Research Article  Open Access
Synergistic Design of the Bipedal LowerLimb through Multiobjective Differential Evolution Algorithm
Abstract
The complexity in the design of bipedal robots has motivated the use of simple mechanisms to accomplish the desired locomotion task with a minimum control effort. Nevertheless, a diverse set of conflictive design criteria must be met to develop the bipedal gait. In this paper, the synergy in the eightbar mechanism design criteria to satisfy the bipedal lowerlimb behavior is promoted by proposing a Paretobased Nonlinear Mixed DiscreteContinuous Constrained Multiobjective Optimization Problem and by improving the search in the optimizer through the inclusion of the Multiselection Strategy into the multiobjective differential evolution algorithm, where the conflictive design objectives are the continuous path generation based on an approximation error to track twenty precision points of the bipedal gait and also the force transmission exerted when the foot reaches the ground. The manufacture of a prototype with a specific design tradeoff experimentally validates the obtained synergistic design with the proposed design approach. In addition, the empirical analysis in simulation through a statistical validation indicates that the Multiselection Strategy explores and exploits the design solutions promoting the diversity, convergence, and capacity features of the obtained Pareto front with respect to other four multiobjective optimizers and consequently improves the reconfigurability in the design such that more alternatives result to the designer decision making.
1. Introduction
Human beings have handled and solved significant problems throughout history, producing theories and technologies even more sophisticated [1]. One of these theoretical and practical aspects is focused on the development of highperformance bipedal robots in terms of energy consumption, design complexity, number of actuators, and so on. Despite the complexity in their mechanical design (interrelation of several aspects: temporalspatial data, kinematic variables, and the like), a significant challenge in the design of bipedal robots is to develop it as simple as possible implying a prototype with low cost, low energy consumption, and easy to control.
In that way, there has been a growing interest in the design of bipedal robots focus mainly on lowerlimb rehabilitation [2] and lowerlimb prostheses [3, 4]. Furthermore, several design studies that involve simple mechanisms have been exploited to reproduce the human gait: sixbar mechanism [5], Chebyshev mechanism [6], and tenbar mechanism [7]. Nevertheless, the limitation of the linkage mechanism to generate continuous paths is that the desired continuous path can only be generated approximately and then, in some of cases, optimization approaches have been stated to satisfy several precision points of designs. The optimization problem was solved by using metaheuristic algorithms instead of gradientbased approaches because the latter are problem dependent (the convergence to a solution is only guaranteed for continuous problem); they present high sensitivity to initial condition which results in the convergence to diverse local minima in multimodal optimization problems, and they require the gradient of the performance function and/or constraints which is not always possible for realworld applications. For instance, Calva et al. [8] synthesized a reconfigurable fourbar linkage to the rehabilitation gait in children with cerebral palsy. A monoobjective constrained optimization problem was formulated and solved by differential evolution (DE) and Sequential Quadratic Programming (SQP) techniques. The DE metaheuristic algorithm provided a better performance than the SQP gradientbased approach due to the fact that the former efficiently explores the search space in the constrained problem.
Therefore, metaheuristic approaches are gaining more attention lately [9, 10] because they can find solutions near to the optimum, be less prone to get stuck in local minima due to the fact that they are populationbased approaches, are merged to other techniques to improve the performance, and are used in nonlinear, discontinuous, continuousdiscrete problems. Approaches that employed metaheuristic algorithms have been recently developed in the design of rehabilitation robots based on mechanisms. Shao et al. [11] performed the dimensional synthesis of a camlinkage mechanism to human gait rehabilitation, considering a multiobjective optimization problem. A weighed sum approach was used to transform the multiobjective optimization problem into a monoobjective one. This problem was solved via genetic algorithm. The main proposal advantage is that the twodegreeoffreedom in the initial design was effectively reduced to a onedegreeoffreedom camlinkage mechanism. The results proved that the final design was capable of tracking the desired path. Bataller et al. [12] synthesized an exoskeleton for hand finger rehabilitation through a lowcost mechanism. The objective designs were the deviation of the generated path, as well as the difference of the three angles of each phalanx. The authors handled the optimization problem with the weighted sum approach which was solved by Malaga University Mechanism Synthesis Algorithm. In order to validate the design, a 3Dprinted prototype was manufactured. Ji and Manna [13] synthesized a planar fourbar mechanism to track a closed ankle trajectory obtained from gait data. By establishing and solving a monooptimization problem was possible to find a design which minimizes the trajectory tracking error. A conceptual design of the linkage mechanism was displayed to expose the idea of creating machines using planar mechanisms. McDaid [14] performed a two degreeoffreedom fivebar linkage to reinstate walking patterns for children with cerebral palsy, to which, a multiobjective optimization problem was formulated and solved via genetic algorithm with a weighted sum approach. The design was experimentally validated with four healthy subjects, as well as with clinical data from three children with cerebral palsy. Singh et al. [15] proposed a novel synthesis method, considering a twostage monoobjective optimization problem and applying to a fourbar exoskeleton lowerlimb to mimic the human gait. The first stage involves the tracking error between the generated and desired trajectories for the design a fourbar linkage. In the second stage the configurations of the linkage for the remaining precision points are determined. A conceptual design of an exoskeleton demonstrated the viability of the method. In [16], the same authors synthesized a crankrocker planar mechanism (human kneeexoskeleton) by formulating a constrained monoobjective optimization problem with the error between generated and prescribed trajectory as the performance function. The optimization problem was solved by teachinglearningbasedoptimization algorithm, considering the penalty method to manage all the constraints. A prototype was used to validate the design, observing that a subject was able to walk smoothly with the kneeexoskeleton designed. Shen et al. [17] proposed an integrated method of type and dimensional synthesis to design one degreeoffreedom mechanisms with only revolute joints based on the fourbar mechanisms. During the design process a monoobjective optimization problem is formulated and solved via genetic algorithms. The proposed method was applied to a leg exoskeleton for hip and knee joint gait reproduction and a finger exoskeleton for curling motion duplication. Pantoja et al. [18] synthesized an eightbaronedegreeoffreedom mechanism to track a trajectory similar to the human gait on sagittal plane. A monoobjective optimization problem was formally stated to fulfill the design requirement. Experimental results verified the design approach through a 3Dprinted prototype.
An important aspect to be considered in the bipedal design is how to synergistically improve the tradeoffs in the design criteria. In previous works related to simple bipedal design the tradeoff among design criteria is a priori given through weighted sum approach. Nevertheless, once the weights are set in the weighted sum approach, a specific tradeoff is found such that the reconfigurability of the design can not be achieved [19]. In addition, it is quite impossible to obtain the suitable synergy between different design criteria because an evenly distributed set of weights do not produce a uniform distribution of the Pareto front points [20]. Then, to perform an efficient synergy among design criteria in the bipedal lowerlimb design, a Paretobased Multiobjective Optimization Problem [10] must be considered to efficiently search several tradeoffs for the specific application. Moreover, the optimization technique to be used in the synergistic design is also a crucial factor to find the appropriate tradeoffs in the design with several criteria [21]. Due to the features and properties of the problem, there is not a universal algorithm that performs better for all classes of multiobjective optimization problems [22]. Hence, to find the best synergy in the design implies the inclusion of strategies which promotes an efficient exploration and exploitation of the design space to fulfill diverse tradeoffs in the design.
Therefore, in this paper, the synergistic design of a bipedal lowerlimb through a onedegreeoffreedom eightbar mechanism is accomplished by (i) Proposing a Paretobased Nonlinear Mixed DiscreteContinuous Constrained Multiobjective Optimization Problem (NMDCCMOOP). The proposed synergistic design approach integrates the continuous path generation based on the tracking of twenty precision points of the bipedal gait and the force transmission (via transmission angle) exerted in the stance phase. (ii) Including a Multiselection Strategy into the selection stage of the multiobjective differential evolution algorithm. This strategy explores specific regions of the search space and promotes to discover a large amount and spread of Pareto solutions, providing more and better alternatives to the designer decision making (more reconfigurability in the bipedal lowerlimb design).
The paper is organized in the following way: Section 2 presents the kinematic model of the Bipedal LowerLimb (BLL). The objective functions, the design variables, and the constraints involved in the Nonlinear Mixed DiscreteContinuous Constrained Multiobjective Optimization Problem are explained in Section 3. Section 4 explains the proposed algorithm to solve the NMDCCMOOP. Section 5 covers an assessment of the effectiveness of the proposed DEbased selection strategy algorithm and other four algorithms. Besides, experiments of the best mechanism found are conducted to verify the synergistic design. Finally, in Section 6 the conclusions are drawn.
2. Analysis of the Bipedal LowerLimb
Two identical mechanisms are required to generate the human gait on the sagittal plane, where each is a bipedal lowerlimb. However, it is sufficient to design only one element instead of the two required due to the symmetrical behavior of the gait. Therefore, the one degreeoffreedom eightbar planar mechanism presented in Figure 1 is designed to generate the human gait path on the sagittal plane. The link lengths is grouped in the vector , the link angular displacements are represented as , the terms are the interior angles of the triangular links marked as A and B, and the Cartesian coordinate corresponds to the point where the desired gait movement would be set. This mechanism is selected due to its scalable design, energy efficiency, bioinspired locomotion, and it has a wide operation range capable of generating closedloop trajectories [23].
2.1. Kinematic Model
Separating the BLL into three submechanisms (i.e., two fourbar and one fivebar mechanism), kinematics equations are derived in (1) and (2) from [18].where
3. Synergistic Design of the BLL
The design of the BLL takes into account the synergy between the generation of the human gait path on the sagittal plane and the force transmission from the input link to the output point PE. Therefore, in order to fulfill a synergistic design approach with different tradeoffs, a multiobjective optimization problem is stated as a mathematical programming problem, where (4) and (5) show its general representation.subject towhere
: vector of design variables;
: number of inequality constraints;
: number of equality constraints.
In the next sections, the functions related in the multiobjective optimization problem for the synergistic design of the BLL are detailed.
3.1. Design Objectives
The first design objective displayed in (6) is related to the accuracy of the path tracking. This considers the error between the desired path , and the Cartesian coordinates tracked by the point , where is the number of precision points to satisfy the point of the BLL. The second criterion displayed in (7) provides a design with the most suitable force transmission during the stance phase of the BLL and also is less prone to inaccuracies due to manufacturing errors, thermal expansion, or material contraction. Then, the second design objective is the deviation of the transmission angle of the three submechanisms from the ideal one (pressure angle [25]).where
3.2. Design Variables
The design variables of the BLL are the link lengths , the angles of the fixed links, the interior angles of the triangular links, the angle deviation , the start coordinate of the gait path, kinematic configurations of submechanisms (elbow up and elbow down configuration), and the angular displacements of the crank . Finally, the vector in (9) collects the design variables, where the discrete variables are given by and the rest consists of real (continuous) ones.
3.3. Constraints
3.3.1. Lower and Upper Bounds of Design Variables
Bounds on design variables delimit the search space to finding an optimal solution in realworld engineering problems. Such bounds are set as inequality constraint in (10), where the upper and lower limits are given in Table 1. It is important to note that the variables can only take the discrete values of or .

3.3.2. Grashof Criterion
Another constraint in the design of the BLL is the Grashof criterion of the fourbar submechanisms [26], which states if the sum of the shortest and longest link lengths is less than or equal to the amount of the two remaining links, then at least one of them can rotate completely. This criterion is set as inequality constraints in (11) and (12).
Besides, the fourbar submechanisms can be locked if the inequality constraints (13)(16) are not taken into account. These constraints force that such mechanisms present crankrocker configuration.
3.3.3. Functional Morphology
The functional morphology constraint is associated with the conditions that the BLL can move freely (i.e., without any of its joints hitting the ground). The inequality constraints given in (17)(19) force that the points , , and are above of the point where the path is performed.where
3.3.4. Real Kinematic Movement
This set of constraints does not allow that submechanisms present nonphysical design (i.e., mechanisms that do not satisfy the closed kinematic chain). Then, each of the crank input angular motions must guarantee that constraints (21) and (22) are satisfied to produce the motion in the point of the BLL.
3.3.5. Desired Path
The desired path uses the sign convention in Figure 2 for the hip and knee angles of the lowerlimb of human being. Moreover, this path includes the torso movement (see Figure 3). Then, the desired path is parameterized and resized in (23) considering the hipknee angles presented in the biomechanical study in [24]. Twenty hipknee angles of a younger group between 6 to 12 years old collected from such study are used and shown in Table 2. As a result, the precision point number for the desired path is , where means that of the gait cycle is in the stance phase, and means that of the gait cycle is in the swing phase.
The desired path expressed in (23) is included as inequality constraints, taken into account the lengths and and the scale factors and yielding a gait path with a step length of about and a step elevation of about . It is important to note that this path presents a closed form, and an importance challenge in the bipedal lowerlimb design is to track this two degreeoffreedom Cartesian space trajectory with one eightbar mechanism.where
4. Multiobjective Differential Evolution Algorithm
Differential Evolution (DE) proposed by Price and Storn in 1995 [27] is a populationbased metaheuristic global search technique. DE is formulated to solve numerical optimization problems coded with real values as in evolution strategies. In the same way as evolutionary algorithms, DE presents operators of recombination, mutation, and selection strategies. DE has had a remarkable acceptance in the diverse areas such as mechanical engineering design [28], mechatronic design [21], automatic control tuning [29], and so on. This fact is because of the suitable convergence and robustness properties, the reduced number of tuning parameters, and the success in solving realworld applications. Nevertheless, the search capability of DE may be compromised in multimodal problems [30]. Accordingly, further improvements must be made to enhance (avoid stagnation to local optima) its performance for a particular optimization problem.
The DE process consists of three stages: mutation, crossover, and selection. In the mutation process, the individuals in the population change the search direction based on the distribution of three solutions and the step size , as is shown in
In the recombination process (26), some design parameters of parent individuals and mutant vectors share their genetic information to generate the offspring population . The parameter influences the probability to obtain information from either parent or the offspring (great values mean less impact of the parent individual).
The generated offspring competes against of its corresponding parent and the most competitive survives to the next generation. The mechanism to know the most suitable individual is based on the proposed Multiselection Strategy explained in the next section. This strategy promotes a better convergence as well as widespread solutions in the Pareto front. Additionally, based on [31], the crowding distance strategy [32] and an External Archive (ExA) are employed to keep the best individuals (based on Pareto dominance) through the optimization process, allowing promoting diversity in the solutions. This strategy interchanges individuals stored in the ExA to provide the information to the mutation process, meaning that the individuals required in the mutation process are taken from the current population or from the ExA, depending on the Normal Selection (NS) value. This process is repeated until a maximum number of generations is fulfilled.
Algorithm 1 displays the DEbased approach with the Multiselection Strategy.
1: Create a random initial population , and  
2: Evaluate in performance functions.  
3:  
4: while do  
5:for to do  
6:if then  
7:Select randomly  
8:else  
9:Choose , and from ExA  
10:end if  
11:  
12:for to do  
13:if or then  
14:  
15:else  
16:  
17:end if  
18:end for  
19:MultiSelection Strategy between and based on Algorithm 2  
20:Update the ExA with the information of the best performance individuals.  
21:  
22: end for  
23: end while 
1: Virtual split of into: ,  
2: if & then  
3:  
4: else if & then  
5:if then  
6:  
7:else  
8:Whether :▷VS1  
9:if then  
10:  
11:end if  
12:Whether :▷VS2  
13:Random choice between and  
14:end if  
15: else if & then  
16:Choose the one with the smaller constraint violation i.e., vs  
.  
17: else  
18:  
19: end if 
The discrete variables in of the design variable vector (9) are handled as in (27); i.e., once the discrete design variable is generated (randomly generated in the population initialization in line 1 or after the crossover process in line 14 of the Algorithm 1) the procedure in (27) is performed.
4.1. Proposed Multiselection Strategy
Some concepts related to multiobjective problems are firstly given to describe the proposed Multiselection Strategy (MSS).
Definition 1. Let be the full set of alternatives for the decision variables, the feasible region is such that .
Definition 2. Given two vectors , we say that dominates (denoted by ) if & , for .
Definition 3. A vector of decision variables is Paretooptimal if it is nondominated with respect to .
Definition 4. The Paretooptimal set is defined by
Definition 5. The Pareto Front is defined by
The proposed MSS promotes the spread, diversity, convergence, and capacity [33] of the solutions through the Pareto front to provide a diverse set of bipedal lowerlimb designs that fulfill different tradeoffs between design objectives. This strategy splits the population of individuals into two subpopulations with the same number of individuals named as Virtual Subpopulation 1 (VS1) and Virtual Subpopulation 2 (VS2). Each subpopulation implements different selection processes. The VS1 enhances the convergence and the spread of a relevant region of the Pareto front (exploitation), and the VS2 explores the solutions space to promote diversity (exploration) in the search. This strategy is included when the offspring and the parent competes (line 19 of Algorithm 1). The MSS follows the following rules:(1)Any feasible solution is preferable over any unfeasible solution.(2)If the two solutions are feasible, the nondominated one is preferred; if both solutions are nondominated, then Individuals in VS1:(i)The one having the smallest value of is preferred. Individuals in VS2:(ii)One of them is chosen considering the same probability to be selected (flip a coin).(3)Among two unfeasible solutions, the one having a smaller constraint violation is preferred.
Algorithm 2 displays the proposed selection. The MSS is similar to the Deb Criterion [34]; however the main differences are the way in which both individuals are nondominated. The VS1 searches in the improvement of the performance function (path tracking accuracy) as can be seen in line 9 of the Algorithm 2, while VS2 promotes a random search into the design space; i.e., any solution is chosen with the equal probability to be selected (observe line 13 of the Algorithm 2).
The MSS encourages a Pareto front more widespread towards the interest regions (to minimize ) of designer interest and also promotes more synergistic solutions because different tradeoffs are found in the region of interest.
4.1.1. Promoting Diversity
The diversity of the Pareto front is promoted in the DE algorithm by using the diversity strategy introduced by the Crowding Distance (CD) in [31]. The CD estimates the perimeter of the nearest neighbors of each solution. Algorithm 3 shows how to compute the crowding distance. The Normal Selection (NS) parameter is introduced to change the selection of vectors in the mutation process. When the generation is greater than of the generation maximum number , the mutation process uses three individuals (, , and ) from an External Archive instead of selecting them from the current population (as in the traditional DE algorithm). Such selection is based on CD. Larger values of CD are preferred since they represent a measure to know uncovered regions of nondominated solutions. Therefore, more uniform Pareto front is obtained by using this diversity strategy.
1: Begin  
2:  
3:Sort in descent order the nondominated set.  
4: and .  
5:for to do  
6:for to do  
7:  
8:end for  
9:end for  
10: End 
Storing in the External Archive. This work uses an External Archive (ExA) to store the nondominated solutions along the search. The ExA allows finding better solutions when the diversity strategy based on CD is applied and thus accomplishes with design requirements appropriately [35]. The ExA allows the entrance of a solution if it is feasible and dominates any of the solutions inside the ExA. The dominated solutions from the ExA are removed. This process is expressed in the following definition.
Definition 6. Let , , then is accepted in the External Archive and , where is the maximum solution numbers of External Archive.
This ExA is limited to a maximum number of nondominated solutions . When the number of nondominated solutions surpasses the maximum number of individuals in ExA , the excesses of individuals are deleted. This removal of individuals is carried out equidistantly every of the nondominated solutions in the ExA. Line 20 of the Algorithm 1 indicates the store and removal process of surplus solutions in ExA.
5. Results
The synergetic design problem of the bipedal lowerlimb is solved by using the proposed multiobjective differential evolution based on Multiselection Strategy explained in Section 4. Other four optimizers are implemented to compare the performance of the proposed optimizer. The first comparative DEbased multiobjective optimizer is denoted by “A1” which does not include either a strategy to promote diversity or a Multiselection Strategy [36]. The second DEbased multiobjective optimizer, denoted by “A2”, uses the crowding distance to promote diversity and furthers elitism by the inclusion of the External Archive which stores nondominated solutions [31]. The third optimizer is termed as “A3” and is related to the proposed multiobjective differential evolution based on Multiselection Strategy. The fourth optimizer is the Nondominated Sorting Genetic Algorithm II (NSGAII) [37] denoted by “A4”. The last optimizer is the Multiobjective Particle Swarm Optimization (MOPSO) [38] termed as “A5”. This includes boundary position update technique with absorb zero velocity update strategies [39] and also the best position of the swarm is taken from the knee solution of the Pareto front in the External Archive. The selection of the comparative optimizers is to provide enough information about the advantages of including the proposed Multiselection Strategy into the Multiobjective DE algorithm.
5.1. Comparative Analysis of Algorithms
An overall number of independent runs of each algorithm are performed to obtain relevant information about their performance. In the case of DEbased optimizers, their parameters are set as follows: (i) the crossover probability is chosen as . This was obtained through a trial and error procedure. (ii) The scale factor is randomly selected in the interval for each generation, as is described in [40]. (iii) The External Archive capacity is limited to solutions with the Normal Selection of for algorithms “A2” and “A3”. (iv) The population size and the maximum number of generations are set as NP = 20 and Gmax = 500000, respectively. In the case of “A4” and “A5”, the optimizer parameters are chosen according to the recommended by the author cited. Only “A5” a population size of is selected. All algorithms are implemented in over a PC with an i73.5 GHz CPU and 16 GB in RAM.
This work uses the normalized Hypervolume metric () [33] to evaluate the performance in the diversity and convergence of the obtained Pareto front by using the five mentioned algorithms. This metric represents the normalized area covered by the Pareto front and a reference point. Whether the reference point is the Nadir vector, higher values of the metric indicate that the Pareto front has higher diversity and convergence to the True Pareto front. The Nadir vector is obtained by taking the worst independently objective function of the nondominated solutions for all algorithm runs. Since the True Pareto front is unknown in realworld engineering problems, in this work all resulting Pareto fronts from all algorithm runs are filtered to generate an approximation of the True Pareto front. The filtered process involves a dominance checking of all solutions given by runs and thus generates one single Pareto front. By using the approximation of the True Pareto front, the Ideal vector can be obtained with the best independently objective function of the nondominated solutions. Figure 4 shows the obtained approximation of the True Pareto front, where the Ideal and Nadir vectors can be observed.
Table 3 shows the value of each algorithm execution, its mean value , and standard deviation . It is observed that the proposed algorithm “A3” presents more reliable results in the convergence and diversity of the obtained Pareto front because it presents larger mean value and standard deviation in metric.
