaneous activity of animals in a new environment. It allows experimental animals to move freely in a certain space with few restrictions and has become an important method for studying the behavior and movement of animals [31,32]. Biologists usually divide the MPs of rats into motions such as going straight,turning,standing up,and staying in one place.Correspondingly,the behaviors of rats are divided into behaviors such as exploring, grooming, and resting to reflect the different activity characteristics of rats [33,34]. In this paper,we observed the activities of individual rats (Rattus norvegicus) in a 1 m×1 m×1 m open field,recorded 10 min daily,and obtained a dataset of the motions of three rats over five days. The rats involved in this study were from the same litter and were seven weeks old. The dataset (150 min) contains 3537 rat movements.

    Here,we marked the nose tip,tail,and key movement joints of the rats, and tracked the coordinates of these points through multiple cameras. We defined the MPs based on the coordinate changes of the marked points.The pitch and yaw of the head joint are defined as MP1and MP2, respectively; the pitch and yaw of body joints are defined as MP3and MP4, respectively; the linear motion of all marked points is defined as MP5; the rotational motion of all marked points is defined as MP6; almost no motion in the coordinates of all marked points is defined as MP7; and the swing of the forelimb joint is defined as MP8. The MPs, behaviors, and states of the rats are presented in Table 1.

    2.1.2. Combinations of MPs and classification of behaviors and states

    An observed sequence of MPs is shown in Fig.2.The sequence of MPs reflects the activity sequence of the rats,and the movements of the rats can be represented by a temporal combination of the above MPs. We extracted the combinations of MPs using the following method:First,we calculate the distance between MPs.A value of 0 indicates that the MPs occur simultaneously.We further extract all MPs that occur simultaneously and form candidate combinations(including the combination that contains only one MP). Secondly,each combination is combined with the subsequent MP or combination(with a distance value of 1)to form a new combination and the frequency of the new combination is calculated.If the frequency of the new combination is very high, the new combination replaces the original combination;if the frequency of the new combination is high, both combinations are retained; if the frequency of the new combination is low,the new combination is discarded.The second step is repeated until no new combination appears.

    We use combination attributes to reflect the features of the combinations of these MPs.Here,C1represents the repeatability of MPs,where 1 represents repetition and 0 represents non-repetition; C2represents the number of MPs generated simultaneously; and C3represents the number of MPs in the combination(juxtaposed MPs count as only one).Then,we classified the behaviors and states on the basis of different combinations of MPs(Fig.2).

    We used supervised learning to obtain the relationship between the combination of MPs and behaviors/states. A labeled dataset was collected to train the softmax classifier.The inputs of the network are MPs (1 indicates that the MPs are activated, whereas 0 indicates that they are not activated) and combination attributes,and the outputs are the corresponding behaviors and states. The number of nodes in the hidden layer is eight. Here, we assumed that different behaviors and states are mutually exclusive, which is conducive to robotic rat control.

    2.1.3. Hierarchical model of states, behaviors, and movements

    The results of the softmax classifier were used to obtain the hierarchical model of states, behaviors, and movements (Section 3.1). Here, λ denotes the initial and transition probabilities between different states and behaviors as well as the observation probability of each combination in each kind of behavior. This hierarchical model effectively reflects the activity characteristics of the rats in the observation. The state layer of the structure is aMarkov model and includes an initial probability,a probability for transition to its own state,and a probability for transition to other states; for the behavior layer, because a continuous behavior is represented by a specific combination of MPs,it only has an initial probability and a probability for transition to other behaviors; for the movement layer,because there is not necessarily a connection between the combinations of MPs corresponding to different behaviors, which is extremely random and dominated by the probability of transition between different behaviors, only an the initial probability(the observation probability)is given.The hierarchical model becomes gradually more refined from top to bottom as the probability system becomes simpler.In the following work,we analyze the convergence of the model. We also adjust the model using the correlation between the robot and rats.

    Table 1 MPs, behaviors, and states of rats.

    Fig. 2. Observed MP sequence. C1, C2, and C3 correspond to different behaviors and states.

    2.1.4. Model convergence

    Information entropy solves the problem of the quantitative measurement of information[35].In this study,we calculated the information entropy of the states and behaviors in the hierarchical model to determine whether the model tends to be stable with respect to probability. We first counted 1700 rat movements,mapped these movements to different states and behaviors, and then calculated the number of occurrences and probability of each state and behavior. Next, the probability distribution of each state and behavior was recalculated for each additional 200 rat movements.When the probability distribution of each state and behavior hardly changed, the information entropy of the states (H(S)) and behaviors (H(B)) also become stable, indicating that the model had converged.The calculation of information entropy is as follows:where nSand nBrepresent the number of states and behaviors in the model, respectively; πSiand πBiare probability of each state and behavior in the model, respectively.

    The calculation results are shown in Fig. 3. The information entropy of states and behaviors in the model is stable at around 3500 rat movements. The increase in information entropy indicates that the difference in probability between different states and behaviors is decreasing. This may reflect the adaptability of rats to a new environment. In the early data, the probabilities of stressed states and exploratory behavior are much higher than the other states; as the amount of data increases, the probability of relaxed states,resting,grooming,and other rat behaviors gradually increased. After the model converged, we stopped collecting data from the rats.

    2.1.5. Model regulation

    The Pearson correlation coefficient reflects the linear correlation between two sets of data, and it has been successfully used

    Fig. 3. Information entropy of states and behaviors.

    To achieve rat-like behavior in the robotic rat,we required that the behavior correlation coefficients be greater than 0.8.When the correlation coefficient of a certain kind of behavior is below this threshold, we first adjust the extraction process of the movement parameters (Section 2.2); if it still fails to meet the threshold, the combinations to which the behavior belongs are split so that they corresponding to different new behaviors, and then the behaviors are reclassified. A realistic example is described in Section 4.

    2.2. Movement generalization

    Our parametric modeling of rats is shown in Fig. 4. The red markers indicate the main DOFs in the pitch direction, the blue markers indicate the main DOFs in the yaw direction,and the black markers indicate the nose tip,tail end,and center of mass(CoM)of the rats.A coordinate system was set up with respect to the tail of the rats, and the Kinovea software was used to track the location information of the marked points.

    We define the static movement parameters,and all angles refer to the movement amplitudes of rats. The pitch angle is zero in the horizontal position,positive upward,and negative downward.The yaw angle is zero at the symmetrical position(where the angles of the rat head and body are in a straight line), the direction is random, and the value is positive (it was observed in the data that the probabilities of rat left and right yaw are roughly equal).Parameters head pitch angle φhpand head yaw angle φhyrepresent the angle between the x-axis and the line from the nose tip to joints J7and J6, respectively. Parameters body pitch angle φbpand body yaw angle φbyrepresent the angle between the x-axis and the line from the CoM to the tail end.Parameter turing angle α represent the rotation angle of the x-axis before and after the turning movement. Parameter φfsrepresents the rat forelimb swing angle,v represents the average forward speed of the rat CoM,f represents the movement frequency (only for combinations with repetitive MPs), and T represents the movement duration.

    Fig. 4. Parametric modeling of rats. The red markers (J1, J2, J5, and J7) and blue markers(J3,J4,and J6)indicate the main DOFs in the pitch and yaw directions of the rats, respectively, and the black markers are used as the reference points to define the static movement parameters.

    To extract the predominant values of the static parameters,because of the small sample size and the difficulty in determining the number of clusters in advance,we used hierarchical clustering to classify these static parameters. First, we treated the different values of each static parameter as a class, and calculated the distances between classes. The two closest classes (i - 1, j - 1) were merged into one,and the average value of the original two classes was taken as the value of the new class. Afterwards, the distances Di,jbetween the new class(j)and other classes(i)were calculated,until the distance between the two classes met the following condition:

    where Amaxdenotes the maximum value of each static parameter and τ determines the density of classification. Finally, the static parameter value corresponding to each class in the clustering results is the extracted dominant value. When the probability corresponding to the predominant values of a static parameter is calculated, it is often affected by other parameters. Therefore, it is necessary to calculate the independence between the two parameters and determine the joint statistics of two non-independent parameters to form the joint probability distribution of the predominant values; otherwise, separate statistics for each parameter are obtained to form the independent probability distributions of the predominant values.

    To extract the predominant values of the dynamic parameters,because of the large amount of data and accurate classification requirements, we used FCM clustering to classify the dynamic parameters. In this paper, the dynamic parameters (θi,j) mainly reflect the position characteristics of the main DOFs over time.For a joint in a certain kind of movement, we can calculate its angular displacement and angular velocity at a discrete time by the marked coordinate values. Because this kind of movement occurs many times, we obtained the data group xj′= (θj′, θ′j′) (j′=1, 2, ???, n) of the joint under a specific set of static parameters.Here, θj′and θ′j′are the angular displacement and velocity of joint Jiat time tj, respectively. The process of cluster analysis is expressed as follows:

    tive function, ui′j′denotes the membership degree of sample xj′belonging to class i′, m is the factor of membership degree, which affects the degree of classification, and ci′denotes the clustering center of class i′. The number of clusters c is determined by the average value of the contour coefficient of each sample point.When the average value of the contour coefficient is closer to 1,this indicates the clustering is better.Parameters ci′and ui′j′iterate continuously,until the objective function O is less than a certain threshold.To better reflect the movement characteristics of rats, we used the weighted method for each sample instead of the average method for the predominant dynamic parameters.The weight of each sample was determined by the proportion of the number of samples of each class as follows:

    where parameter Ci′denotes the set of all samples of class i′and Ni′denotes the number of samples contained in class i′.

    Because functions satisfying Dirichlet’s condition can be approximately expressed by a finite Fourier series,we used a Fourier series to simulate the angular displacement curve of each joint in rats[38].After experiments,we found that a second-order Fourier series is sufficiently accurate (R-square > 0.9). Taking the pitching movement(MP1MP3)under exploring behavior(B2)as an example,the input and output are normalized and the following formula is obtained:

    where, wiand ki1–ki5are fitted joint trajectory parameters (i = 1, 2,5, and 7). For a repetitive movement such as MP*1, wiare represented by frequency f.

    A neural network is usually used to generalize joint movements[39–41].In this study,we adopted the BP neural network with two hidden layers.The inputs of the network are static parameters,and the outputs are joint trajectory parameters.The hyperbolic tangent(tanh) activation function is used to accelerate the convergence of the network.According to previous studies [42,43], the number of nodes in the hidden layer should be between the number of input nodes and output nodes, and the same number of neurons is used for all hidden layers. In this study, for each combination of MPs, a corresponding neural network was used to participate in the training process. We adjusted the number of nodes in the hidden layer to avoid the phenomena of underfitting or overfitting. Still taking the pitching movement (MP1MP3) under exploring behavior (B2)as an example, the number of input nodes of the network was two (φbp, T), the number of output nodes was six (wi, ki1–ki5),and the number of nodes in the hidden layer was four (as shown in Fig. 5). The cost function was the mean squared error loss function and the network was trained by a batch sample parallel gradient algorithm. The clustering results were modified to adjust the static and dynamic parameters. In the initial setting, to reduce the complexity and computation of the model, we used a larger value τ and a smaller value c. When the correlation coefficient of a certain kind of behavior is below a threshold, the clustering results of movements are modified in descending order of observation probability while reducing the τ value and increasing the c value.

    2.3. Deployment on the robotic rat

    2.3.1. Robotic rat platform

    Fig. 5. BP neural network used in the pitching movement (MP1MP3) under exploring behavior (B2).

    The prototype of our recently developed robotic rat is shown in Fig.6(a).The robotic rat mainly consists of a head,forelimb,waist,and hip. The head and hip joints are driven by a servo motor, the waist joints and wheels are driven by a direct current (DC) motor,and the forelimb is driven by a micro deceleration stepper motor.Its shape and size are similar to those of actual rats, and the total mass is approximately 400 g. The spinal joints of the robot correspond to the main DOFs of the rats. Fig. 6(b) and Table 2 show the distribution of each joint of the robotic rat and the corresponding relationship between the MPs and joints.

    The movement parameters of the robot are mainly affected by the movement parameters of rats and the structural constraints of the robot. The task space constraints on the MP mean that by analyzing the workspace of the robot and the interference between joints, we can obtain the amplitude of the movement parameters of the robot, such as the maximum body pitch angle and body yaw angle. We described this analysis in detail in Ref. [22]. Concerning the constraints on the control,the robot needs to maintain dynamic balance in movements without vibration or sideslip.Therefore, the robot usually has a minimum movement duration under various conditions. We analyzed this in detail in Ref. [44].When determining the movement parameters of the robot, we depend on whether the extracted parameters of rats exceed the limits of the robot itself.If the limit is not exceeded,the movement parameters of the robot are equal to those of the rat;if the limit is exceeded, the movement parameters of the robot are equal to the maximum or minimum value of the limit. Hence, we defined the parameters of the robot movements as

    where U and Ceqrepresent the task space and dynamic balance constraints,respectively;(φbp)maxand(φby)maxrepresent the maximum body pitch angle and body yaw angle satisfying the task space constraints,respectively;and Tminrepresents the minimum movement duration satisfying the dynamic balance constraints.Using the prototype,we built a rigid-body simulation model in the robot operating system(ROS)environment(Gazebo simulator)and used it in the training process.

    2.3.2. Wheel model of the robotic rat

    Fig. 6. (a) Robotic rat and (b) motion coordinate system of the robotic rat. JF: forelimb swing joint; JWL, JWR: left and right drive wheel joint, respectively.

    Table 2 Joints of robotic rat used to perform the MPs.

    Fig. 7. Turning model of the robotic rat. Ii¨θi is the moment applied to link i.

    The static friction and driving force of the wheels are the reaction forces,which are equal in value and not more than the sum of the maximum static friction on each wheel.The angular velocity of the CoM caused by the difference in rotation speed of the driving wheels is generated by the normal static friction. Based on the statistics of the static parameters of rats,the robotic rat must reach turning angle α within duration T. Assuming that the rotational speeds of the driving wheels a and d are ω1and ω2, respectively;τ1and τ2are the actual output torques of the two driving wheels;and P1and P2are the actual output power of the two driving wheels. We then obtain the set of equations in Eq. (11):

    where rdis the radius of the driving wheel, k1and k2are power factors.

    Because Fcoand γ can be expressed by the angular displacement θiof each yaw joint, which depends on parameters α and T, the relationship between the driving wheel rotational speeds ω1and ω2and the parameters α and T can be obtained using Eqs. (10)and (11).

    3. Results

    3.1. Probability distribution and transition in hierarchical model

    On the basis of behavior and state classification,we determined the probabilities of the states, behaviors, and movements by the number of occurrences, as shown in Fig. 8(a). Among them, the dotted lines represent the initial probability, the solid lines represent the transition probability, and different colors represent different probability intervals. For walking behavior, which includes going straight, turning, and staying MPs, because the behavior can include various combinations with different numbers of MPs,we created separate probability statistics. Fig. 8(b) shows the initial probability(black dotted line)and transition probability(black solid line) of each MP in walking behavior; Fig. 8(c) shows the probability distribution of the number of MPs in the combinations included in walking behavior.

    3.2. Distribution of static parameters

    Table 3 shows the relevant static parameters used to describe the MPs and combinations of MPs. The parameters in brackets are not independent of each other. For different behaviors, the same MPs and static parameters often appear. Fig. 9 presents the distribution of the static parameters in different behaviors.Fig.9(a)compares average forward speed during walking and trotting.In both cases, the average forward speed has a similar range(0.125–0.250 m?s-1),but a high speed when trotting is more likely.Fig. 9(b) compares the staying times of walking and resting. The staying time in walking is short,approximately 0.5–1.5 s,whereas the staying time is longer in resting,approximately 3–7 s.Fig.9(c)shows the distribution of the turning angle and movement duration in walking and trotting. The upper-right ‘‘*” indicates that the data has a high probability of occurrence(the sum of the occurrence probability of all data with ‘‘*” is greater than 0.7). It can be seen that the turning movement when trotting has a faster average angular velocity than that when walking.The distributions of these parameters support the rationality of the behavior classification.

    3.3. Generalizable movement planning of the robotic rat

    Fig. 8. Probability distributions and transitions in the hierarchical model. (a) The different colors correspond to different probability intervals, the juxtaposition of MPs indicates that these MPs are generated simultaneously, the symbol ‘‘–” indicates the time sequence of MPs and ‘‘*” indicates that the MPs occur repeatedly in the combination;(b)initial probability(black dotted line)and transition probability(black solid line)of each MP in walking behavior;(c)probability distribution of the number of MPs in the combinations included in walking behavior.

    Table 3 Static parameters used to describe the MPs/combination of MPs.

    Fig.10 presents the main joint trajectory curves of the pitching,yawing,and turning movements.The light red areas represent the envelope of the test sample trajectory, and the black solid lines represent the mean predicted value of the test sample trajectory given by the neural network based on the training samples. For each movement, the number of training samples is approximately twice that of the test samples. For pitching and yawing movements,because the lifting/yawing time accounts for approximately half of the total time, and the lifting/yawing trajectory is basically symmetrical to that of the other half (falling/returning);we hence only give the trajectory of each joint of the robotic rat from t=0 to T/2.It can be seen that the black solid line is almost always within the red area, which indicates that the neural network can realize the generalization of rat movement. In the pitching movement,the trajectory distribution of each joint is relatively concentrated,which indicates that the movement of the pitching joints of rats under different static parameters is relatively similar, and the displacement mainly occurs in the hip and front waist.In the yawing and turning movements, the distribution of the head joint trajectory is relatively loose,which indicates that the head movement is complex and diverse. However, the trajectory distributions of other joints are more concentrated; moreover, the head accounts for a small proportion of body length, and hence the head joint error is acceptable. The displacement mainly occurs in the head and front waist.

    Fig. 11 shows the angular velocity of the driving wheel in the turning movement of the robotic rat.The red dotted lines represent the theoretical calculation results,and the black dots represent the value obtained in the simulation environment (ω1= 0). The simulation value is close to the theoretical value. Moreover,ω2is positively correlated with α and negatively correlated with T.

    3.4. Similarity between the robotic rat and actual rats

    Fig.9. Distributions of the static parameters in different behaviors. (a)Comparison of average forward speed in walking and trotting; (b)comparison of the staying time in walking and resting; (c) distribution of the turning angle and movement duration in walking and trotting. The label ‘‘*” indicates that the data has a high probability of occurrence.

    Fig.10. Joint trajectory curves of the robotic rat.(a)Trajectory curves of pitching joints for B2,MP1MP3,the horizontal axis is the proportion of pitching cycle T and the vertical axis is the proportion of body pitch angle φbp;(b)trajectory curves of yawing joints for B2,MP2MP4,the horizontal axis is the proportion of yawing cycle T and the vertical axis is the proportion of body yaw angle φby;(c)the trajectory curves of turning joints for B3,MP6.The horizontal axis is the proportion of turning cycle T and the vertical axis is the proportion of turning angle α.

    Fig. 11. Angular velocity of the driving wheel.

    In the simulation and experiment,we first determined the robot the corresponding behavior and combination of MPs based on the obtained behavior-movement hierarchical model.Furthermore,for each combination of MPs, we set its static movement parameters according to the results of clustering analysis. From these static movement parameters, we obtained the generalized trajectory of each joint using the BP neural network.In fact,we have established a robot joint motion library.Once the static movement parameters have been determined,we can obtain the trajectories of each joint using a look-up table. Thus, we obtained the joint state (position,velocity,and acceleration)planned by the robot in advance.In the actual control,we calculated the joint state of the robot in real time and compared it with the preplanned joint state. According to the difference between them, the joint torque was calculated through impedance control, and then each joint of the robot was driven to produce a new joint state. We controlled the robot to produce 50 movements in each group,recorded a total of ten groups,and then calculated the correlation coefficients of various behaviors and movements.

    Fig. 12 shows behaviors and movements of the robot over 10 s and the comparison results with the behaviors and movements of actual rats. The results show that the robot has movements that are often relatively similar to those of actual rats. Fig. 13 shows the relationship between correlation coefficient ρrat–robotand τ,and c in cluster analysis (taking B2, MP1MP3as an example). Coefficient τ represents the scale coefficient in hierarchical clustering,which is used to control the density of classification; c represents the number of clusters in FCM clustering.As τ increases,the correlation coefficient first decreases slowly, and then decreases sharply. Here, τ = 0 indicates that each data point is a class, whereas τ = 1 means that all data points are only in one class. When τ = 0, although it has a higher correlation coefficient, the number of samples of the neural network also increases, which increases the training cost of the model. Therefore, we believe that τ = 0.1 is a better choice to reduce the training cost of the model while maintaining a high correlation coefficient. For c, c = 1 and c = n(n = 15, the number of data points) mean that all data points are averaged to obtain the dynamic parameters, which will reduce the correlation coefficient. When c is exactly between 1 to n, the correlation coefficient is higher, indicating that the weighted method is better than the average method. For B2, MP1MP3; c = 3 is a better choice.

    Fig. 12. Behaviors and movements of the robot over 10 s.

    Fig. 13. Relationship between ρrat–robot,τ, and c in cluster analysis.

    Fig. 14. Correlation coefficients of different behaviors and movements. (a) Movement correlation coefficients in exploration, the sizes of the sectors reflect the observation probability of movements during exploration and the data in every sector are expressed as mean plus or minus SD;(b)correlation coefficients of behaviors,error bars show one SD. SD: standard deviation.

    The correlation coefficients of different behaviors and movements between the robot and rats are shown in Fig. 14. Because the joints of MP7and B5do not produce displacement, their correlation coefficient was not calculated. Fig. 14(a) shows the movement correlation coefficients during exploration.The sizes of the sectors reflect the observation probability of movements during exploration. The data in every sector are expressed as mean plus or minus standard deviation (SD). Because of the complexity and diversity of head yaw movement in rats, the correlation coefficient of pitching movement (MP1MP3) is higher than that of the yawing movement (MP1MP2, MP2MP4), with a small SD. Fig. 14(b)shows the correlation coefficients of different behaviors. Strongly regular behaviors, such as B1and B6, which only include one repetitive movement, have a higher correlation coefficient and smaller deviation. In contrast, because of the complexity and diversity of combinations in B3,the correlation coefficient is lower with larger deviation.For more complex and diverse behaviors and movements,more rat data need to be collected and used for training to further improve the correlation coefficients between the behaviors of the robotic rat and rats.

    4. Discussion

    In the process of behavior classification,we initially divided the behaviors into sniffing, exploring, walking, resting, and grooming.Because of the low correlation coefficient,we subdivided the walking behavior. We made a distinction between continuous straight walking or turning and stop-and-go movement in rats, and matched the four combinations of MPs (MP5, MP6, MP5MP6, and MP6MP5) to the trotting behavior. The above combinations did not include staying (MP7) and had a relatively quick velocity(Figs. 9(a) and (c)). Next, the behaviors were reclassified, and the movement parameters in walking and trotting were extracted separately. After reclassification, we obtained higher correlation coefficients for the walking and trotting behaviors.Because we only changed part of the original walking behavior into trotting behavior without changing the order of the rat observations, the transition probability between the walking and trotting behaviors is zero when counting the transition probabilities between behaviors.

    Different behaviors and states in the hierarchical model have obvious characteristics.Sniffing is mainly manifested in the movement of repeatedly touching the ground with the tip of the rat’s nose. Exploring is mainly manifested in the pitching and yawing movements of the head and body.Walking is mainly characterized by slower straight and turning speeds,accompanied by short staying periods. In addition, the number of MPs in a combination corresponding to this behavior is usually more than three. Trotting is mainly characterized by faster straight and turning speeds. The number of MPs in a combination corresponding to this behavior is generally no more than two. Resting is mainly manifested by the long staying of rats. Grooming is mainly manifested in the slight head and forelimb movements of rats. The stressed state is mainly indicated by obvious displacement or frequent sniffing and exploring behaviors. The relaxed state is mainly indicated by no obvious displacement,with no movement or only partial,small movements of the head and forelimbs.

    Because the limbs of rats rarely move in the reverse direction during the turning movement, we did not drive the robotic rat model in reverse when designing the rotation speeds of the driving wheels to avoid reducing the biomimicry of the robotic rat. At the same time, because of the difference between the wheeled base of the robotic rat and the limbs of rats,the turning radius of the robotic rat is larger than that of actual rats. Under the same driving wheel speed differences,the turning radius of the robotic rat increases substantially as the driving wheel speeds increase.Hence,fixing ω1=0 and only controlling ω2to coordinate the motion of the yaw joints in the turning movement is a more reasonable design.

    5. Conclusions

    For biomimetic robots, similar behavior models are more conducive to effective interaction between robots and animals. The main focus of this study was to establish the behavior–movement law of rats by extracting different combinations of MPs and corresponding them to different behaviors. The predominant parameters of the MPs were extracted, and the trajectory of each joint was learned and generalized. The correlation coefficient between the robot and rats was used to measure the similarity and control the process of behavior classification and the extraction of movement parameters. A robotic rat model in the ROS environment was used to learn the data and train the robot. The simulation results show that the robot can achieve six typical rat-like behaviors. For each behavior, the robot presents high similarity (the behavior correlation coefficient is greater than 0.8).

    In future work,we will implement two modes in the robot:individual mode and interaction mode.In the individual mode,the robot moves based on the law given by the behavior–movement hierarchical model in this paper; in the interaction mode, the robot recognizes the current behavior of the interactive target, and then performs interactive behaviors such as tracking, imitating and contacting. Whether the robot operates in the individual mode or interaction mode will be determined by the interaction probability,that is,the observed probability of interaction between two rats over a series of activities. The behavior and movement of the robot in interaction mode can be further defined on the basis of the approach in this paper. Through interactive control between the robot and rats, we can explore the social behavior mechanism of rats.

    Acknowledgments

    This work was supported in part by the National Natural Science Foundation of China (62022014) and in part by the National Key Research and Development Program of China(2017YFE0117000).

    Compliance with ethics guidelines

    Zihang Gao, Guanglu Jia, Hongzhao Xie, Qiang Huang, Toshio Fukuda,and Qing Shi declare that they have no conflict of interest or financial conflicts to disclose.

    亚洲天堂国产精品一区在线| 国模一区二区三区四区视频| 波多野结衣高清作品| 天堂中文最新版在线下载 | 日韩大尺度精品在线看网址| 欧美成人a在线观看| 国产蜜桃级精品一区二区三区| 国产精品国产高清国产av| 蜜桃久久精品国产亚洲av| 一区二区三区免费毛片| 日韩欧美三级三区| 日本-黄色视频高清免费观看| 日韩欧美在线乱码| 干丝袜人妻中文字幕| 一边亲一边摸免费视频| 亚洲自拍偷在线| 22中文网久久字幕| 亚洲无线观看免费| 麻豆av噜噜一区二区三区| 一夜夜www| 国产男人的电影天堂91| 国产高清激情床上av| 亚洲一区二区三区色噜噜| 特级一级黄色大片| 免费看日本二区| 亚洲av一区综合| 精品一区二区免费观看| 国产高潮美女av| 亚洲人成网站在线播放欧美日韩| 免费黄网站久久成人精品| 在现免费观看毛片| 91久久精品国产一区二区成人| 久久久久久久午夜电影| 亚洲美女视频黄频| 欧美日本亚洲视频在线播放| 国产熟女欧美一区二区| 免费看光身美女| 亚洲精华国产精华液的使用体验 | 国产成人a区在线观看| 少妇裸体淫交视频免费看高清| 亚洲成人久久性| 乱码一卡2卡4卡精品| 欧美色视频一区免费| АⅤ资源中文在线天堂| 少妇人妻一区二区三区视频| 国内精品一区二区在线观看| 天堂网av新在线| 别揉我奶头 嗯啊视频| 国产单亲对白刺激| 三级毛片av免费| 国产人妻一区二区三区在| 欧美xxxx黑人xx丫x性爽| 久久久精品欧美日韩精品| 国产精品三级大全| 99久久久亚洲精品蜜臀av| 成人综合一区亚洲| 又爽又黄无遮挡网站| 少妇猛男粗大的猛烈进出视频 | 欧美日本亚洲视频在线播放| 婷婷精品国产亚洲av| 99国产极品粉嫩在线观看| 日韩,欧美,国产一区二区三区 | 桃色一区二区三区在线观看| 国产老妇伦熟女老妇高清| 一区二区三区四区激情视频 | 亚洲精品乱码久久久v下载方式| 在线免费十八禁| 久久韩国三级中文字幕| 成人午夜高清在线视频| 黄色视频,在线免费观看| 91久久精品国产一区二区三区| 精品午夜福利在线看| 亚洲精品成人久久久久久| 欧美不卡视频在线免费观看| 赤兔流量卡办理| 国产成人精品一,二区 | 国内少妇人妻偷人精品xxx网站| 成人欧美大片| 又黄又爽又刺激的免费视频.| 中出人妻视频一区二区| 日本撒尿小便嘘嘘汇集6| 在线免费观看不下载黄p国产| 日韩 亚洲 欧美在线| 给我免费播放毛片高清在线观看| 99riav亚洲国产免费| 日韩成人av中文字幕在线观看| 欧美日韩综合久久久久久| 高清在线视频一区二区三区 | avwww免费| 久久鲁丝午夜福利片| 中文字幕制服av| 免费观看的影片在线观看| 夜夜爽天天搞| 国国产精品蜜臀av免费| 国产黄a三级三级三级人| 97超视频在线观看视频| 91麻豆精品激情在线观看国产| 我的女老师完整版在线观看| 高清日韩中文字幕在线| 美女国产视频在线观看| 国产精品一及| 免费不卡的大黄色大毛片视频在线观看 | 黄色一级大片看看| 日韩视频在线欧美| 99热全是精品| 老司机影院成人| av卡一久久| or卡值多少钱| 亚洲av免费在线观看| 深夜a级毛片| 久久人人精品亚洲av| 神马国产精品三级电影在线观看| h日本视频在线播放| 欧美日韩国产亚洲二区| 成熟少妇高潮喷水视频| 午夜激情福利司机影院| 日韩 亚洲 欧美在线| 99久久无色码亚洲精品果冻| 99国产精品一区二区蜜桃av| 免费搜索国产男女视频| 日韩高清综合在线| 亚洲欧美中文字幕日韩二区| 日日干狠狠操夜夜爽| 深夜精品福利| 搡女人真爽免费视频火全软件| 亚洲国产精品sss在线观看| 亚洲国产精品sss在线观看| 99久久久亚洲精品蜜臀av| 国产极品精品免费视频能看的| 国产精品一区二区性色av| 亚洲欧美精品自产自拍| 国产午夜福利久久久久久| 成人毛片60女人毛片免费| 国产爱豆传媒在线观看| 中文字幕av成人在线电影| 久久欧美精品欧美久久欧美| 蜜臀久久99精品久久宅男| 免费一级毛片在线播放高清视频| 久久精品国产亚洲网站| 一边摸一边抽搐一进一小说| 深夜精品福利| 日日啪夜夜撸| 黄色日韩在线| 中文字幕av在线有码专区| 赤兔流量卡办理| 十八禁国产超污无遮挡网站| 校园春色视频在线观看| 国产三级中文精品| 两个人的视频大全免费| 日本-黄色视频高清免费观看| 久久精品91蜜桃| ponron亚洲| 久久热精品热| 成人毛片a级毛片在线播放| 国产毛片a区久久久久| 亚洲性久久影院| 可以在线观看毛片的网站| 最近最新中文字幕大全电影3| 亚洲久久久久久中文字幕| 欧美色欧美亚洲另类二区| 51国产日韩欧美| 麻豆成人午夜福利视频| 熟女人妻精品中文字幕| 午夜激情欧美在线| 国产美女午夜福利| 午夜精品在线福利| 最近的中文字幕免费完整| 亚洲va在线va天堂va国产| 91久久精品国产一区二区成人| 18禁在线无遮挡免费观看视频| 国产亚洲精品久久久com| 51国产日韩欧美| 国产精品蜜桃在线观看 | 国内久久婷婷六月综合欲色啪| 精品人妻一区二区三区麻豆| 久久精品国产99精品国产亚洲性色| 国产精品伦人一区二区| 99热这里只有是精品在线观看| 欧美3d第一页| 女的被弄到高潮叫床怎么办| 中文字幕制服av| 国产又黄又爽又无遮挡在线| 小蜜桃在线观看免费完整版高清| 大香蕉久久网| 99久久久亚洲精品蜜臀av| 少妇人妻一区二区三区视频| 国产精品一及| 淫秽高清视频在线观看| av视频在线观看入口| 久久久色成人| 午夜亚洲福利在线播放| 一级二级三级毛片免费看| a级毛片a级免费在线| 国产成人影院久久av| 午夜激情福利司机影院| 亚洲五月天丁香| videossex国产| 激情 狠狠 欧美| 国产精品国产三级国产av玫瑰| 午夜a级毛片| 久久中文看片网| 中文精品一卡2卡3卡4更新| 精品久久国产蜜桃| 亚洲av第一区精品v没综合| 一级毛片我不卡| 日韩一区二区三区影片| 在线播放国产精品三级| 午夜免费男女啪啪视频观看| 黄色一级大片看看| 在线a可以看的网站| 精品不卡国产一区二区三区| 国产成人精品久久久久久| 成人亚洲欧美一区二区av| 寂寞人妻少妇视频99o| 高清午夜精品一区二区三区 | 国产精品久久久久久精品电影| 99久久精品热视频| 免费av毛片视频| 久久中文看片网| 国产av不卡久久| a级毛片免费高清观看在线播放| 最近视频中文字幕2019在线8| 成人午夜高清在线视频| 久久久色成人| 国产精品一区www在线观看| 久久久久性生活片| 亚洲第一电影网av| 国产午夜精品一二区理论片| 久久久久免费精品人妻一区二区| 看十八女毛片水多多多| 人妻夜夜爽99麻豆av| 精华霜和精华液先用哪个| 亚洲内射少妇av| 女人十人毛片免费观看3o分钟| 亚洲va在线va天堂va国产| 秋霞在线观看毛片| 国产精品久久久久久久久免| 观看免费一级毛片| 国产成人午夜福利电影在线观看| 岛国毛片在线播放| 悠悠久久av| 男女下面进入的视频免费午夜| 精品久久久久久久久亚洲| 黄片wwwwww| 天堂√8在线中文| 卡戴珊不雅视频在线播放| 亚洲国产精品成人久久小说 | 小说图片视频综合网站| 欧美日本亚洲视频在线播放| 亚洲av免费在线观看| 精品久久久噜噜| 搞女人的毛片| 欧美潮喷喷水| 波多野结衣巨乳人妻| 亚洲在线观看片| 精品久久久久久久久久久久久| 亚洲图色成人| 高清毛片免费看| 极品教师在线视频| 99热这里只有是精品在线观看| 亚洲av成人精品一区久久| 久久久久免费精品人妻一区二区| 成人特级黄色片久久久久久久| 国产精品福利在线免费观看| 国产成人精品婷婷| 99国产精品一区二区蜜桃av| 精品人妻熟女av久视频| 级片在线观看| 日日摸夜夜添夜夜添av毛片| 美女脱内裤让男人舔精品视频 | 日韩一本色道免费dvd| 一进一出抽搐gif免费好疼| 少妇人妻一区二区三区视频| 26uuu在线亚洲综合色| 日韩欧美三级三区| 18禁在线播放成人免费| 性色avwww在线观看| 青春草亚洲视频在线观看| 国产黄色小视频在线观看| 最近手机中文字幕大全| 国产视频内射| av黄色大香蕉| 婷婷精品国产亚洲av| 亚洲成人中文字幕在线播放| 国产精品爽爽va在线观看网站| 成人午夜高清在线视频| 亚洲成人久久爱视频| 亚洲丝袜综合中文字幕| 亚洲人成网站在线播放欧美日韩| 婷婷六月久久综合丁香| 99国产精品一区二区蜜桃av| 特大巨黑吊av在线直播| 亚洲在久久综合| 精品人妻视频免费看| 亚洲aⅴ乱码一区二区在线播放| 99在线视频只有这里精品首页| 国内久久婷婷六月综合欲色啪| 精品久久久久久久人妻蜜臀av| 三级男女做爰猛烈吃奶摸视频| 看非洲黑人一级黄片| 免费人成视频x8x8入口观看| 色尼玛亚洲综合影院| 看十八女毛片水多多多| 男女下面进入的视频免费午夜| 69人妻影院| 1024手机看黄色片| 亚洲人成网站高清观看| 免费观看在线日韩| 亚洲av成人av| 成人av在线播放网站| 国产探花在线观看一区二区| 一个人看的www免费观看视频| 天天躁夜夜躁狠狠久久av| 小说图片视频综合网站| 久久欧美精品欧美久久欧美| 久久精品国产亚洲av香蕉五月| 国产成年人精品一区二区| 少妇人妻一区二区三区视频| av在线观看视频网站免费| 成人毛片a级毛片在线播放| 国产老妇女一区| 三级经典国产精品| 91精品一卡2卡3卡4卡| 性欧美人与动物交配| 夜夜夜夜夜久久久久| 亚洲不卡免费看| 久久久久久久亚洲中文字幕| 欧美最黄视频在线播放免费| 国产成人午夜福利电影在线观看| 波野结衣二区三区在线| 成人毛片60女人毛片免费| 日韩欧美精品v在线| 国产精品久久久久久精品电影| 国产一区亚洲一区在线观看| 一级二级三级毛片免费看| 岛国在线免费视频观看| 欧美激情在线99| 午夜福利在线观看吧| 亚洲va在线va天堂va国产| 久久久色成人| 久久人人爽人人片av| 美女高潮的动态| 性欧美人与动物交配| 成人无遮挡网站| 日韩精品有码人妻一区| av在线蜜桃| 能在线免费看毛片的网站| 久久6这里有精品| 久久久精品94久久精品| 国产精品日韩av在线免费观看| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 99热这里只有是精品在线观看| 国产成人影院久久av| 午夜福利在线观看吧| 精品欧美国产一区二区三| 最近的中文字幕免费完整| 校园春色视频在线观看| 高清毛片免费看| 美女高潮的动态| 国产在线精品亚洲第一网站| 亚洲无线观看免费| 午夜福利在线观看免费完整高清在 | 亚洲无线观看免费| 国产老妇女一区| 亚洲av二区三区四区| 悠悠久久av| 日韩欧美 国产精品| 成人毛片60女人毛片免费| 欧美日本亚洲视频在线播放| 日本五十路高清| 少妇丰满av| 久久久久久伊人网av| 国产精品,欧美在线| 一级二级三级毛片免费看| 六月丁香七月| 午夜激情欧美在线| 十八禁国产超污无遮挡网站| 日韩国内少妇激情av| 久久久久久九九精品二区国产| 精品久久久久久久久亚洲| 国产成人福利小说| 好男人视频免费观看在线| 毛片一级片免费看久久久久| 乱系列少妇在线播放| 91精品国产九色| 一卡2卡三卡四卡精品乱码亚洲| 简卡轻食公司| 亚洲国产精品久久男人天堂| 可以在线观看毛片的网站| 久久久久国产网址| 边亲边吃奶的免费视频| 免费观看a级毛片全部| 亚洲熟妇中文字幕五十中出| 日日干狠狠操夜夜爽| 精品欧美国产一区二区三| av黄色大香蕉| 伦理电影大哥的女人| 亚洲精品日韩在线中文字幕 | 亚洲av第一区精品v没综合| 最近的中文字幕免费完整| 国产亚洲av片在线观看秒播厂 | 午夜亚洲福利在线播放| 99久国产av精品| 国产探花在线观看一区二区| 成人特级黄色片久久久久久久| 哪个播放器可以免费观看大片| 亚洲内射少妇av| 性插视频无遮挡在线免费观看| 乱系列少妇在线播放| 欧美极品一区二区三区四区| 麻豆av噜噜一区二区三区| 成人一区二区视频在线观看| 国产亚洲精品久久久久久毛片| 村上凉子中文字幕在线| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 国产精品久久久久久久电影| 久久久久久九九精品二区国产| 国产爱豆传媒在线观看| 草草在线视频免费看| 校园人妻丝袜中文字幕| 国产亚洲欧美98| 中文字幕制服av| 亚洲无线观看免费| 欧美又色又爽又黄视频| 国产成年人精品一区二区| 91在线精品国自产拍蜜月| 狂野欧美白嫩少妇大欣赏| 国内精品久久久久精免费| 在线观看66精品国产| 日韩欧美精品v在线| 床上黄色一级片| 天天躁日日操中文字幕| 女人被狂操c到高潮| 亚洲av中文字字幕乱码综合| 国产欧美日韩精品一区二区| 黄色日韩在线| 国产不卡一卡二| 少妇的逼好多水| 日本一二三区视频观看| 国产综合懂色| 国产精品一区二区三区四区免费观看| 国产免费一级a男人的天堂| 日韩欧美精品v在线| 日韩欧美三级三区| 一级毛片电影观看 | 免费观看的影片在线观看| 99热这里只有是精品在线观看| 久久草成人影院| 成人欧美大片| 久久鲁丝午夜福利片| 午夜福利高清视频| 97超视频在线观看视频| 亚洲第一电影网av| 级片在线观看| 黄色日韩在线| 久久精品国产亚洲av香蕉五月| 深爱激情五月婷婷| 乱系列少妇在线播放| 日本在线视频免费播放| 久久久久久久久久成人| 午夜福利高清视频| av在线观看视频网站免费| 99久久中文字幕三级久久日本| 免费大片18禁| 久久精品夜色国产| 国产精品电影一区二区三区| 亚洲第一区二区三区不卡| 色吧在线观看| 亚洲精品影视一区二区三区av| 免费看av在线观看网站| 日本三级黄在线观看| 丝袜喷水一区| 国产精品国产高清国产av| 日本色播在线视频| 国产精品一区二区三区四区久久| av卡一久久| 久久久国产成人免费| 麻豆乱淫一区二区| 国产伦一二天堂av在线观看| 亚洲中文字幕一区二区三区有码在线看| 亚洲18禁久久av| 男人的好看免费观看在线视频| 久久精品国产鲁丝片午夜精品| 人妻制服诱惑在线中文字幕| 亚洲欧美日韩无卡精品| 国产高清有码在线观看视频| a级毛色黄片| 国产成人一区二区在线| 18禁黄网站禁片免费观看直播| av视频在线观看入口| 亚洲国产精品久久男人天堂| 在线观看66精品国产| 免费电影在线观看免费观看| 好男人在线观看高清免费视频| 精品久久久久久久久久免费视频| 国产高潮美女av| 国产成人精品婷婷| 亚洲人与动物交配视频| 国产午夜精品论理片| av免费观看日本| 亚洲七黄色美女视频| 久久久国产成人精品二区| 国产成人午夜福利电影在线观看| 人妻夜夜爽99麻豆av| 久久久久久国产a免费观看| 一进一出抽搐动态| 真实男女啪啪啪动态图| 不卡视频在线观看欧美| 在线播放无遮挡| 人妻夜夜爽99麻豆av| 久久久成人免费电影| 精品久久久久久久末码| 成熟少妇高潮喷水视频| 日韩,欧美,国产一区二区三区 | 免费看美女性在线毛片视频| 欧美色欧美亚洲另类二区| 伦精品一区二区三区| 婷婷色综合大香蕉| 午夜老司机福利剧场| 在线播放国产精品三级| 大香蕉久久网| 美女国产视频在线观看| 久久午夜福利片| 中文在线观看免费www的网站| 在线观看66精品国产| 91麻豆精品激情在线观看国产| 日本一二三区视频观看| 国产精品野战在线观看| 国产av在哪里看| 偷拍熟女少妇极品色| 成熟少妇高潮喷水视频| 婷婷精品国产亚洲av| 国产高清激情床上av| 99久久精品国产国产毛片| 久久人妻av系列| 午夜福利在线观看免费完整高清在 | 国产色爽女视频免费观看| 亚洲人成网站在线播放欧美日韩| 爱豆传媒免费全集在线观看| 国产成人a∨麻豆精品| 最近2019中文字幕mv第一页| 国产精品人妻久久久久久| 国产精品久久久久久亚洲av鲁大| 亚洲精品国产成人久久av| 夜夜夜夜夜久久久久| 九九爱精品视频在线观看| 变态另类成人亚洲欧美熟女| 99热网站在线观看| 好男人在线观看高清免费视频| 麻豆一二三区av精品| av在线观看视频网站免费| av在线播放精品| 中文在线观看免费www的网站| 悠悠久久av| 久久久精品94久久精品| 欧美3d第一页| 欧美日韩综合久久久久久| 国产成人精品婷婷| 蜜桃亚洲精品一区二区三区| 在线观看免费视频日本深夜| 亚洲精品成人久久久久久| 午夜亚洲福利在线播放| 亚洲精品456在线播放app| 变态另类成人亚洲欧美熟女| 日韩欧美一区二区三区在线观看| 午夜福利视频1000在线观看| 欧美丝袜亚洲另类| 少妇的逼水好多| av在线观看视频网站免费| 两个人视频免费观看高清| av在线老鸭窝| 波多野结衣高清无吗| av在线播放精品| 精品久久久久久久久久久久久| 久久精品人妻少妇| 日韩av在线大香蕉| 成人欧美大片| 日韩精品有码人妻一区| 国产精品嫩草影院av在线观看| 久久亚洲国产成人精品v| 国产伦精品一区二区三区视频9| 成人性生交大片免费视频hd| 国产免费男女视频| 26uuu在线亚洲综合色| 欧美成人一区二区免费高清观看| 久久久精品大字幕| 日本色播在线视频| videossex国产| 久久人妻av系列| 中文字幕av成人在线电影| 观看美女的网站| 中出人妻视频一区二区| 欧美不卡视频在线免费观看| 久久婷婷人人爽人人干人人爱| 日韩亚洲欧美综合| av专区在线播放| 最好的美女福利视频网| 色综合站精品国产| 一本久久精品| 99热这里只有是精品在线观看| 色综合色国产| 综合色av麻豆| 国内久久婷婷六月综合欲色啪| 在线观看av片永久免费下载| 岛国在线免费视频观看| 国产在线精品亚洲第一网站| 哪个播放器可以免费观看大片| 少妇猛男粗大的猛烈进出视频 | 少妇熟女欧美另类| 国产白丝娇喘喷水9色精品| 亚洲国产欧美人成| 性欧美人与动物交配| 三级国产精品欧美在线观看| 亚洲七黄色美女视频| 最近最新中文字幕大全电影3|