Journal of Biomechanics 123 (2021) 110530

Contents lists available at ScienceDirect

Journal of Biomechanics
journal homepage: www.elsevier.com/locate/jbiomech
www.JBiomech.com

Evaluating cost function criteria in predicting healthy gait

K. Veerkamp a,b,c,⇑, N.F.J. Waterval a,d, T. Geijtenbeek e, C.P. Carty b,c,f,g, D.G. Lloyd b,c,
J. Harlaar e,h, M.M. van der Krogt a
a
Amsterdam UMC, Vrije Universiteit Amsterdam, Department of Rehabilitation Medicine, Amsterdam Movement Sciences, de Boelelaan 1117, Amsterdam, the Netherlands
b
School of Allied Health Sciences, Griffith University, Gold Coast, Australia
c
Griffith Centre for Biomedical & Rehabilitation Engineering (GCORE), Menzies Health Institute Queensland, and Advanced Design and Prototyping Technologies Institute
(ADAPT), Griffith University Gold Coast, Australia
d
Amsterdam UMC, Univ of Amsterdam, Rehabilitation Medicine, Amsterdam Movement Sciences, Meibergdreef 9, Amsterdam, the Netherlands
e
Department of Biomechanical Engineering, Delft University of Technology, Delft, the Netherlands
f
Department of Orthopaedics, Children’s Health Queensland Hospital and Health Service, Queensland Children’s Hospital, Brisbane, Australia
g
Research Development Unit, Caboolture and Kilcoy Hospitals, Metro North Health, Australia
h
Department of Orthopaedics, Rotterdam, Erasmus Medical Center, the Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: Accurate predictive simulations of human gait rely on optimisation criteria to solve the system’s redun-
Accepted 10 May 2021 dancy. Defining such criteria is challenging, as the objectives driving the optimization of human gait are
unclear. This study evaluated how minimising various physiologically-based criteria (i.e., cost of trans-
port, muscle activity, head stability, foot–ground impact, and knee ligament use) affects the predicted
Keywords: gait, and developed and evaluated a combined, weighted cost function tuned to predict healthy gait.
Human locomotion A generic planar musculoskeletal model with 18 Hill-type muscles was actuated using a reflex-based,
Forward dynamic simulations
parameterized controller. First, the criteria were applied into the base simulation framework separately.
Neuro-musculoskeletal modelling
Computational biomechanics
The gait pattern predicted by minimising each criterion was compared to experimental data of healthy
Feedback control gait using coefficients of determination (R2) and root mean square errors (RMSE) averaged over all biome-
chanical variables. Second, the optimal weighted combined cost function was created through stepwise
addition of the criteria. Third, performance of the resulting combined cost function was evaluated by
comparing the predicted gait to a simulation that was optimised solely to track experimental data.
Optimising for each of the criteria separately showed their individual contribution to distinct aspects of
gait (overall R2: 0.37–0.56; RMSE: 3.47–4.63 SD). An optimally weighted combined cost function pro-
vided improved overall agreement with experimental data (overall R2: 0.72; RMSE: 2.10 SD), and its per-
formance was close to what is maximally achievable for the underlying simulation framework. This study
showed how various optimisation criteria contribute to synthesising gait and that careful weighting of
them is essential in predicting healthy gait.
Ó 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://
creativecommons.org/licenses/by/4.0/).

1. Introduction known how human gait is optimised, and how to specify and
weigh the involved criteria.
Forward dynamic simulations can be used to predict new pat- A number of different physiologically-based criteria have been
terns of human movement without requiring experimental data. used in predictive simulations of human gait. Many experimental
Thus, such simulations have the potential to investigate the mech- studies have shown that the energetic cost of transport (CoT) is
anistic cause-and-effect relationships in movement impairments, involved in human gait selection (Abram et al., 2019; Bertram
optimise the prescription of assistive devices, and predict out- and Ruina, 2001; Donelan et al., 2001; Minetti et al., 2020;
comes of interventions. A critical aspect in predictive simulations Selinger et al., 2015; Zarrugh et al., 1974) and, consequently, it is
is the cost function, which sets the goal of the movement and aims also the most commonly considered criterion in predictive simula-
to replicate human gait optimisation. At present, it is not clearly tions of gait (Anderson and Pandy, 2001; Dorn et al., 2015; Falisse
et al., 2019; Lai et al., 2018; Ong et al., 2019; Song and Geyer, 2015;
Wang et al., 2012). Besides energy cost, human gait might also be
⇑ Corresponding author. optimised to prevent muscle fatigue by minimising muscle
E-mail address: [email protected] (K. Veerkamp).

https://doi.org/10.1016/j.jbiomech.2021.110530
0021-9290/Ó 2021 The Authors. Published by Elsevier Ltd.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

activation, as suggested by the muscle force-endurance relation- ques (Dorn et al., 2015; Geyer and Herr, 2010; Ong et al., 2019),
ship (Crowninshield and Brand, 1981; Müller and Grosse- which is particularly relevant for the knee joint that may approach
Lordemann, 1937), and as applied in several predictive simulation end-of-range of motion during gait.
studies (Ackermann and van den Bogert, 2010; Falisse et al., 2019; Although it is likely that human gait is optimised for a variety of
Lai et al., 2018). Head stability measures have also been used in these physiologically-based criteria, it is unclear how each of these
simulation studies (Dorn et al., 2015; Nguyen et al., 2019; Ong criteria contribute to gait. To achieve more representative predic-
et al., 2019) supported by the finding that human head accelera- tions, it is necessary to better understand the individual effects
tions, detected by the otolith organs in the ear (Kandel et al., of these criteria, as well as their relative contribution to human
2013), remain similar across different walking conditions (Menz gait. Therefore, this study’s two aims were to 1) evaluate the effects
et al., 2003). Further, a recent commentary has suggested how of minimising the criteria CoT, muscle activity (MusAct), head sta-
the derivative of force, defined as ‘yank’, is an important factor in bility (HeadStab), foot–ground impact (FGImpact), and knee hyper-
sensorimotor systems (Lin et al., 2019). Indeed, mitigating the extension (KneeExt) on gait using predictive simulations, and 2)
impact of external forces, or yank, which could be detected by develop and evaluate an optimally weighted, combined cost func-
the cutaneous mechanoreceptors in the foot (Kavounoudias and tion tuned to predict healthy gait.
Roll, 1998; Kennedy and Inglis, 2002), may be involved in human
walking, as a high rate of loading has been suggested to contribute 2. Methodology
to injuries by applying high stresses to the leg’s tissues
(Mündermann et al., 2005; Nigg, 1985). Finally, the musculoskele- 2.1. Experimental gait data
tal models in previous simulation studies included a joint limit tor-
que, representing ligaments, that are activated during joint hyper- Motion capture data was collected using the Plug-In-Gait mar-
flexion or -extension (Dorn et al., 2015; Geyer and Herr, 2010; Ong ker set (Vicon Nexus 2.6.1; Davis, Õunpuu, Tyburski, & Gage, 1991;
et al., 2019). Mimicking type III joint receptors firing when reach- Kadaba, Ramakrishnan, & Wootten, 1990) for ten healthy adults (4
ing these extreme joint angles (Nyland et al., 1994; Zimny and males, age 26.8 ± 2.6 years, mass 67.2 ± 8.5 kg, height 1.76 ± 0.0
Wink, 1991), those studies minimised using these joint limit tor- 8 m) walking overground at their comfortable walking speed.

Table 1
Overview of the musculotendon parameters that were used in the musculoskeletal model.

MTU in model Gait2392 Tendon slack Optimal fiber Maximum Pennation angle at optimal Tendon strain at maximum Slow twitch
pathway length length isometric force fiber length isometric force ratio
Hamstrings bifemlh 0.326 0.109 2594 0.000 0.049 0.499
Biceps femoris bifemsh 0.089 0.173 1122 0.401 0.049 0.529
short head
Gluteus maximus glut_max2 0.127 0.147 4759 0.000 0.049 0.550
Iliopsoas psoas 0.160 0.100 5148 0.144 0.049 0.500
Rectus femoris rect_fem 0.310 0.114 1169 0.052 0.049 0.387
Vasti vas_int 0.136 0.087 4530 0.087 0.049 0.484
Gastrocnemius med_gas 0.390 0.060 2241 0.297 0.100 0.546
Soleus soleus 0.250 0.050 7147 0.436 0.100 0.759
Tibialis anterior tib_ant 0.223 0.098 1597 0.087 0.049 0.721

Table 2
Overview of the gait phases and muscle reflexes used in our controller.

Early Stance Late Stance Pre-swing Swing Late swing

Threshold measured to initiate GRF greater than Sagittal distance Contralateral foot enters GRF lower than Sagittal distance swing
gait phase threshold* stance foot* early stance threshold* foot*
Hamstrings C0 C0 C0 F+ F+
PD from pelvis tilt PD from pelvis tilt PD from pelvis tilt
Biceps femoris short head C0 C0
L+ L+
Gluteus maximus C0 C0 C0 F+ F+
PD from pelvis tilt PD from pelvis tilt PD from pelvis tilt
Iliopsoas C0 C0 C0 C0 C0
PD from pelvis tilt PD from pelvis tilt PD from pelvis tilt PD from pelvis tilt PD from pelvis tilt
L+ L+
L- from hamstrings L- from hamstrings
Rectus femoris C0 C0 C0 C0 C0
L+ L+ L+ L+ L+
Vasti C0 C0 C0
F+ F+ F+
Gastrocnemius F+ F+ F+
Soleus F+ F+ F+
Tibialis anterior C0 C0 C0 C0 C0
L+ L+ L+ L+ L+
F- from soleus F- from soleus F- from soleus

* indicates thresholds differentiating gait phases that were optimised. C0 was a constant stimulation value. PD was a proportional-derivative reflex, based on the pelvis tilt
angle. F was a force-based reflex, L a length-based reflex. + and – were positive and negative feedback, respectively. Reflex loops mostly worked within muscles, but also
between muscles (i.e., iliopsoas-hamstrings and soleus-tibialis anterior), to prevent co-contraction.

2
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Ground reaction forces (GRFs) were simultaneously collected from trunk merged) and nine musculotendon units per leg. For each
two embedded force plates (AMTI, Watertown, MA, USA). Muscu- group of muscles working around the same degrees of freedom,
loskeletal modelling was conducted in OpenSim (version 4.0; the musculotendon pathway and parameters from a representative
Seth et al., 2018), scaling the gait2392 musculoskeletal model muscle were selected, and its maximum isometric force was
(Delp et al., 1990) to each participant’s proportions using a static updated by adding the forces from the other muscles to that group.
T-pose trial. Inverse kinematic and inverse dynamic tools were The Millard-equilibrium muscle model was used (Millard et al.,
used to compute joint angles and moments from walking trials. 2013) with Table 1 providing an overview of its parameters. Of
Joint powers were calculated from these angles and moments note, tendon strain at maximum isometric force was 4.9% for all
using custom-written scripts in MATLAB 2016a (The MathWorks muscles (Millard et al., 2013; Rajagopal et al., 2016), except for
Inc., MA). The ensembled averaged GRFs and joint angles, moments the gastrocnemius and soleus that were set to 10% (Arnold et al.,
and powers were used in further comparisons as normative exper- 2013; Ong et al., 2019). The ratio between slow and fast twitch
imental data for healthy gait. Additionally, normative muscle acti- fibres for each muscle was similar to Ong et al. (2019). A knee
vation patterns were derived from Bovi et al. (2011), including end-of-range of motion limit torque was added that was activated
electromyography data for the gluteus maximus, rectus femoris, when the knee extended below 5 degrees of flexion (Markolf et al.,
biceps femoris, vastus medialis, gastrocnemius medialis, soleus 1990), with a stiffness of 2 N*m/degree and a dampening of 0.2 N
and tibialis anterior muscles using a ZeroWire system (Aurion, *m/(degrees*second). A Hunt-Crossley foot–ground contact model
Milano, Italy). with two viscoelastic spheres at each foot was added (Hunt and
Crossley, 1975) and the geometry of the spheres was initially opti-
2.2. Base simulation framework mised minimising the error between simulated and experimental
GRFs and kinematics (see Appendix A for details).
Our base simulation framework was developed using the open- A reflex-based controller (Geyer and Herr, 2010) was imple-
source software SCONE (version 1.4.0; Geijtenbeek, 2019). A gen- mented into our base simulation framework, detailed in Table 2.
eric musculoskeletal model (gait2392; Delp et al., 1990) was sim- This type of controller results in intrinsically stable neuromuscular
plified to nine degrees of freedom (sagittal plane only; pelvis and simulations, and was therefore favoured over open loop controller

Fig. 1. The gait patterns predicted by minimising for different criteria, compared to normative experimental data. The variables from the simulated gait were averaged over
time-normalised gait cycles in the ten-second simulation, excluding the first two gait cycles.

3
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

in this study. The controller consisted of different types of muscle more than 0.01% per generation. The simulation with the lowest
force- and length-based reflex loops working within and between score over all six optimisations was selected for subsequent evalu-
muscles, activating the musculotendon units in different gait ation. From the ten-second simulation, the first two gait cycles
phases (i.e., early stance, late stance, pre-swing, swing and late were excluded, and waveforms were time-normalised and aver-
swing). The trunk was controlled by a proportional-derivative aged over all subsequent gait cycles, which were typically all very
(PD) reflex controller based on the pelvis + trunk tilt angle. Reflex similar.
delays were also set as in Geyer and Herr (2010). For the initial
state, the initial pose rate of change was selected from pilot exper- 2.3. Evaluation cost functions
iments. The initial muscle activation was set according to initial
controller output, while the initial length of the contractile ele- Various cost functions were added to the base simulation
ment was found after equilibration of the muscle tendon dynamics, framework, consisting of the following criteria in different
based on initial muscle–tendon length, velocity and activation. The compositions:
model’s initial pose, the thresholds between gait phases, and the i. CoT, an overall effort measure, was calculated using the mus-
reflex gains and offsets of each muscle for each gait phase were cle metabolic model by Umberger, Gerritsen, & Martin (2003) and
optimised. There were 98 design variables: seven for the model’s implemented based on Uchida, Hicks, Dembia, & Delp (2016). The
initial pose, four to differentiate gait phases, and 87 for the muscle calculated energy rate was summed for all muscles (m) over the
reflexes. Except for the initial pose, all design variables were sym- simulation (until tend ), and divided by the mass of the model and
metric for the left and right legs. distance travelled, i.e.:
The design variables were optimised by adding different cost Z t end X
18
1
functions, specified below, to the base simulation framework. CoT ¼  ½ E_ m ðt Þdt
When optimising for a cost function, the score, which is the distance  mass 0 m¼1
weighted sum of the involved cost function criteria over the ten-
Muscle fatigue (MusAct) was quantified by the muscle activa-
second simulation, was minimised using a Covariance Matrix
tion squared. The activation of each muscle (m) was squared,
Adaptation Evolution Strategy (CMA-ES; Hansen 2016). A 24-core
summed over the simulation, and divided by the distance travelled,
Intel Xeon CPU E5-2690 2.60 GHz processor was used. In the opti-
i.e.:
misation, penalties were applied when, (1) the model walked
below a speed of 0.5 m/s, (2) the ankle angle was beyond 60
Z t end X
18
1
activ ationm ðtÞ dt
2
degrees dorsi- or plantarflexion, (3) simulation time was less than MusAct ¼  ½
distance 0 m¼1
ten seconds, and (4) the centre of mass was below 0.75 m. These
penalties were used to avoid unrealistic local minima, with the lat- iii. Head stability (HeadStab) was represented by the accelera-
ter measure, indicating a fall, also prematurely ending simulations tion of the head segment. The sum of the horizontal and vertical
to increase optimisation speed. For each cost function, six optimi- accelerations of the head centre of mass (ax and ay , respectively)
sations with different random seeds were performed, until, aver- during the simulation was divided by the distance travelled, i.e.:
aged over the last 500 generations, the cost had not improved

Fig. 2. Quantified agreement (R2 and RMSE) of the gait patterns predicted by minimising for different criteria with normative experimental data. RMSE was normalized to the
standard deviation of each variable. The right column shows the walking speed for each criterion.

4
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Z t end
1 v. KneeExt involved a penalty to the use of the knee limit tor-
HeadStab ¼  ½jax ðtÞj þ jay ðtÞjdt
distance 0 que, representing ligaments, (F limit ) that was summed over the
whole simulation for both the left and right knee, i.e.: KneeExt
iv. The foot–ground impact (FGImpact) was assessed by the
became zero when the knee limit torque was not used (i.e., more
GRFs derivative or yank, penalizing high increases/decreases in
than 5 degrees of flexion during the whole simulation) and for this
the GRF. The sum of the time-derivative of horizontal and vertical
reason, it could not be optimised independently.
GRFs (GRF x and GRF y ) from both the left and right leg during the
simulation was divided by the distance travelled, i.e.: Z t end
Z KneeExt ¼ ½F limit;left þ F limit;right dt
1 dGRF x;left ðtÞ
t end
dGRF y;left ðt Þ
FGImpact ¼  ½j jþj j 0
distance 0 dt dt
Note that tend was ten seconds for each simulation. Besides
dGRF x;right ðt Þ dGRF y;right ðt Þ these five physiologically-based criteria, a cost function criterion
þj jþj jdt
dt dt was used to examine the base simulation framework’s best possi-

Fig. 3. The gait patterns predicted by each step in the combined cost function, compared to normative experimental data. The variables from the simulated gait were
averaged over time-normalised gait cycles in the ten-second simulation, excluding the first two gait cycles.

5
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

ble agreement with experimental data (ExpTrack). In this, we min- optimised independently and their quantified performance was
imised the root mean square error (RMSE) between predicted and evaluated. Second, these criteria were ranked by their overall
experimental GRFs, joint angles, and muscle excitations. To give average R2 when optimised independently and then added in
equal weighting to each biomechanical category, the weighting a stepwise way to create a combined cost function. The crite-
of each variable was divided by its experimental standard devia- rion with the highest average R2 was selected with a set
tion, and then divided by the number of variables within each weighting of 1, and the criterion with the second-highest aver-
category. age R2 was added using five different systematically chosen
Each cost functions’ performance was assessed by the agree- weightings. Out of these, the weighting providing the best aver-
ment between the average predicted and normative experimental age R2 and RMSE was selected, after which the next criterion
waveforms for different categories of biomechanical variables, i.e. was added with five weightings, and so on until all criteria
GRF, joint angles, moments, powers, and muscle excitations. The were combined into an optimal, combined cost function.
performance metrics were coefficients of determination (R2) and KneeExt’s weighting was tuned when the knee showed more
RMSE. RMSE of the excitations was excluded since matching mag- extension than seen in experimental data. For each criterion,
nitudes is not straightforward as these rely on many factors, such its tested weightings were normalised to the criterion’s value
as EMG normalisation (Devaprakash et al., 2016). R2 was consid- when independently optimised, except for KneeExt since its
ered very weak when smaller than 0.3, weak between 0.3 and minimal value was zero, to be able to evaluate the contribution
0.5, moderate between 0.5 and 0.7 and strong above 0.7 (Hair Jr of each criterion in the combined cost function. When a new
et al., 2016). The RMSE for each variable was divided by its average criterion was added, the optimised design variables from the
experimental standard deviation, to allow more fair comparison of previous stepwise combination of cost functions were used as
the RMSE between variables with different magnitudes. R2 and initial values. In stage three, ExpTrack was independently opti-
RMSE were evaluated on three levels: for each variable separately, mised for and the performance of the optimised, combined cost
averaged over each biomechanical category and averaged over all function was compared to this simulation. Here, R2 and RMSE
categories. Walking speed was also assessed. for the moments and powers were excluded in the evaluation,
The cost functions were analysed in three stages. In the first since they could not be tracked using ExpTrack.
stage, criteria i-iv (CoT, MusAct, HeadStab and FGImpact) were

Fig. 4. The muscle excitation patterns predicted by each step in the combined cost function, compared to normative experimental data (Bovi et al., 2011). The excitation
patterns from the simulated gait were averaged over time-normalised gait cycles in the ten-second simulation, excluding the first two gait cycles.

6
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Fig. 5. Quantified agreement (R2 and RMSE) with normative experimental data of the gait patterns predicted in each step of combining the cost functions. RMSE was
normalized to the standard deviation of each variable. The right column shows the walking speed in each step.

Table 3
Overview of composition of the cost function combining all criteria, and the minimised cost function score.

Criteria Criterion’s score when optimised independently Criterion’s optimal cost function weighting* Criterion’s
weighted score
in best
simulation
optimal
combined cost
function
FGImpact Anterior-posterior 1.34 1.0 0.47 15.4%
Vertical 2.67 1.10
HeadStab Anterior-posterior 0.63 1.5 0.92 18.4%
Vertical 0.62 0.96
CoT 3.82 5.0 6.12 60.1%
KneeExt n/a 0.25 0.25 2.5%
MusAct 103.55 0.10 0.37 3.6%

* Normalized to its score when optimised independently (third column), except for KneeExt.

3. Results dicting GRFs (strong average R2: 0.92; RMSE: 3.22 SD) and
moments (moderate average R2: 0.68; RMSE: 2.48 SD), performing
When optimising for each physiologically-based criterion inde- best for the knee moment (moderate R2: 0.54: RMSE: 2.33 SD). It
pendently, different criteria performed best for distinct biome- resulted in a walking speed of 0.92 m/s. FGImpact minimisation
chanical categories (Figs. 1 and 2; more detailed in Appendix B). was best for R2 of the joint powers (weak average R2: 0.46), for
Minimising CoT was best for predicting kinematics as it scored which it resulted in the best ankle push-off power (moderate R2:
highest R2 and lowest RMSE for all angles (strong average R2: 0.67; RMSE: 3.30 SD). Its walking speed was 1.13 m/s. Overall,
0.80; RMSE: 1.54 SD), although, it was worst for predicting GRFs average R2 was highest for FGImpact (0.56), followed by HeadStab
compared to the other criteria (moderate average R2: 0.53; RMSE: (0.51), CoT (0.49) and MusAct (0.37). Hence, the cost functions
9.09 SD). It resulted in the highest walking speed of 1.36 m/s. Min- were combined in this stepwise order.
imising for MusAct resulted in the lowest average RMSE for joint When combining the criteria, the average R2 and RMSE
powers (3.66 SD), and resulted in the slowest walking speed improved in each step, indicating that each criterion contributed
(0.51 m/s). Over all criteria, minimising HeadStab was best for pre- positively to the combined cost function (Figs. 3, 4 and 5; more

7
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

detailed in Appendix C). When combining all cost function criteria, 4. Discussion
the average RMSE was 2.10 SD and the average R2 was 0.72, per-
forming strong for GRFs, kinematics and moments (average R2: In this study we demonstrated that optimising gait for FGIm-
0.93, 0.80 and 0.83, respectively), moderate for powers (average pact, HeadStab, CoT, and MusAct independently predicted different
R2: 0.67) and weak for excitations (average R2: 0.36). In this cost gait patterns which showed good agreement with experimental
function, CoT had the biggest contribution (Table 3) and it also data for distinct biomechanical categories. Combining
had the highest normalised weighting. physiologically-based criteria with tuned weightings in a stepwise
The performance of the combined cost function was close to the approach resulted in an overall improved agreement with experi-
maximal quantified agreement obtained from the ExpTrack simu- mental data. Compared to ExpTrack simulations, the combined
lation, (Figs. 6 and 7; comparison of excitation patterns in Appen- cost function performed very well, which indicates that its agree-
dix D). When comparing GRFs, joint angles and excitation patterns, ment with experimental data is close to its maximum given the
the average R2 (0.70) was only slightly smaller than for the Exp- underlying constraints of the framework.
Track simulation (0.73), while the average RMSE (2.15 SD) was When optimising for the criteria separately, FGImpact overall
slightly higher than for ExpTrack (1.15 SD). showed the best R2 compared to experimental data and its pre-
dicted walking speed was closest to 1.2 m/s, which in general cor-

Fig. 6. The gait pattern and muscle excitation patterns predicted by the cost function combining all criteria, compared to a simulation tracking normative experimental GRFs,
joint angles and muscle excitations. The variables from the simulated gait were averaged over time-normalised gait cycles in the ten-second simulation, excluding the first
two gait cycles.

8
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Fig. 7. Quantified agreement of the gait predicted by the cost function combining all criteria and a simulation tracking normative experimental data with normative
experimental data. RMSE was normalized to the standard deviation of each variable.

responds to the preferred walking speed of humans (Ralston, 1958). teria but also optimising their relative contributions to get an
Its main benefit was the improved ankle push-off, with a better tim- optimal agreement with experimental data.
ing of the ankle power compared to the other criteria, which Overall, performance of our framework was worst for tracking
showed a peak in ankle power generation either too early (CoT the experimental muscle excitations, even in the ExpTrack simula-
and HeadStab) or too late (MusAct). Remarkably, even though CoT tion. The reflex-based controller may be constraining the match
had only a weak R2 when minimised individually, it had the highest with experimental EMG, not allowing a wide enough range of mus-
contribution in the optimal combined cost function (Table 3), cle excitation pattern for better tracking. It has been suggested that
underlying its importance in human gait as also shown before in reflexes, and their spinal loops, are important during walking
predictive simulations and human experiments. Optimising for (Duysens and Forner-Cordero, 2019), but it is hard to validate
CoT by itself led to relatively poor GRFs with a high peak directly which type of reflexes are predominant for each muscle during
after initial contact for both the anterior-posterior and vertical walking, as well as their gain values. We would suggest for future
components. Ackermann and van den Bogert (2010) showed com- studies to further tune the controller, by broadly testing and imple-
parable peaks in the GRFs when optimising for comparable effort menting different sets of reflexes. For example, anticipatory activa-
and fatigue criteria. Contrarily, they showed that minimising for a tion patterns that act to stabilize the knee in the frontal plane
MusAct criterion contributed to a physiologically consistent early (Besier et al., 2001; Lloyd and Buchanan, 2001) may be under-
stance knee flexion, whereas we did not see this when optimising pinned by ligamento-muscular reflexes (Kim et al., 1995). Addi-
for MusAct. This difference could possibly be explained by a differ- tionally, the musculotendon parameters of the musculoskeletal
ence in walking speed, which was set to 1.1 m/s in the study of Ack- model in our framework were based on generic values and may
ermann and van den Bogert, but was self-selected in our study and not be accurate for our population, possibly explaining the overall
went down to 0.5 m/s for MusAct. bad performance for muscle excitations while performing moder-
Our combined, optimal cost function improved upon previous ate and strong for other biomechanical categories. Previous studies
work using a reflex-based controller in SCONE by showing a better have shown the importance of personalising these musculotendon
agreement with experimental data (current study vs Ong et al., parameters (Lloyd and Besier, 2003) and we would suggest for
2019: R2: 0.72 vs 0.63; RMSE: 2.10 vs 2.47 SD). The main improve- future studies to apply this personalisation, especially when work-
ment was obtained for the knee moment, which lacked an external ing with pathological gait (Davico et al., 2020; Veerkamp et al.,
knee flexion moment in early stance in their results, even though 2019). Additionally, the match with the experimental moments
their framework included minimisation of the head acceleration and powers was generally not as good as for the GRFs and kinemat-
(HeadStab) and of using the knee limit torque (KneeExt), which ics. This could have been affected by the impact artifacts as seen in
we found to be important for obtaining knee flexion in early stance. the experimental data, which are too rapid to be produced by the
Other recent work, using open-loop control with a more complex muscle model. Since this concerns only a small part of the gait
model and a manually-tuned five-term cost function, also found cycle (~3%), effects are expected to be minimal. Likely, it is also
limited knee flexion in early stance after including metabolic caused by only having two foot–ground contact points, limiting
energy rate in the cost function (Falisse et al., 2019). These compar- the path of the centre of pressure, directly affecting the joint
isons underline the importance of not only including multiple cri- moments and thus also joint powers.

9
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Several factors need to be considered to ascertain whether ical what-if questions and to get closer to the ultimate goal of pre-
our weighted, combined cost function is indeed optimal. A more dicting treatment outcomes.
ideal approach to obtain an optimal cost function matching
experimental data is to perform a bilevel optimisation, in which Declaration of Competing Interest
the weightings are optimised such that an optimal agreement
with experimental data is obtained (Mombaur and Truong, The authors declare that they have no known competing finan-
2010; Nguyen et al., 2019). We were limited by computational cial interests or personal relationships that could have appeared
power to do so, as solving for a cost function took around to influence the work reported in this paper.
14 hours using a shooting approach. The use of open loop con-
trol with trajectory optimization methods, such as direct colloca- Acknowledgement
tion, is generally much faster, enabling the use of such bilevel
optimisation. However, those simulations are not intrinsically The authors would like to thank Wouter Schallig for providing
stable, meaning they cannot handle uncertainty (such as pertur- the experimental walking data. This study was supported by
bations and noise) like reflex-based controllers (Groote and Amsterdam Movement Sciences, under the Innovation Call 2018,
Falisse, 2021), limiting their application. Nevertheless, both bile- Tenure Development grant to MK, and by a Queensland Advancing
vel optimisation and the stepwise approach we used could be Clinical Research Fellowship to CC.
considered a fitting process, with the risk of overfitting, espe-
cially given the low dimensionality of human gait (Schwartz
and Rozumalski, 2008). In our stepwise design, when adding a Appendix A
criterion to the cost function, the match with experimental data
could not get worse than the previous result that lacked this cri- The foot–ground contact model detects when the foot hits the
terion (assuming the global minimum was found each time), and ground and calculates the ground reaction forces. The model is
it got close to optimal for our simulation framework. However,
this does not mean that this cost function is necessarily best
in predicting gait in all new conditions. Hence, to be able to
conclude what the best cost function in terms of predictive abil-
ity is, cross-validation on other gait conditions is needed. For
example, in a validation test, a person could walk at different
speeds and on different slopes, or with assistive/resistive devices,
and the best cost function should be able to capture all observa-
tions in the different experimental conditions. Also, before clini-
cal application, it is important to establish whether the used cost
function is valid for a range of pathological conditions.
A shortcoming of our framework is that it is only two-
dimensional and will therefore only be applicable to answer ques-
tions in which gait deviations occur mostly in the sagittal plane. A
three-dimensional analysis would become far more complex, and
other factors, such as medio-lateral stability, as discussed above,
will come into play. Also, more muscles and more control param-
eters will have to be added to the framework, which will consider-
ably slow down optimisation times even further. Another
challenge of our framework is that within predictive simulations
the risk of ending up in local instead of global minima depends
on chance. Therefore, we optimised each scenario for six random
seeds enabling broad exploration. Even though we found differ-
ences between the outcomes of these seeds to be minimal (Appen-
dix E), we cannot be sure that when the number of seeds would
have been increased, possibly lower and slightly different minima
would have been found. However, given the fact that we observed
clear trends in the stepwise combinations, and both the RMSE and
Fig. A1. The resulting ground reaction forces and joint angles from the simulation
R2 clearly improved when combining criteria, we are confident that tracking experimental ground reaction forces and joint angles to optimise the
it would not have affected the interpretation of the results. Lastly, contact sphere geometry.
in this study, the speed and step length differed between simula-
tions with different cost functions, which allows the evaluation
of how these variables change in different conditions in future Table A1
The foot–ground contact model parameters used in the predictive simulations.
work (i.e., in pathology). However, when interpreting the mechan-
ical effects of the different cost functions, it needs to be realized Heel sphere Toe sphere
that the speed and step length can be confounding factors. Contact geometry
In conclusion, this study showed how various optimisation cri- x-position* (m) 0.015 0.00
teria can contribute to different gait features and that careful y-position* (m) 0.019 0.20
z-position (m) 0.005 0.014
weighting of them is essential in predicting healthy gait. Remark-
Radius* (m) 0.040 0.020
ably, a high contribution of CoT to gait only comes to expression Hunt-Crossley force parameters
when combined with other criteria. After developing our frame- Stiffness (N/m) 2,000,000 2,000,000
work incorporating an optimal cost function that is tuned to per- Dissipation (s/m) 1 1
form well in predicting healthy gait, a next step is to validate its Frictions (static, dynamic, viscous) 1 1

predictive ability. This step is important to be able to answer clin- * indicates optimised parameters.

10
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

composed of spheres of specified size and force parameters, which pation, and friction of both spheres were set to pre-determined
are positioned at specified locations on the foot. Previous studies values, based on the values used in previous studies.
used different number of spheres, location of the spheres and the We performed the optimisation with three random initial
force parameters (i.e. stiffness and dampening), which are often guesses and the results from the best simulation (i.e. lowest RMSE)
determined by manual tuning and might therefore be suboptimal were selected, which are shown in Fig. A1. Table A1 provides all
and not consistent for the experimental data. Given recent research parameters for the foot–ground contact model used in the predic-
suggesting that the predicted gait is highly sensitive to changes in tive simulations in this study.
the foot contact model (Millard and Mombaur, 2019), we decided Ground reaction forces joint angles.
to tune the foot–ground contact model in such a way that it can
fit with both experimental kinematics and GRFs as good as possi-
ble. In order to do this, we setup a tracking optimisation within
SCONE, in which the root mean square error (RMSE) between the
predicted and experimental kinematics and GRFs was minimised Appendix B
while the geometries of the spheres (i.e. x-position, y-position,
and radius) were added to the design variables. The stiffness, dissi- There was no EMG available for the iliopsoas and the biceps
femoris short head (see Fig. B1).

Fig. B1. The agreements between predicted and experimental data from each criterion’s independent optimised simulation.

11
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

GRF x: anterior-posterior GRF y: vertical: Rect fem: rectus GRF x: anterior-posterior: GRF y: vertical: Rect fem: rectus
femoris: Glut: gluteus maximus: Bifemsh: biceps femoris short femoris: Glut: gluteus maximus: Bifemsh: biceps femoris short
head: Tib ant: tibialis anterior: Gastroc: gastrocnemius head: Tib ant: tibialis anterior: Gastroc: gastrocnemius.

Appendix C

See There was no experimental data available for the activa- Appendix D
tions of the iliopsoas and the biceps femoris short head (see
Fig. C1). See Fig. D1.

Fig. C1. The agreements between predicted and experimental data produced by the optimised simulations from each stepwise combination of criteria.

12
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Fig. D1. The muscle excitation patterns predicted by the cost function tracking experimental data (ExpTrack) and the optimal, combined cost function.

Appendix E outcome with the lowest score was also the one with the best
agreement with experimental data (Figs. E1 and E2). Over all opti-
There was some variation in the simulation outcomes when six misations, the lowest R2 was 0.63, and the highest RMSE was 2.83
different random seeds were used (Fig. E1), although gait features (Fig. E2).
were generally similar between the different optimisations. The

13
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Fig. E1. The variation between the simulation outcomes of the optimal combined cost function, in which six different random seeds were used. Blue indicates the best score
while red is the worse score.

14
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Fig. E2. The variation in the R2 and RMSE produced by the simulation outcomes of the final, combined cost function, in which six different random seeds were used.

Appendix F. Supplementary data Delp, S.L., Loan, J.P., Hoy, M.G., Zajac, F.E., Topp, E.L., Rosen, J.M., 1990. An Interactive
Graphics-Based Model of the Lower Extremity to Study Orthopaedic Surgical
Procedures. IEEE Trans. Biomed. Eng.
Supplementary data to this article can be found online at Devaprakash, D., Weir, G.J., Dunne, J.J., Alderson, J.A., Donnelly, C.J., 2016. The
https://doi.org/10.1016/j.jbiomech.2021.110530. influence of digital filter type, amplitude normalisation method, and co-
contraction algorithm on clinically relevant surface electromyography data
during clinical movement assessments. J. Electromyogr. Kinesiol. 31, 126–135.
https://doi.org/10.1016/j.jelekin.2016.10.001.
Donelan, J.M., Kram, R., Kuo, A.D., 2001. Mechanical and metabolic determinants of
References the preferred step width in human walking 1985–1992. https://doi.org/
10.1098/rspb.2001.1761
Dorn, T.W., Wang, J.M., Hicks, J.L., Delp, S.L., 2015. Predictive simulation generates
Abram, S.J., Selinger, J.C., Donelan, J.M., 2019. Energy optimization is a major human adaptations during loaded and inclined walking. PLoS ONE 10, 1–16.
objective in the real-time control of step width in human walking. J. Biomech. https://doi.org/10.1371/journal.pone.0121407.
91, 85–91. https://doi.org/10.1016/j.jbiomech.2019.05.010. Duysens, J., Forner-Cordero, A., 2019. A controller perspective on biological gait
Ackermann, M., van den Bogert, A.J., 2010. Optimality principles for model-based control: Reflexes and central pattern generators. Annu. Rev. Control 48, 392–
prediction of human gait. J. Biomech. 43, 1055–1060. https://doi.org/10.1016/j. 400. https://doi.org/10.1016/j.arcontrol.2019.04.004.
jbiomech.2009.12.012. Falisse, A., Serrancolí, G., Dembia, C.L., Gillis, J., Jonkers, I., De Groote, F., 2019. Rapid
Anderson, F.C., Pandy, M.G., 2001. Dynamic Optimization of Human Walking. J. predictive simulations with complex musculoskeletal models suggest that
Biomech. Eng. 123, 381. https://doi.org/10.1115/1.1392310. diverse healthy and pathological human gaits can emerge from similar control
Arnold, E.M., Hamner, S.R., Seth, A., Millard, M., Delp, S.L., 2013. How muscle fiber strategies. J. R. Soc. Interface 16. https://doi.org/10.1098/rsif.2019.0402.
lengths and velocities affect muscle force generation as humans walk and run at Geijtenbeek, T., 2019. SCONE: Open Source Software for Predictive Simulation of
different speeds. J. Exp. Biol. 216, 2150–2160. https://doi.org/ Biological Motion. J. Open Source Softw. 4, 1421. https://doi.org/10.21105/
10.1242/jeb.075697. joss.01421.
Bertram, J.E.A., Ruina, A., 2001. Multiple walking speed-frequency relations are Geyer, H., Herr, H., 2010. A Muscle-reflex model that encodes principles of legged
predicted by constrained optimization. J. Theor. Biol. 209, 445–453. https://doi. mechanics produces human walking dynamics and muscle activities. IEEE
org/10.1006/jtbi.2001.2279. Trans. Neural Syst. Rehabil. Eng. 18, 263–273. https://doi.org/10.1109/
Besier, T.F., Lloyd, D.G., Ackland, T.R., Cochrane, J.L., 2001. Anticipatory effects on TNSRE.2010.2047592.
knee joint loading during running and cutting maneuvers. Med. Sci. Sports Groote, F. De, Falisse, A., 2021. Perspective on musculoskeletal modelling and
Exerc. 33, 1176–1181. https://doi.org/10.1097/00005768-200107000-00015. predictive simulations of human movement to assess the neuromechanics of
Bovi, G., Rabuffetti, M., Mazzoleni, P., Ferrarin, M., 2011. A multiple-task gait gait. Proc. R. Soc. B Biol. Sci. 288, 20202432.
analysis approach: Kinematic, kinetic and EMG reference data for healthy Hair Jr, J.F., Hult, G.T.M., Ringle, C., Sarstedt, M., 2016. A primer on partial least
young and adult subjects. Gait Posture 33, 6–13. https://doi.org/10.1016/ squares structural equation modeling (PLS-SEM). Sage. Publications.
j.gaitpost.2010.08.009. Hansen, N., 2016. The CMA Evolution Strategy: A Tutorial 102, 75–102.
Crowninshield, R.D., Brand, R.A., 1981. A physiologically based criterion of muscle Hunt, K., Crossley, E., 1975. Coefficient of restitution interpreted as damping in
force prediction in locomotion. J. Biomech. 14, 793–801. https://doi.org/ vibroimpact To cite this version : HAL Id : hal-01333795 Coefficient of
10.1016/0021-9290(81)90035-X. Restitution Interpreted as Damping in Vibroimpact. J. Appl. Mech. Am. Soc.
Davico, G., Pizzolato, C., Lloyd, D.G., Obst, S.J., Walsh, H.P.J., Carty, C.P., Health, A., Mech. Eng. 42, 440–445.
Centre, G., Gcore, R.E., Health, M., 2020. Increasing level of Kadaba, M.P., Ramakrishnan, H.K., Wootten, M.E., 1990. Measurement of lower
neuromusculoskeletal model personalisation to investigate joint contact extremity kinematics during level walking. J. Orthop. Res. 8, 397–398. https://
forces in cerebral palsy : A twin case study. Clin. Biomech. 72, 141–149. doi.org/10.1007/978-1-4471-5451-8_100.
https://doi.org/10.1016/j.clinbiomech.2019.12.011. Kandel, E.R., Schwartz, J.H., Jessell, T.M., Siegelbaum, S.A., Hudspeth, A.J., 2013.
Davis, R.B., Õunpuu, S., Tyburski, D., Gage, J.R., 1991. A gait analysis data collection Principles of Neural Science 5th edition.
and reduction technique. Hum. Mov. Sci. 10, 575–587.

15
K. Veerkamp, N.F.J. Waterval, T. Geijtenbeek et al. Journal of Biomechanics 123 (2021) 110530

Kavounoudias, A., Roll, C.A.R., 1998. The plantar sole is a ‘dynamometric map’ for Nyland, J., Brosky, T., Currier, D., Nitz, A., Carborn, D., 1994. Review of the afferent
human balance control 9, 3247–3252. neural system of the knee and its contribution to motor learning. J. Orthop.
Kennedy, P.M., Inglis, J.T., 2002. Distribution and behaviour of glabrous cutaneous Sport. Phys. Ther. 19, 2–11.
receptors in the human foot sole. J. Physiol. 538, 995–1002. https://doi.org/ Ong, C.F., Geijtenbeek, T., Hicks, J.L., Delp, S.L., 2019. Predicting gait adaptations due
10.1013/jphysiol.2001.013087. to ankle plantarflexor muscle weakness and contracture using physics-based
Kim, A.W., Rosen, A.M., Brander, V.A., Buchanan, T.S., 1995. Selective muscle musculoskeletal simulations. PLoS Comput. Biol. 15,. https://doi.org/10.1371/
activation following electrical stimulation of the collateral ligaments of the journal.pcbi.1006993 e1006993.
human knee joint. Arch. Phys. Med. Rehabil. 76, 750. Rajagopal, A., Dembia, C.L., DeMers, M.S., Delp, D.D., Hicks, J.L., Delp, S.L., 2016. Full-
Lai, A., Biewener, A.A., Wakeling, J.M., 2018. Metabolic cost underlies task- body musculoskeletal model for muscle-driven simulation of human gait. IEEE
dependent variations in motor unit recruitment. https://doi.org/10.1098/ Trans. Biomed. Eng. 63, 2068–2079. https://doi.org/10.1109/
rsif.2018.0541. TBME.2016.2586891.
Lin, D.C., McGowan, C.P., Blum, K.P., Ting, L.H., 2019. Yank: The time derivative of Ralston, H.J., 1958. Energy-speed relation and optimal speed during level walking.
force is an important biomechanical variable in sensorimotor systems. J. Exp. Int. Zeitschrift für Angew. Physiol. Einschl. Arbeitsphysiologie 17, 277–283.
Biol. 222. https://doi.org/10.1242/jeb.180414. Schwartz, M.H., Rozumalski, A., 2008. The gait deviation index: a new
Lloyd, D.G., Besier, T.F., 2003. An EMG-driven musculoskeletal model to estimate comprehensive index of gait pathology. Gait Posture 28, 351–357. https://doi.
muscle forces and knee joint moments in vivo. J. Biomech. 36, 765–776. https:// org/10.1016/j.gaitpost.2008.05.001.
doi.org/10.1016/S0021-9290(03)00010-1. Selinger, J.C., O’Connor, S.M., Wong, J.D., Donelan, J.M., 2015. Humans can
Lloyd, D.G., Buchanan, T.S., 2001. 2001_Strategies-of-muscular-support-of-varus- continuously optimize energetic cost during walking. Curr. Biol. 25, 2452–
and-valgus-isometric-loads-at-the-human-knee_Lloyd_Journal-of- 2456. https://doi.org/10.1016/j.cub.2015.08.016.
Biomechanics_10.pdf 34, 1257–1267. Seth, A., Hicks, J.L., Uchida, T.K., Habib, A., Dembia, C.L., Dunne, J.J., Ong, C.F., DeMers,
Markolf, K.L., Gorek, J.F., Kabo, J.M., Shapiro, M.S., 1990. Direct measurement of M.S., Rajagopal, A., Millard, M., Hamner, S.R., Arnold, E.M., Yong, J.R.,
resultant forces in the anterior cruciate ligament. J. Bone Jt. Surg. 72, 557– Lakshmikanth, S.K., Sherman, M.A., Ku, J.P., Delp, S.L., 2018. OpenSim:
567. Simulating musculoskeletal dynamics and neuromuscular control to study
Menz, H.B., Lord, S.R., Fitzpatrick, R.C., 2003. Acceleration patterns of the head and human and animal movement. PLoS Comput. Biol. 14,. https://doi.org/10.3758/
pelvis when walking on level and irregular surfaces. Gait Posture 18, 35–46. BF03326891 e1006223.
Millard, M., Uchida, T., Seth, A., Delp, S.L., 2013. Flexing Computational Muscle : Song, S., Geyer, H., 2015. A neural circuitry that emphasizes spinal feedback
Modeling and Simulation of Musculotendon. Dynamics 135, 1–11. https://doi. generates diverse behaviours of human locomotion. J. Physiol. 593, 3493–3511.
org/10.1115/1.4023390. https://doi.org/10.1113/JP270228.
Minetti, A.E., Moia, C., Roi, G.S., Susta, D., Ferretti, G., Alberto, E., Moia, C., Roi, G.S., Uchida, T.K., Hicks, J.L., Dembia, C.L., Delp, S.L., 2016. Stretching your energetic
Susta, D., Ferretti, G., 2020. Energy cost of walking and running at extreme budget: How tendon compliance affects the metabolic cost of running. PLoS
uphill and downhill slopes 1039–1046. ONE 11, 1–19. https://doi.org/10.1371/journal.pone.0150378.
Mombaur, K., Truong, A., 2010. From human to humanoid locomotion — an inverse Umberger, B.R., Gerritsen, K.G.M., Martin, P.E., 2003. A model of human muscle
optimal control approach 369–383. https://doi.org/10.1007/s10514-009-9170- energy expenditure a model of human muscle energy expenditure. Comput.
7 Methods Biomech. Biomed. Engin. 6, 99–111. https://doi.org/10.1080/
Müller, E.A., Grosse-Lordemann, H., 1937. Der EinfluB der Tretkurbellänge auf das 1025584031000091678.
Arbeitsmaximum und den Wirkungsgrad beim Radfahren. Kaiser Wilhelm- Veerkamp, K., Schallig, W., Harlaar, J., Pizzolato, C., Carty, C.P., Lloyd, D.G., van der
Institut fiir Arbeitsphysiologie. Dortmund-l~Iiinste. Krogt, M.M., 2019. The effects of electromyography-assisted modelling in
Mündermann, A., Dyrby, C.O., Andriacchi, T.P., 2005. Secondary gait changes in estimating musculotendon forces during gait in children with cerebral palsy. J.
patients with medial compartment knee osteoarthritis: Increased load at the Biomech. https://doi.org/10.1016/j.jbiomech.2019.05.026.
ankle, knee, and hip during walking. Arthritis Rheum. 52, 2835–2844. https:// Wang, J.M., Hamner, S.R., Delp, S.L., Koltun, V., 2012. Optimizing locomotion
doi.org/10.1002/art.21262. controllers using biologically-based actuators and objectives. ACM Trans.
Nguyen, V.Q., Johnson, R.T., Sup, F.C., Umberger, B.R., 2019. Bilevel Optimization for Graph. 31, 1–11. https://doi.org/10.1145/2185520.2335376.
Cost Function Determination in Dynamic Simulation of Human Gait. IEEE Trans. Zarrugh, M.Y., Todd, F.N., Ralston, I.J., 1974. Optimization of Energy Expenditure
Neural Syst. Rehabil. Eng. 27, 1–1. https://doi.org/10.1109/tnsre.2019.2922942. during Level Walking 306.
Nigg, B.M., 1985. Biomechanics, Load Analysis and Sports Injuries in the Lower Zimny, M.L., Wink, C.S., 1991. Neuroreceptors in the Tissues of the Knee Joint 1,
Extremities. Sport. Med. An Int. J. Appl. Med. Sci. Sport Exerc. 2, 367–379. 148–157.
https://doi.org/10.2165/00007256-198502050-00005.

16