Novel Control Algorithm for the Foot Placement of a Walking Bipedal Robot

A novel control algorithm for the foot placement of walking bipedal robots is proposed which can output the optimal step time and step location to obtain a desired walking gait from every feasible robot state. The step time and step location are determined by approximating the robot dynamics with the 3D linear inverted pendulum model and analytically solving the constraint equations. Intensive simulation studies are conducted to check the validity of the theoretical results. The results of this study show that the proposed control algorithm can get the system to a desired gait cycle from every feasible state within a finite number of steps.


Introduction
In considering the control methods for walking bipedal robots, two fundamentally different approaches can be observed: the technical approach and the biological approach.Technical control approaches are understood as the class of methods that are mainly based on insights from industrial robotics and mechanical engineering.They draw upon sound mathematical concepts such as dynamics, linear and nonlinear control theory, and established joint constructions and materials from industrial robots.Biologically inspired approaches try to transfer results from human motion analysis, biomechanics and neuro-scientific research into technical systems.It will cause a difficult control problem when generating a robust walking bipedal robot gait -the robot systemʹs central velocity can be controlled through the centre of the pressure location, but it will be affected by the support polygon imposed upon the constraint in this location.In particular, when the single support polygon is small, it results in a robot system that resembles a highly unstable inverted pendulum.For the walking bipedal robot to remain stable, an appropriate new control algorithm has to be developed [1][2][3].
Most of the control systems for walking bipedal robots are based upon a biological approach.Recently, biologically inspired studies of approaches have found that the walking bipedal robotʹs stable step locations, centre of mass position and velocity information, can be predicted based upon simple inverted pendulum dynamics.Alghooneh Mansoor [4] and Johannes Englsberger [5] have built upon the capture point concept and exploited the simple form of the dynamical equations

R E T R A C T E D
of the linear inverted pendulum model when formulated in terms of the centre of mass and the capture point.
Using the capture point to generate the walking gait of a bipedal robot is less straightforward, since this obviously requires a non-zero forward speed.This "constant offset control" approach is proposed by Dehghani Reza [6], and it seems logical since stepping ahead or behind the indicated point will, in general, cause the legged system to decelerate or accelerate respectively [7][8][9].
However, the constant offset control approach for a bipedal robot has three significant shortcomings, as follows: (1) the desired gait will only be reached in an infinite time.This can be problematic because the system may not have recovered from a perturbation before a new perturbation occurs.(2) The control algorithm of foot placement will only work for a limited set of all feasible system states.This limits the ability to maintain stability in the presence of large perturbations.(3) The control algorithm for foot placement assumes a constant step time.This reduces the possible step strategies and, therefore, limits the control performance [10][11][12][13][14].
In this paper, in order to overcome these shortcomings, we introduce a novel control algorithm for dynamic foot placement that gives a direct relation between the desired system state and the required step location and step time to reach that state within a finite number of steps.The control algorithm is derived using the linear inverted pendulum model.The simple nature of this model results in simple control rules that allow for real-time implementation on a real walking bipedal robot.We will demonstrate the effectiveness of our proposed foot placement control algorithm and show that the control algorithm can be used to obtain a desired cyclic walking gait from any feasible system state.

B 2. Modelling of a Walking Bipedal Robot
In order to obtain the novel control algorithm, we modelled the dynamics of a walking bipedal robot, as shown in Figure 1: the 3D linear inverted pendulum model is comprised of the point foot, a point mass with m at the coordinate position (x, y,z) with respect to the local reference frame of x y z (e ,e ,e ) and a telescoping leg that keeps the point mass at a constant height 0 z z  in contact with the ground (the leg has a point foot, which cannot apply torques on the ground).The state of the model is uniquely described by the set of state variables T q (x, y,x,y)    .The gravitational acceleration vector is T g (0,0, g)   .
The state of the walking bipedal robot model can be described by the state variables of  The continuous dynamics describe the evolution of the initial state 0,i q to state 0,f q over time 0 t .The transition dynamics describe the instantaneous change of state 1,f q to state 1,i q due to step 0 S .This sequence is repeated for consecutive stance phases and steps.Any state that lies will make the model come to a stop in the x -direction, t lim x 0    .The model can only come to a stop for initial states n ,i q that lie inside the 1step viable-capture basin due to the imposed stepping constraints.With , for which g is the gravitational constant and 0 z is the height of the point mass.Solving Eq.( 1) results in a closed form solution for the stance phase dynamics: where n t is the duration of the stance phase and subscript i and f refer to the initial and final model state respectively of a stance phase, which is indexed with subscript n , with 0 n N  .
The transition dynamics are instantaneous and describe the changes in foot positions that occur when taking a step.It is assumed that a step has no instantaneous effect on the velocity of the point mass.The instantaneous state change is given by: describes the step size in the x and y directions.
To model the limitations of a real robot, we impose two stepping constraints on this model.First, we model actuator saturation on a real walking bipedal robot by introducing a lower limit on the time between foot location changes: Second, we model the limited kinematic workspace of a real robot by introducing an upper limit on step length, i.e., the distance between subsequent point foot locations: Due to these constraints, the model can only operate in a subset of the state space.The subset consists of all states for which the model has the ability to come to a stop.For states outside of this subset, the model will accelerate without the possibility of deceleration.The subset is spanned by model states n,i q that lie within the 1-step viable-capture basin, as given by: In Figure 2, the basin boundary is indicated by the dashed lines labelled d  .

Walking Gait
Our proposed foot placement algorithm has the ability to bring the model to any desired feasible state.In this paper, we select a more practical application of the controller by bringing the model to a state that is part of a desired walking gait.This requires that the model should not only arrive at the desired state, but that it should also be capable of maintaining the desired gait cycle.This requires a passed minimum step time when the gait cycle requires another step to be taken.
The desired gait cycle is set to a humanlike two-step cyclic gait with an alternating left and right step.If the model has obtained the desired gait cycle, then: with a corresponding step time # n 2 n t t t    .The step size of the desired gait cycle is: with x and y being the forward and lateral directions of movement respectively.
The control objective is a desired state within the gait cycle.For this paper, we select the state at the end of the stance phase as a reference state within the gait cycle: For all the figures throughout this paper, we use an example desired gait cycle, which is depicted in Figure 3.For this gait cycle, # t 1.5 .These three parameters uniquely describe the gait cycle and set the average forward velocity to # # x S / t 0.5  .We set the model parameters g , 0 z , max l and max t to the unit magnitude.

Control Algorithm
To let the model reach the desired state # f q within the gait cycle, the control algorithm should adjust the step time n t as well as the step location n,x S and n,y S over a finite number of N steps, while satisfying the stepping constraints.This results in the following formal description of the control algorithm: f(q ,t , t ,S , S ) q t t ,for n 0 N S l ,for n 0 N 1

R E T R A C T E D
in which function f gives the state of the system after the N th step.This function can be derived from Eq.( 2) and Eq.( 3), and is given by: S cosh( (T(N) T(i))) f(q ,t , t ,S , S ) with T(i) t p sinh( T(N)) p cosh( T(N)) S sinh( (T(N) T(i))) Finding the control algorithm that satisfies Eq.( 10) for a minimal number of steps N can be done using numerical optimization methods.However, the nondeterministic nature of these optimization methods makes it unattractive for real-time implementation on a walking bipedal robot.
Therefore, we adopt a novel approach to synthesize a foot placement.We start by checking if there is a solution to the control problem for N 0  .If there is no solution, we repeat the procedure after an incremental increase of N .For each N , the problem that needs to be solved is different, since N determines the number of control parameters and constraints, as in Table 1.
The dynamic control algorithm for foot placement outputs the next step location and step time, expressed in terms of the control parameters n,x S , n,y S and n t .We will derive expressions for these control parameters for each initial model state.The foot placement control algorithm consists of four N -step strategies with N 0,1,2,(N 2)   .Each N -step strategy will result in an N 1  -step strategy up to a 0-step strategy, as demonstrated for an example 3-step strategy in Figure 4.

0-step algorithm
If the initial state is already on the gait cycle or will arrive at the gait cycle without taking a step, then the following should hold: The control parameters are then given by # 0,x x S S  , # 0,y y S S  and: which is found by solving Eq.( 12) for t , in which 0,f q is given by Eq. ( 2).

1-step algorithm
The 1-step strategy is fully constrained (Table 1), which means that a unique solution exists for the control parameters x,0 S , y,0 S , 0 t and 1 t .To derive an expression for these parameters, it is important to acknowledge the following: since a step has no instantaneous effect on the model velocities, we know that the velocities in the x-and y-directions at the moment of the step should match the velocities of the desired gait cycle at some instant of the gait.To further clarify this fact, let us consider the model velocities p  shown in the phase space diagram depicted in Figure 4. Here, we see a 1-step strategy that is part of a 3-step strategy.We see that, in going from a 1-step strategy to a 0-step strategy, the phase space trajectory of the model velocities intersects the phase space trajectory of the possible gait-cycle velocities.At the instant that the velocities match, a step is taken and the model evolves to the state # f q  .Consequently, for a 1-step strategy to be feasible, an intersection 0,f should occur.This means that the following expression, based on Eq.( 2), should hold, This expression can be solved to find the stance phase durations 0 t and 1 t .We omit the expressions for 0 t and 1 t because of their length.
With the known stance phase durations t0 and t1, we can also derive the expression for the step size 0 S from Eq.( 2) and Eq.( 3): The last part of the 1-step strategy is to check whether the step durations found, 0 t and 1 t , and the step size 0 S , are within the stepping constraints, Eq. ( 4) and Eq. ( 5).If this is not the case, the 2-step strategy should be considered.

2-step algorithm
A 2-step strategy is under-constrained (Table 1), meaning that multiple solutions can exist for the available control parameters 0 S and 1 S and the stance durations 0 t , 1 t and 2 t .The constraint equation (11) in case of a 2-step strategy becomes: # 0,i 0 1 2 0 1 f f(q ,t ,t ,t ,S ,S ) q  (16) Solving Eq.( 16) for 0 S , 1 S results in: We can set three of the seven free control parameters (Table 1) and thereby derive a value for all of the parameters.An example option would be to set all of the stance durations 0 t , 1 t and 2 t to min t , which we have done for the visualization of the control performance in the above Section.Though this approach might result in a stepping strategy, it does not use the absolute minimum number of steps for certain model states.

(N>2)-step algorithm
For all those model states that cannot evolve into the gait cycle using N -step strategies for N 2  , we adopt a single step strategy.Though more suitable stepping strategies for N 2  can be derived, these were found to be complex and result in a negligible increase in stepping performance.
The model state should quickly be brought to a region in a state space for which N -step strategies for N 2  do apply; consequently, n t is set to min t .A corresponding effective step location is a point at an offset from the instantaneous capture point and was introduced as part of the constant offset controller: If this point is not within reach due to stepping constraints, we step on the instantaneous capture point or else do so maximally towards this point if it is also out of reach.

Simulation and Discussion
The projects of the Natural Science Foundation of Jiangsu Province (BK2011226) and the Specialized Research Fund for the Doctoral Program of Higher Education of China (20100095120006) sponsor these simulations.The goal of this study is to synthesize a novel foot placement control algorithm that can predict an appropriate step location and step time that will bring a bipedal robot to any desired gait cycle in a minimal number of steps.To arrive at comprehensible control strategies, we used a simple gait model and adopted an incremental approach to solve the control problem.The use of this approach and the model has its benefits, but it also introduces some limitations.Incrementally solving the control problem for different N , which allowed us to derive closed-form expressions for the control parameters.These expressions give direct insight in the relation between the walking objectives (in this case, a desired walking gait in a minimal number of steps) and the foot placement control (step time and step location) that is required to reach this objective.We demonstrated that a total of only four step strategies were required to effectively bring the model to the desired gait cycle from any feasible model state.The Figure 6 shown that a change in walking direction is achieved by rotating the local coordinate frame in the x y  plane over an angle.The depicted 90 turn is, in this case, achieved by rotating the coordinate frame over 45 after the first step and again 45 after the second step.
A limitation of the approach is that we optimized our control algorithm to a minimum number of steps, though other optimization criteria might also be relevant.These could be, for example, the minimization of the time required to reach the state or the minimization of the amount of mechanical work required by the actuators.In

R E T R A C T E D
the near future, we plan to study how the foot placement strategies are influenced by the selected optimization criterion or by combinations of criteria.
Although the 3D-model enabled us to derive these simple stepping strategies, it is a very simplistic representation of a real robot.By modelling only the centre of mass dynamics and not taking impact dynamics into account, we neglect many dynamic aspects of true dynamic walking.However, we believe it could be successfully applied to control a complex biped robot.No attempt should be made to forcefully mimic the 3D-model dynamics on the complex biped, as is typically done for 'Zero Moment Point'-based control approaches.The evolution of the complex biped will inevitably deviate from the evolution that is predicted by the simple model.The simple nature of the dynamic foot placement controller allows for the real-time execution on the robot, which in turn allows the controller to continuously adapt to the current state of the robot.The output of the dynamic foot placement controller can be used as an approximation for an adequate step time and step location for the real robot.Biomechanical studies have shown that such simple foot placement approximations, based upon simple inverted pendulum models, display a good correlation with human foot placement strategies.

Conclusions
In this paper, we presented a novel control algorithm for the foot placement of walking bipedal robots which can output the optimal step time and step location to obtain a desired walking gait from every feasible robot state.The step time and step location are determined by approximating the robot dynamics with the 3D linear inverted pendulum model and analytically solving the constraint equations.The simulation results of this study show that the proposed control algorithm can get the system to a desired gait cycle from every feasible state within a finite number of steps.
T q (p,p)   , with T p (x,y)  ，and the x and y representing the position of the point mass with respect to a local reference frame of x y z (e ,e ,e ) .The dynamics of the model consists of a stance phase and A transition from one stance phase to the next.

Figure 1 .Figure 2 .
Figure 1.The schematic representation of the walking bipedal robot foot Figure 2 shows the dynamics of the walking bipedal robot model in the x -direction, with a walking motion that consists of two steps and two stance phases.The stance phase dynamics are calculated by: 2 0 p p   (1)

Figure 3 .
Figure 3. Two-step gait cycle for the walking bipedal robot

Figure 4 .
Figure 4. 3-step strategy of the model in transition from the initial state to the desired state

Figure 5 .
Figure 5. Simulation results of a change in walking speed using the dynamic foot placement control algorithmAnother benefit of this approach is that the walking objective does not have to be a gait cycle per se.A highlevel controller can use the dynamic foot placement controller to track a desired gait pattern, which could include reaching a standstill or changing walking speed (Figure5); a speed change is achieved by changing the desired gait parameters.For the depicted speed change, the desired speed # # x S / t is changed from 0.5 to 0.75.The new desired speed is reached within three steps after the desired speed change.

Figure 6 .
Figure 6.Simulation results of a change in walking direction using the dynamic foot placement control algorithm

Table 1 .
Parameters and constraints for different numbers of steps