Control Strategies of Human Interactive Robot Under Uncertain Environments

Actually, the research on human interactive robot (HIR) has been a topic of both science fiction and academic speculation for a long time. The origin of HIR as a discrete issue was stated by 20th century author Isaac Asimov in 1941, in his novel “I, Robot”. He stated the Three Laws of Robotics1 as, “ a. A robot may not injure a human being or, through inaction, allow a human being to come to harm. b. A robot must obey any orders given to it by human beings, except where such orders would conflict with the First Law. c. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law. ” The three laws of robotics determine the idea of safe interaction which constitutes the basic rules of HIR. With the advances of artificial intelligence (AI), the HIR could eventually have more proactive behaviours, planning their motion in complex unknown environments. Nowadays, HIR are artificial agents with capacities of perception and action in the human’s environment. Their use has been tended to be found in the most technologically advanced societies in critical domains as search and rescue, military battle, law enforcement, entertainment, hospital care, etc. These domains of applications imply a closer interaction with human. The concept of closeness is to be taken in its full meaning, HIR and humans not only share the workspace but also share goals in terms of task achievement. The HIR has to adapt itself to human’s way of expressing desires and fulfill its task. Taking lifting up human in elder care for example, the human interactive robot RI-MAN, designed by the RIKEN Bio-Mimetic Control Research Center, communicates with human by listening and speaking, which makes it understand the human will (Onish, Luo et al. 2007). To fulfil the task, it also estimates the attitude of human body in real-time by tactile sense (Mukai, Onishi et al. 2008). This example contains two aspects of HIR, one is to understand the human mind and the other is to accomplish the manipulation. The former is based on AI techniques, like language comprehension, and the latter is relied on force control. On the other hand, human’s environments are much more complex. Thus, the HIR needs perceiving and understanding capacities to build dynamic models of its surroundings. It


Introduction
Actually, the research on human interactive robot (HIR) has been a topic of both science fiction and academic speculation for a long time.The origin of HIR as a discrete issue was stated by 20th century author Isaac Asimov in 1941, in his novel "I, Robot".He stated the Three Laws of Robotics1 as, " a.A robot may not injure a human being or, through inaction, allow a human being to come to harm.b.A robot must obey any orders given to it by human beings, except where such orders would conflict with the First Law.c.A robot must protect its own existence as long as such protection does not conflict with the First or Second Law." The three laws of robotics determine the idea of safe interaction which constitutes the basic rules of HIR.With the advances of artificial intelligence (AI), the HIR could eventually have more proactive behaviours, planning their motion in complex unknown environments.Nowadays, HIR are artificial agents with capacities of perception and action in the human's environment.Their use has been tended to be found in the most technologically advanced societies in critical domains as search and rescue, military battle, law enforcement, entertainment, hospital care, etc.These domains of applications imply a closer interaction with human.The concept of closeness is to be taken in its full meaning, HIR and humans not only share the workspace but also share goals in terms of task achievement.The HIR has to adapt itself to human's way of expressing desires and fulfill its task.Taking lifting up human in elder care for example, the human interactive robot RI-MAN, designed by the RIKEN Bio-Mimetic Control Research Center, communicates with human by listening and speaking, which makes it understand the human will (Onish, Luo et al. 2007).To fulfil the task, it also estimates the attitude of human body in real-time by tactile sense (Mukai, Onishi et al. 2008).This example contains two aspects of HIR, one is to understand the human mind and the other is to accomplish the manipulation.The former is based on AI techniques, like language comprehension, and the latter is relied on force control.On the other hand, human's environments are much more complex.Thus, the HIR needs perceiving and understanding capacities to build dynamic models of its surroundings.It Mobile Robots -Control Architectures, Bio-Interfacing, Navigation, Multi Robot Motion Planning and Operator Training 56 needs to categorize objects, recognize and locate humans and further their emotions.Also in the case of human interactive robot RI-MAN (Mukai, Onishi et al. 2008), if he cannot navigate in the hospital and further more cannot locate the human, the understanding and good manipulation mentioned above make no sense.In our opinion, with the development of human society and robotics technology, HIR research becomes much important (Fig. 1).As HIR interacts with human, sometimes we can model the process of human activity accurately and the robot-human interaction process can be simulated and further done by theoretical approach; sometimes human movement is too complex to model and in that case, the experimental approach is a good way.However, the above thing is to make the HIR capable to navigate in the human environment.The chapter focuses on the issue of designing control strategies of human interactive robot where dealing with uncertainties is a critical issue.Actually, although there are many researches on HIR, these works do not concern too much on the uncertainty.The fact is that there are various uncertainties in the world which comes from robot, human, environment, etc. Developing human interactive robot, in some concept, is dealing with uncertainties.Actually, there are kinds of solving approaches for uncertainties according to circumstances: some uncertain can be assumed as Gaussian noise and based on the property of Gaussian noise, we can do estimation more accurately; some uncertainty coming from model reduction and this kind of uncertainty can be solved by control theory; some uncertainty is neither hard to model, nor difficult to determine the source.In that case, it is a better way to consider it as a black box and recognize it by system identification.In our opinion, the issue of uncertainty has become a major stumbling block for the design of capable human interactive robot.To develop a human interactive robot, it is evitable to deal with uncertainties.Furthermore, managing uncertainty is possibly the most important step towards robust real-world HIR systems (Thrun, Burgard et al. 2005).The basic principle of this chapter is to design control for HIR by dealing with uncertainties.Specifically, there are a number of factors that contribute to the uncertainty of robot, human and environment: Firstly, robot environment is inherently unpredictable.While the degree of uncertainty in well-structured environment such as assembly lines is small, environments such as

58
environment because of the time-consuming computation.When solving the first basic problem, the main topic is to propose an efficient SLAM approach for large scale unknown environment.The second basic problem contains two issues potentially: the interactive object is human dynamics with large degree of freedom; the interactive manipulation is physically done by force.As the first issue, human body has numerous bones and joints and the human model is a very complex one with large degrees of freedom.Exerting force on such a big model is very complex and such dynamic process is impossible to be calculated in real time.Here we use model reduction to decrease the DOF of human model whereas the reduction error (i.e.model uncertainty) comes out.For the second issue, external force is exerted on human.Unlike common object, human has passive moment in human joint and more complicated, sometimes moves at his or her will.Such force character provides a big challenge to control.To solve the basic problem two, the main topic is to propose an adaptive force control approach for HIR when physically contacting with human.More specifically, we take a typical case to research, i.e. how to lift human by HIR in nursing care.The third basic problem focuses on obtaining the human's intention.Actually, the human's intention is very difficult to measure.The model for human mind is also extremely hard to build right now.For us, the human's intention is almost full of unmodeled uncertainty, i.e., we have little knowledge about it.In this case, treating it as a black box is a good way.We stimulate the black box (i.e.human dynamics system) and measure output.By choosing a suitable function to link the stimulus signal and output, we can obtain the intention model experimentally.The main topic here is to design an approach for extracting human's intention.Without loss of generality, in this chapter we consider the problem of estimating the human's intended walking speed.To conclude, in this chapter, we design controls for human interactive robot by dealing with the environment uncertainty, sensor uncertainty, actuator uncertainty, model uncertainty and unmodeled uncertainty.Specifically, the typical problems of HIR we focus on are designing an efficient SLAM approach for large unknown environment; proposing an adaptive force control for lifting human up; estimating human's intended walking speed.The above three typical problems involve the basic problems when designing controls for HIR.The solution of them also provides a general solving framework for HIR, which is of great importance both in research and in application.The preceding researches relating with the above three typical problems are shown as follows.

Background
Extended Information Filter SLAM (EIF-SLAM) estimates the positions of robot and landmarks by updating information matrix and information vector.The total element numbers of information matrix and information vector are where lm n denotes the mapped landmarks.() dof landmark and () dof robot denote landmarks' degree of freedom (DOF) and robot's DOF, respectively.Actually, the dimension of the information matrix increases rapidly with the increase of the landmark number in the environment.For example, if a three-dimensional environment has 100 landmarks, the information matrix is a huge matrix with dimension of 303 by 303.Actually, the computational burden in EIF-SLAM is mainly due to the calculation of the information matrix.
Previous research has proven that information matrix is a naturally sparse matrix (Eustice, Singh et al. 2005).Hence, it is a good way to make use of this feature and change information matrix into a real sparse matrix for computation reduction.Until now, the successful research includes Thrun et al.'s work which solved a relatively large environment (i.e.Victoria Park) for the first time (Thrun, Koller et al. 2002); Eustice et al.'s work which constructed the map of Titanic ship in the dark ocean (Eustice, Singh et al. 2005).Among this work, the sparsification of information matrix was obtained by constructing a proper topological structure of Bayes network, which needs to classify the landmarks in advance.This is a time-consuming work and against the enhancement of efficiency.Actually, the intuitive idea is to set the near-zero elements to zero directly.Thrun tried this way (Thrun, Liu et al. 2004), however, Eustice proved that direct sparsification leads the algorithm diverge (Eustice, Walter et al. 2005), which indicates the improper sparsification process mentioned may cause the algorithm corrupt.If we want to pursue this kind of idea, we have to clarify the condition under which the algorithm maintains stable which constitutes the basis of this section.This section focuses on proposing an efficient stable sparsification SLAM approach for large scale environment.

Sparsing information matrix 2.2.1 Characters of information matrix
In fact, information matrix in EIF-SLAM has special structural characters (Dong, Luo et al. 2009) where the value denotes the numerical value of the element in information matrix located at (, ) row index column index .It is convenient to divide the elements of information matrix into two parts: Part I denotes larger part where the elements have larger values.The elements along the main-diagonal and near the endpoints of sub diagonal line belong to this part.Part II is composed of other elements in the information matrix.Three main characters of information matrix can be stated as a. Information matrix is symmetric along the main diagonal line; b.Elements with huge values distribute in the neighborhood of the main diagonal line and the end points of the sub diagonal line (Part I); c.For the elements of Part II, the value of the element decreases with the distance from the main diagonal line.The reasons for character 1) to 3) can be explained as follows, respectively (Smith, Self et al. 1990;Liu and Thrun 2003;Eustice, Singh et al. 2005;Eustice, Singh et al. 2006).a.Each element in the information matrix denotes the link strength between the corresponding landmark and robot.As we all know that the link strength has symmetry, i.e. if the link strength between A and B is  , then the link strength both from A to B and from B to A are  .Therefore, information matrix is a symmetric matrix; b. From the viewpoint of Bayes network, the link strength shows the correlation.In SLAM problem, the largest correlation is the relation of robot itself and landmark itself.The correlation between the robot and landmark becomes weak with time.If the robot is currently observing a landmark, the correlation between the robot and the landmark is strong.The elements in diagonal line denote the link strength of robot versus itself, or landmark versus itself while the elements near the endpoint of information matrix show the link strength of robot versus current observing landmark.Hence, the elements with large value centralize in Part II; c.The correlation becomes weak when time passes after observing it.When the robot observes the landmark again, the correlation increases again.The most important characteristic of information matrix is that most elements of it are nearly zero.In other words, an information matrix is a nearly sparse matrix.The information matrix sparsification by different threshold (1 6 (1) Fig. 2. Information matrix sparsification.
After computing the sparse ratios of the information matrices in Fig. 2, we can get the graph that illustrates how the sparse ratio changes with the threshold (Fig. 3) where the x-axis denotes the threshold and the y-axis denotes the sparse ratio.By using a curve fitting method, it shows that the distribution of elements in the information matrix satisfies a linear distribution in logarithmic coordinates.The sparse ratio increases with the threshold.However, sparsification error arises from the sparsification process.Furthermore, inappropriate sparsification may lead EIF-SLAM to diverge.Therefore, the key issue here is to find a condition under which the estimation results converge all the time.From the illustrations above, the information matrix in EIF-SLAM algorithm is naturally sparse.Hence, the sparsification structure of the information matrix is very suitable to be sparsed.Moreover, because the computations are mainly on the calculation relating with information estimation variables, high efficiency is predicted by sparsing the information matrix.The following sparsification approach utilizes the characters of the information matrix to decrease the computational burden of EIF-SLAM.From the illustrations above, we obtain that the information matrix in EIF-SLAM algorithm is naturally sparse.Hence, the sparsification structure of the information matrix is very suitable to be sparsed.Moreover, because the computations are mainly on the calculation relating with information estimation variables, high efficiency is predicted by sparsing the information matrix.The following sparsification approach utilizes the characters of the information matrix to decrease the computational burden of EIF-SLAM.

Sparsification approach
The mean vector  and the covariance matrix  of the environment state vector  are written in the form Assume estimation error is a Gaussian noise, we have

By taking the relation between time estimation variables and information estimation variables
Here we propose Theorem 1 to clarify the conditions for sparsification for guaranteeing the convergence.In other words, from the viewpoint of mathematics, Theorem 1 gives the consistency condition; from the viewpoint of geometric, the mean remains in the range of the corresponding covariance under consistency condition.

Theorem 1
Assume the new information matrix after sparsification is written in the form of E  , where E is the sparsification error matrix.If where  ,  ,  and E are defined as , 00 where the element positions of  are the transport element position of E .That is, the portioned matrices E ,  ,  and  are of order of 12 mm  , 21 mm  , 2 nm  and 1 mn  , respectively.The proof of Theorem 1 is shown in (Dong, Tang et al. 2010).Consider Equation ( 6) with respect to Equation (5), it is shown that after sparsification, each estimated mean still inside the scope of corresponding covariance.That is to say, the estimated position of the robot or landmark is inside the estimation range, which indicates the estimation is reliable.In addition, the proof of Theorem 1 also gives a quite efficient way to get covariance matrix from information matrix without directly computing the inversion.According to Theorem 1, we can eliminate the information matrix under the condition of stabilization.In practice, loop-closure is used to compensate the estimation error which comes from sparsification.After sparsification, another theorem was proposed to evaluate sparsification error by deriving the upper bound of relative error ratios, including mean vector error ratio and covariance matrix error ratio (Dong, Tang et al. 2010).

Large complex environment simulation
The proposed sparsification SLAM approach was simulated in a large scale environment (about 400 landmarks).As the landmarks are very dense, the environment is a complicated one.Such environment can further verify the effectiveness of the proposed approach.In the simulation, the robot applied is a two-wheel robot with three degrees of freedom and the sensor arranged on the robot is range-bearing type sensor (such as laser, camera, etc.).The robot moves counterclockwise for two laps while utilizing sparsification approach to process SLAM.

Efficiency analysis
The main computational burden of EIF-SLAM is primarily from the three steps: motion update, features addition and observation update.The comparison of computational time between EIF-SLAM (denoted as stars) and sparsification method (denoted as crosses) is shown in Fig. 4. It shows that the sparsification approach is able to solve SLAM in the environment with 400 landmarks.Furthermore, according to the curve trend of computational time, we can predict that our sparsification method has an efficiency advantage for large environment.

Accuracy analysis
The comparison of error and covariance is shown in Fig. 5 where dash line and solid line denote estimation error and self-covariance of robot's position, respectively.Horizontal axis shows the time.Vertical axis denotes the estimation error and covariance where x and y denote the moving directions, respectively.It shows that estimation error and covariance decrease sharply at time 250, which indicates that there is a loop-closure at this moment.In fact, during the first lap, the error and covariance by the sparsification method nearly have the same magnitude, (in Fig. 5 (b), the solid line nearly coincides with the dash line), which indicates that the sparsification method eliminates considerable number of elements under the condition of consistency.In the second lap, the estimation error by EIF-SLAM (Fig. 5 (a)) and the sparsification approach (Fig. 5 (b)) seems the same.Thus, it can be concluded that by using loop-closure properly, the sparsification approach can obtain satisfying accuracy with high efficiency.

Summary
In this section, a sparsification EIF-SLAM approach was proposed to enable human interactive robot to navigate in the large unknown environment.According to the normal structural features of information matrix, we eliminate many of the near-zero elements in the information matrix.Such sparsification process is under the condition of consistency.
Hence the sparsification approach stays stable.In addition, the upper bound of estimation error is also given for evaluation.The large complex environment simulation indicates that the sparsification approach has the advantage of high efficiency and accuracy.The outdoor car-park experiment shows the ability to realize consistent estimation by the sparsification approach (Dong, Luo et al. 2009).Compared to the previous researches on efficient navigation of human interactive robot, the sparsification approach proposed in this section gives a direct but effective way to obtain efficiency.Concerning parameters, we can compromise the efficiency and accuracy under the condition of stabilization.It is also noted that the consistency condition derived in this section also has great meaning for the future research on sparsification.This section dealt with the environment uncertainty, sensor uncertainty, actuator uncertainty, and assumed them as Gaussian noises.The proposed sparsification SLAM approach eliminated these uncertainties and realized high efficiency of SLAM estimation.

Background
It has been thought that the human body is composed of 206 bones and numerous human joints connecting adjacent bones.Based on the physiological structure of human joints, the human joints can be mainly divided into three types as the hinge (1 DOF), the pivot (1 DOF) and the ball socket (3 DOF).In dynamic equations, each DOF is expressed as one differential equation, which indicates that the overall set of equations of human body dynamics is a very huge one.The force interaction with such a big model needs considerable computation for human interactive robot.As illustrated in Section 1, in the process of lifting human, the human body has to be considered as a free-floating multi-link rigid object with passive moments.Compared with one-end-fixed object, like manipulator, free-floating object is much more complex.The force acting on any part of the object would affect the attitude of the entire object.Moreover, although the multi-link object has been researched a lot, the one with large DOF and large redundancy has not fully been considered yet.Another problem is calculation; any computation on such a huge model needs much computational time whereas the safe lifting requires real-time computation.In consideration of the difficulties mentioned above, our basic idea for lifting human comes from the daily experience.When we human lift a person, we do not care too much about the detailed dynamics, like the change of ankle angle, hand position and so on.What we do care about are the head position, the vertical deflection of upper limb and the hip angle.Here we call them "states of interest".From the viewpoint of system theorem, we treat the human body as a large redundant system whose dimensions are reduced by diverting the effects of other "joints of noninterest" to the "interested ones".The resulting body model is a reduced one with less DOF but unfortunately, has huge uncertainties (i.e.model uncertainty).Here we focus on proper model reduction and methods for dealing with the generating model uncertainty, which leads to a force control for lifting human.

Force interaction modeling
If we consider human body as a rigid multi-link object, each bone corresponds to a link and each human joint corresponds to a joint connecting adjacent links.Moreover, human joints have passive torques corresponding to the constriction forces and moments developed by ligaments, joint capsules and other soft tissues.Hence, we write the dynamics of human in the from where Generalized body states, including the position of head and the angles of all the joints.() nn Hq  Inertia matrix, defined as a positive semi-definite symmetric matrix, containing information on the body's instantaneous mass distribution.(,) nn Cqq   Centripetal and coriolis torques. (,) Cqq  term contains products of angular speeds.When the degree of freedom is rotational, the terms of (,)  Cqq  represent the moments of centrifugal forces.Here the HIR's task is to lift human, i.e. to control the position and posture of the body to the desired states by external forces.Adding the force exertion in Equation ( 9), leads to: where 1 rob n   The torques acted by the robot arms which is controllable.Actually, the human interactive robot considered in this section has two manipulation arms, like RI-MAN (Mukai, Onishi et al. 2008) J and 2 J are the Jacobian matrices of the human body model.

Human dynamics reduction
The basic idea in this section is to reduce the human body model into a small one with less degree of freedom, which includes the following three steps: a. Choose "states of interest".These states include the fundamental performance indexes of the task.In other words, based on these states, we can determine whether the task is complete or not and furthermore, evaluate the performance is good or not.
Let us define as the "states of interest" and where the dimensions of sub block matrices of 11 Considering that the dynamic model is time-varying, after arranging Equation ( 16), we obtain Ht q Ct q Gt H t q Ct q tt By defining the inertia matrix, centripetal matrix, gravitational matrix and torque vector of the reduced human dynamics as we obtain the general mechanical form of the reduced human dynamics where the subscript s denotes the small system.We consider the influences from "uninterested human joints" (in this case from state 2 q ) as perturbations.We change the attitude of reduced human model adaptively by estimating the parameters of s H , s C and s G in real time.The detailed estimation meanings are as follows.


Estimating s H and s C ---make the system adaptively adjust itself to various people with different weights.


Estimating s G ---eliminate the perturbations from other "uninterested joints".Considering the basic principle above, the approach to be proposed in this section is to identify and control the reduced human dynamics at the same time.Assuming that the human parameters are totally unknown in advance, for the safety in the nursing activity, the identification process needs to be performed in real time.On the other hand, the weights and heights etc. of the human bodies are different between individuals.Hence, the strategy also has to be able to tolerate these individual difference.

Attitude control and human parameter identification
First of all, we assume that we do not have any priori knowledge before lifting human, i.e. the initial value of s H , s C , s G are set to be zero matrices and zero vectors.The benefit of such assumption is that the proposed trajectory is much more robust and can be adaptive to various people with different heights and weights.Whereas, such assumption also leads to a problem, i.e. it generates much model uncertainties in the dynamics.To solve the above problem, we use robust controller to change the human attitude to the desired states.Moreover, online human parameter identification is also done so as to estimate the human body in real time.For the convenience of mathematical derivation, we define the actual human parameter vector Then the estimation error vector can be defined as In fact, not any combination of H , C and G corresponds to a physical system.Therefore, the first step is to prove that the reduced human model represents a real physical system.It is easy to verify that by proving that 2 ss HC   is a skew-symmetric matrix.In other words, the reduced human model satisfies energy conservation where the detailed derivation is in (Dong, Luo et al. 2010).We propose a theorem to change the "states of interest" of human body.It is composed of a human attitude control law and a human parameter identification law.In fact, the control process and identification process run at the same time.In the proof of Theorem 2, the global stability is shown by proving that the derivative of the constructed Lyapunov function candidate is less than zero.

Theorem 2
Consider a time-varying system with m-order without any pre-knowledge about s H , s C and s G .The vector , sd q means the desired states.
Define the sliding term s as ,, () ' () where  is a positive diagonal matrix.Define the reference velocity , sr q  and reference acceleration ,   is a signal function.
The proof of Theorem 2 is shown in (Dong, Luo et al. 2010).It is noted that the signals required in the control law and identification law are s , , sr q  , , sr q  (Equation (32) and Equation ( 33)).According to the definitions of s , , sr q  , , sr q  (Equation (30) and Equation ( 31)), the basic signals required are s q , s q  , s q  .Actually, the "states of interest" s q , s q  represent the basic attitude element of human, i.e. position, angle, linear velocity, angular velocity.ss qf q w    , i.e., the acceleration signal can be obtained from velocity signal.

Simulation 3.5.1 Simulation results
The simulation was implemented by coordination of three software packages, including AUTOLEV, MATLAB and VORTEX.The detailed cooperation relations are explained as follows.AUTOLEV is used to construct the human model (Kane and Levinson 1985).We choose MATLAB to process the main computation of solving ordinary differential equations; Although VORTEX is able to do physical simulation, the programming grammar is a bit complex.Here, we just use its stereoscopic presentation function to make animations.
By applying 1 F and 2 F , human attitude control is achieved.The energy, position and angle changes are shown in Fig. 6.It is seen that it takes about 1 second to accomplish the process of attitude change.There is a peak of kinematics energy at the time of about 0.2 second which means at that time, the attitude changes very quickly (Fig. 6 (a)).One reason is that we assume no pre-knowledge of the human body at the beginning of the simulation.The head position changed to the desired (0.2m, 0.8m, 0.01m) at about 1 second (Fig. 6 (b)).Compared with other joints rotating in x or y direction, the joints rotating in z direction change significantly.Thus, the angle changes of these joints affect the head position in x direction greater.In (Dong, Luo et al. 2010), it is shown that 2 ss HC   is a skew-symmetric matrix which indicates that parts of the states (or their linear combination) can be controlled as a whole.In the simulation, we constructed a new state which is the angle sum of head, chest, mid-trunk, and lower-trunk.The angle drift off the horizontal line of the new state changes to the desired -0.7854 rad (i.e.-45 degrees) at about 1 second as shown in Fig. 6 (c).The angle between lower-trunk and upper-leg changes to the desired 1.5708 rad (i.e. 90 degrees) at about 1 second (Fig. 6 (d  www.intechopen.com

Animation
We imported the computed motion data into the VORTEX human model to make animation.
Here the human model is the visible skeleton which was built in (Dong, Luo et al. 2010).
Compared with the human model for simulation (16 links and 35 DOF), the human model for animation is a redundant one (54 links and 61 DOF).Here we just make the corresponding human joints rotate and maintain the redundant human joints fixed.As assumed in the simulation that the force contact type is contingence, we choose two cylindrical objects to represent robot arms.
The position of the head, angle drift off the horizontal line of lower-trunk, angle of lower-trunk and upper-leg are chosen as the "states of interest" (Fig. 7

Summary
In this section, an attitude control approach was proposed to lift human without regard to the individual differences, such as height, weight, and so on.We used robust adaptive control to eliminate the effects from the "uninterested joints" and identify the human parameters in real time.In addition, the convergence analysis, including tracking time and static tracking error, was also given.The approach was simulated by lifting normal human body with two robot arms, which verifies the efficiency and effectiveness of the proposed strategy.
Compared with the previous researches, there are two novelties in the proposed approach.
First is that it is not necessary to measure human, like height and weight, in advance because the approach can automatically identify the human parameter online.Second is that the attitude control law ensures the accuracy.Moreover, the robust controller which we proposed also has the ability to tolerate the model uncertainty of human.Actually, lifting human is a typical problem when interacting with human by force.As human is such a big redundant model, a good solution is to reduce it and design controls.
Here, there are two sources of model uncertainties: one is from the unmodeled dynamics which we cannot measure; the other is the generating model uncertainty which comes from model reduction.The effectiveness of the adaptive force control shows that by designing robust controller and online estimator, the two uncertainties mentioned can be solved.

Background
Human walking servo robot is a robot which adaptively adjusts its velocity to human's walking speed.In this section, we consider controlled treadmill as human walking servo robot.In the previous human walking servo robot applications, the subjects had to passively follow the given speed (Mcdermott, Ades et al. 2009;Powell, Stevens et al. 2009).However, in many application cases, the human walking servo robot control strategy motivated by human will is very desirable.However, human will is very complex and hard to be measured.Even if certain models are obtained by brain signal, like brain computer interface (BCI), the models are limited because of great unmodeled uncertainty.Such a huge unmodeled uncertainty is very difficult to handle.Although the adaptive control with identification can tolerate some degree of unmodeled dynamics, it requires the unmodeled uncertainty is in some degree of limit.Once the model uncertainty is beyond the threshold, the control strategy corrupts.Considering this, we have to find another way which does not rely on model much.One constructive perspective is to consider the human process as a black box with multi-input and multi-output.Based on the experimental approach, we can evaluate the human process by the input stimulation (from 1 u to m u ) and output human response (from 1 x to n x ).The objective is to find a relation i f to satisfy 1 (,, ) where i f can be linear or nonlinear function.By determing i f (1 ) in  , we can evaluate the human process.This section focuses on designing control for human walking servo robot by human's walking intention.As we consider the human process as a multi-input-muilti-output (MIMO) black box, the main topic is to identify the MIMO human process.

Human walking intention extraction 4.2.1 Characteristic index
The coordinate definition is shown in Fig. 8.Here based on the real product of force plate, we assume that the true origin of the strain gauge force plate is not at the geometric center of the plate surface.There is a pad on the force plate.After a series of calibrations of the true origin, the true origin ' O is at (0,0, ) h .According to the coordinate, the reaction force of ground to feet was simulated where the human model is built by OPENSIM in Section 4.4.It is shown that in one dynamic circle, the force z F is a bell-shape signal; y F is a sine-shape signal (Fig. 9 (a)).We can explain the reason of such curve shape as follows.When foot gets in touch with the surface of force plate, z F increases very rapidly.At the same time, foot has to make a break to adjust its speed to the velocity of the force plate by fiction.After break process, foot applies a force with inverse direction to drive the leg to take a step, i.e. make a preparation for higher speed of leg in the next moment.It is noted that, compared with the previous break process, y F changes its direction at this time.Until now, the foot is on the force plate.Hence, z F maintains large (for a normal person with 70 kg weight, z F is about 700 N).Finally, the subject alternates the other foot to support body and z F decreases rapidly.Considering the fact that when z F is large enough, one foot is firmly on the human walking servo robot.We define a ratio index where  is a threshold.In normal cases,  is set as max{ } 80% z F  .According to the curve shapes of y F and z F , the curve shape of , y z R is a composite signal of connection of sineshape signals and zero signals.
To prove the above simulation result and corresponding analysis, we use Bertec treadmill TM07-B to complete a verification experiment.It is noted that under the treadmill, there are two force plates individually measuring the interaction force and moment.In the experiment, the reaction data are measured when the treadmill velocity varies from 1.0 m/s to 1.6 m/s.Without loss of generality, the zero signals are ignored in analysis, which leads to that

Walking intention modeling
where w  is a parameter vector.The random variable w  is independent of intend V and on average it is equal to zero, i.e. ()0 w E   .We want to find w f that fits the measurement data best and we define the loss function to measure the quality of the fit as In this section, we choose the other three model candidates: quadratic, cubic, 4th degree polynomial.In the experiment, we obtain , y z R ten times when the treadmill velocity varies from 0.1 m/s to 1.0 m/s.The observation samples are shown in Table 1.Good performance can be obtained by using the above filter for many times.Fig. 11 shows a dynamic walking process with acceleration motion, deceleration motion and uniform motion (constant speed motion) in the time interval [28,34] , [55,65] , [15,24] , respectively.

Summary
This section proposed a method for extracting human's walking intention and based on it, a speed control was proposed for adaptively driving human walking servo robot by human's will.The ground reaction force is considered as the indicator of the human's intended walking speed.By analyzing the walking simulation, we found a characteristic index which has significant relevance with the intended walking speed.After processing least-squares regression, four kinds of candidate models were obtained.For simplicity, we chose the linear model.The extracted human intention was used to control the human walking servo robot.The control performance shows the effectiveness of the proposed method for extracting human walking intention.
Compared with previous researches, there are two novelties to be illustrated.First is that the proposed method for human walking intention extraction simply uses the ground reaction forces to do estimation.The method is so easy-to-use that it even can be used on single-chip.
The second is that in previous human walking servo robot control, subject had to follow the speed of human walking servo robot passively.While the proposed speed control can adaptively adjust the velocity of human walking servo robot to the subject's, which shows great potential both in research and real-world applications.
Actually, when HIR interacts with human by human's intention, it is extremely hard to model or identify the unmodeled uncertainty because we know so little about human's intention.In this case, it is a good way to analyze it just from the relation between inputs and outputs.The proposed speed control based on human's walking intention shows that by identifying the relation mentioned, we can deal with the unmodeled uncertainty from human's intention.

Conclusion
In this chapter, we designed kinds of control strategies of HIR and solved three typical problems for HIR.A sparse extended information filter SLAM approach was proposed for HIR's navigation in a large unknown environment, which ensures HIR can fulfil its task in human surrounding environments.An adaptive force control was designed for HIR To lift human solving the physical interaction under an uncertainty of human dynamics of dealing with the complex human dynamics during interaction.A speed control by human's intended walking speed was designed, which gives a solution to extracting human's intention.Compared with other HIR approach, this chapter takes fully into account of uncertainties HIR encounters, including environment uncertainty, sensor uncertainty, actuator uncertainty, model uncertainty and unmodeled uncertainty.The contribution of this chapter is not only to give specific HIR controls for specific cases, but also to provide a good solving framework for HIR problem.


Gq  Gravitational torques.Because () Gq changes as the posture configuration of the human body model, its terms are functions of the generalized states. 1 pass n Passive joint torque, including torques and moments arising from muscular activations and passive elastic structures surrounding the human joints.
www.intechopen.comMobile Robots -Control Architectures, Bio-Interfacing, Navigation, Multi Robot Motion Planning and Operator Training 68 angle drift off the horizontal line of lower-trunk (denoted as 1,x
Fig. 7. "States of interest" in lifting human and Snapshots of the lifting human process (Dong, Luo et al. 2010).
connection of various sine-shape signals with different magnitude.It can be inferred that , y z R is a characteristic index for the walking intention: when the subject intends to speed up, , yz R  becomes large, i.e. both the magnitudes of peak value and valley value become large.After acceleration, , y z R returns to a new equilibrium state (Fig. 9 (b)).In other words, the envelope curve of , y z R determines the intended walking speed.The deceleration case can be explained in a similar way.
Mobile Robots -Control Architectures, Bio-Interfacing, Navigation, Multi Robot Motion Planning and Operator Training . Force is applied to the fixed point at back and knees, shown as 1 13) as the "uninterested states" consisting of the other states.The overall state can be written as )).

Table 1 .
Observation samples.For the purpose of evaluating the candidate models, we define residual norm as the sum of square of deviations

Table 2 .
Regression results.Because the residual norms of the four types do not have big difference, any model of the four is able to describe the human's intended walking speed.Therefore, for simplicity we choose the linear model as follows.