Marcos Duarte

Laboratory of Biomechanics and Motor Control (http://demotu.org/)

Federal University of ABC, Brazil

With respect to the equations of motion to determine the dynamics of a system, there are two general approaches: forward (or direct) and inverse dynamics. For example, consider the solution of Newton's second law for a particle. If we know the force(s) and want to find the trajectory, this is **forward dynamics**. If instead, we know the trajectory and want to find the force(s), this is **inverse dynamics**:

Let's estimate the joint force and moments of force at the lower limb during locomotion using the inverse dynamics approach.

We will model the lower limbs at the right side as composed by three rigid bodies (foot, leg, and thigh) articulated by three hinge joints (ankle, knee, and hip) and perform a two-dimensional analysis.

The free body diagrams of the lower limbs are:

The equilibrium equations for the forces and moments of force around the center of mass are:

For body 1 (foot):

$$ \begin{array}{l l} \mathbf{F_1} + m_1\mathbf{g} + \mathbf{GRF} = m_1\mathbf{a_1} \\ \\ \mathbf{M_1} + \mathbf{r_{cmp1}}\times\mathbf{F_1} + \mathbf{r_{cmCOP}}\times\mathbf{GRF} = I_1\mathbf{\alpha_1} \end{array} $$For body 2 (leg):

$$ \begin{array}{l l} \mathbf{F_2} + m_2\mathbf{g} - \mathbf{F_1} = m_2\mathbf{a_2} \\ \\ \mathbf{M_2} + \mathbf{r_{cmp2}}\times\mathbf{F_2} + \mathbf{r_{cmd2}}\times-\mathbf{F_{1}} - \mathbf{M_1} = I_2\mathbf{\alpha_2} \end{array} $$For body 3 (thigh):

$$ \begin{array}{l l} \mathbf{F_3} + m_3\mathbf{g} - \mathbf{F_2} = m_3\mathbf{a_3} \\ \\ \mathbf{M_3} + \mathbf{r_{cmp3}}\times\mathbf{F_3} + \mathbf{r_{cmd3}}\times-\mathbf{F_{2}} - \mathbf{M_2} = I_3\mathbf{\alpha_3} \end{array} $$Where **p** and **d** stands for proximal and distal joints (with respect to the fixed extremity), $\mathbf{r_{cm\:ji}}$ is the position vector from the center of mass of body **i** to the joint **j**, **COP** is the position of application of the ground reaction force (GRF), $\mathbf{\alpha}$ is the angular acceleration, and g= -9.8 m/s$^2.$

Note that the pattern of the equations is the same for the three segments: distal and proximal forces and moments of force and the weight force are present in all segments.

The only exception is with the foot in contact with the ground. As the ground only pushes the foot, it can not generate a moment of force over the foot. Because of that we model the interaction foot-ground as a resultant ground reaction force (GRF) applied on the foot at the center of pressure (COP) position.

Both GRF and COP quantities are measured with a force platform and are assumed as known quantities.

Because of that the system of equations above is only solvable if we start by the body 1, from bottom to top.

The system of equations above is simple and straightforward to solve, it is just a matter of being systematic.

We start by segment 1, find $\mathbf{F_1}$ and $\mathbf{M_1}$, substitute these values on the equations for segment 2, find $\mathbf{F_2}$ and $\mathbf{M_2}$, substitute them in the equations for segment 3 and find $\mathbf{F_3}$ and $\mathbf{M_3}\:$:

For body 1 (foot):

$$ \begin{array}{l l} \mathbf{F_1} = m_1\mathbf{a_1} - m_1\mathbf{g} - \mathbf{GRF} \\ \\ \mathbf{M_1} = I_1\mathbf{\alpha_1} - \mathbf{r_{cmp1}}\times\left(m_1\mathbf{a_1} - m_1\mathbf{g} - \mathbf{GRF}\right) - \mathbf{r_{cmCOP}}\times\mathbf{GRF} \end{array} $$For body 2 (leg):

$$ \begin{array}{l l} \mathbf{F_2} = m_1\mathbf{a_1} + m_2\mathbf{a_2} - (m_1+m_2)\mathbf{g} - \mathbf{GRF} \\ \\ \mathbf{M_2} = I_1\mathbf{\alpha_1} + I_2\mathbf{\alpha_2} - \mathbf{r_{cmp2}}\times\left(m_1\mathbf{a_1} + m_2\mathbf{a_2} - (m_1+m_2)\mathbf{g} - \mathbf{GRF}\right) + \mathbf{r_{cgd2}}\times\left(m_1\mathbf{a_1} - m_1\mathbf{g} - \mathbf{GRF}\right) - \mathbf{r_{cmp1}}\times\left(m_1\mathbf{a_1} - m_1\mathbf{g} - \mathbf{GRF}\right) - \mathbf{r_{cmCOP}}\times\mathbf{GRF} \end{array} $$For body 3 (thigh):

$$ \begin{array}{l l} \mathbf{F_3} = m_1\mathbf{a_1} + m_2\mathbf{a_2} + m_3\mathbf{a_3} - (m_1+m_2+m_3)\mathbf{g} - \mathbf{GRF} \\ \\ \mathbf{M_3} = I_1\mathbf{\alpha_1} + I_2\mathbf{\alpha_2} + I_3\mathbf{\alpha_3} - \mathbf{r_{cmp3}}\times\left(m_1\mathbf{a_1} + m_2\mathbf{a_2} + m_3\mathbf{a_3} - (m_1+m_2+m_3)\mathbf{g} - \mathbf{GRF}\right) + \mathbf{r_{cmd3}}\times\left(m_1\mathbf{a_1} + m_2\mathbf{a_2} - (m_1+m_2)\mathbf{g} - \mathbf{GRF}\right) - \mathbf{r_{cmp2}}\times\left(m_1\mathbf{a_1} + m_2\mathbf{a_2} - (m_1+m_2)\mathbf{g} - \mathbf{GRF}\right) + \mathbf{r_{cmd2}}\times\left(m_1\mathbf{a_1} - m_1\mathbf{g} - \mathbf{GRF}\right) - \mathbf{r_{cmp1}}\times\left(m_1\mathbf{a_1} - m_1\mathbf{g} - \mathbf{GRF}\right) - \mathbf{r_{cmCOP}}\times\mathbf{GRF} \end{array} $$The calculation above is tedious, error prone, useless, and probably it's wrong.

To make some use of it, we can clearly see that forces act on far segments, which are not directly in contact with these forces. In fact, this is true for all stuff happening on a segment: note that $\mathbf{F_1}$ and $\mathbf{M_1}$ are present in the expression for $\mathbf{F_3}$ and $\mathbf{M_3}$ and that the acceleration of segment 1 matters for the calculations of segment 3.

Instead, we can use the power of computer programming (like this one right now!) and solve these equations recursively hence they have the same pattern.

We could write a function that it would have as inputs the body-segment parameters, the kinematic data, and the distal joint force and moment of force and output the proximal joint force and moment of force.

Then, we would call this function for each segment, starting with the segment that has a free extremity or that has the force and moment of force measured by some instrument (i,e, use a force plate for the foot-ground interface).

This function would be called in the following manner:

```
Fp, Mp = invdyn2d(rcm, rd, rp, acm, alfa, mass, Icm, Fd, Md)
```

So, let's write such function:

In [ ]:

```
# %load ./../functions/invdyn2d.py
"""Two-dimensional inverse-dynamics calculations of one segment."""
__author__ = 'Marcos Duarte, https://github.com/demotu/BMC'
__version__ = 'invdyn2d.py v.2 2015/11/13'
def invdyn2d(rcm, rd, rp, acm, alpha, mass, Icm, Fd, Md):
"""Two-dimensional inverse-dynamics calculations of one segment
Parameters
----------
rcm : array_like [x,y]
center of mass position (y is vertical)
rd : array_like [x,y]
distal joint position
rp : array_like [x,y]
proximal joint position
acm : array_like [x,y]
center of mass acceleration
alpha : array_like [x,y]
segment angular acceleration
mass : number
mass of the segment
Icm : number
rotational inertia around the center of mass of the segment
Fd : array_like [x,y]
force on the distal joint of the segment
Md : array_like [x,y]
moment of force on the distal joint of the segment
Returns
-------
Fp : array_like [x,y]
force on the proximal joint of the segment (y is vertical)
Mp : array_like [x,y]
moment of force on the proximal joint of the segment
Notes
-----
To use this function recursevely, the outputs [Fp, Mp] must be inputed as
[-Fp, -Mp] on the next call to represent [Fd, Md] on the distal joint of the
next segment (action-reaction).
This code was inspired by a similar code written by Ton van den Bogert [1]_.
See this notebook [2]_.
References
----------
.. [1] http://isbweb.org/data/invdyn/index.html
.. [2] http://nbviewer.ipython.org/github/demotu/BMC/blob/master/notebooks/GaitAnalysis2D.ipynb
"""
from numpy import cross
g = 9.80665 # m/s2, standard acceleration of free fall (ISO 80000-3:2006)
# Force and moment of force on the proximal joint
Fp = mass*acm - Fd - [0, -g*mass]
Mp = Icm*alpha - Md - cross(rd-rcm, Fd) - cross(rp-rcm, Fp)
return Fp, Mp
```

The calculations are implemented in only two lines of code at the end, the first part of the code is the help on how to use the function. The help is long because it's supposed to be helpful :), see the style guide for NumPy/SciPy documentation.

The real problem is to measure or estimate the experimental variables: the body-segment parameters, the ground reaction forces, and the kinematics of each segment. For such, it is necessary some expensive equipments, but they are typical of a biomechanics laboratory.

Let's work with some data of kinematic position of the segments and ground reaction forces in order to compute the joint forces and moments of force.

The data we will work are in fact from a computer simulation of running created by Ton van den Bogert. The nice thing about these data is that as a simulation, the true joint forces and moments of force are known and we will be able to compare our estimation with these true values.

All the data can be downloaded from a page at the ISB website:

In [2]:

```
from IPython.display import IFrame
IFrame('http://isbweb.org/data/invdyn/index.html', width='100%', height=400)
```

Out[2]:

In [1]:

```
# import the necessary libraries
import numpy as np
%matplotlib notebook
import matplotlib.pyplot as plt
import sys
sys.path.insert(1, r'./../functions')
```

In [2]:

```
# load file with ground reaction force data
grf = np.loadtxt('./../data/all.frc') # [Fx, Fy, COPx]
# load file with kinematic data
kin = np.loadtxt('./../data/all.kin') # [Hip(x,y), knee(x,y), ankle(x,y), toe(x,y)]
freq = 10000
time = np.linspace(0, grf.shape[0]/freq, grf.shape[0])
```

In [3]:

```
# this is simulated data with no noise, filtering doesn't matter
if False:
# filter data
from scipy.signal import butter, filtfilt
# Butterworth filter
b, a = butter(2, (100/(freq/2)))
for col in np.arange(grf.shape[1]-1):
grf[:, col] = filtfilt(b, a, grf[:, col])
b, a = butter(2, (100/(freq/2)))
for col in np.arange(kin.shape[1]):
kin[:, col] = filtfilt(b, a, kin[:, col])
```

In [4]:

```
# heel strike occurs at sample 3001
time = time[3001 - int(freq/40):-int(freq/20)]
grf, kin = grf[3001 - int(freq/40):-int(freq/20), :], kin[3001 - int(freq/40):-int(freq/20), :]
```

In [8]:

```
# plot data
hfig, hax = plt.subplots(2, 2, sharex = True, squeeze=True, figsize=(9, 5))
hax[0, 0].plot(time, grf[:, [0, 1]], linewidth=2)
hax[0, 0].legend(('Fx','Fy'), frameon=False)
hax[0, 0].set_ylabel('Force [N]')
hax[0, 1].plot(time, grf[:, 2], linewidth=2)
hax[0, 1].legend(['COPx'], frameon=False)
hax[0, 1].set_ylabel('Amplitude [m]')
hax[1, 0].plot(time, kin[:, 0::2], linewidth=2)
hax[1, 0].legend(('Hip x','Knee x','Ankle x','Toe x'), frameon=False)
hax[1, 0].set_ylabel('Amplitude [m]')
hax[1, 1].plot(time, kin[:, 1::2], linewidth=2)
hax[1, 1].legend(('Hip y','Knee y','Ankle y','Toe y'), frameon=False)
hax[1, 1].set_ylabel('Amplitude [m]')
hax[1, 0].set_xlabel('Time [s]'), hax[1, 1].set_xlabel('Time [s]')
plt.tight_layout()
plt.show()
```

In [6]:

```
# body-segment parameters [thigh, shank, foot]
mass = [6.85, 2.86, 1.00] # mass [kg]
Icm = [0.145361267, 0.042996389, 0.0200] # rotational inertia [kg.m2]
cmpr = [0.4323725, 0.4334975, 0.0] # CM [m] wrt. prox. joint [frac. of segment length]
```

In [7]:

```
# Kinematic data
# center of mass position of the thigh, shank, foot segments
rcm = np.hstack((kin[:, (0, 1)] + cmpr[0]*(kin[:, (2, 3)] - kin[:, (0, 1)]),
kin[:, (2, 3)] + cmpr[1]*(kin[:, (4, 5)] - kin[:, (2, 3)]),
kin[:, (4, 5)] + cmpr[2]*(kin[:, (6, 7)] - kin[:, (4, 5)])))
# center of mass linear acceleration of the thigh, shank, foot segments
acm = np.diff(rcm, n=2, axis=0)*freq*freq
acm = np.vstack((acm, acm[-1, :], acm[-1, :]))
# thigh, shank, foot segment angle
ang = np.vstack((np.arctan2(kin[:, 1] - kin[:, 3], kin[:, 0] - kin[:, 2]),
np.arctan2(kin[:, 3] - kin[:, 5], kin[:, 2] - kin[:, 4]),
np.arctan2(kin[:, 5] - kin[:, 7], kin[:, 4] - kin[:, 6]))).T
# hip, knee, and ankle joint angles
angj = np.vstack((-(ang[:, 0]-ang[:, 1]), np.unwrap(ang[:, 1] - ang[:, 2] + np.pi/2))).T*180/np.pi
# thigh, shank, foot segment angular acceleration
aang = np.diff(ang, n=2, axis=0)*freq*freq
aang = np.vstack((aang, aang[-1, :], aang[-1, :]))
```

In [10]:

```
# plot hip, knee, and ankle joint angles
hfig, (hax1, hax2) = plt.subplots(2, 1, sharex = True, squeeze=True, figsize=(9, 4))
hax1.plot(time, angj[:, 0], linewidth=2, label='Knee')
hax1.legend(frameon=False, loc='upper left'), hax1.grid()
hax2.plot(time, angj[:, 1], linewidth=2, label='Ankle')
hax2.legend(frameon=False, loc='upper left'), hax2.grid()
hax1.set_ylabel('Joint angle $[^o]$')
hax2.set_ylabel('Joint angle $[^o]$')
hax2.set_xlabel('Time [s]')
plt.tight_layout()
plt.show()
```

In [11]:

```
# inverse dynamics
# invdyn2d(rcm, rd, rp, acm, alpha, mass, Icm, Fd, Md)
from invdyn2d import invdyn2d
# ankle
[Fa, Ma] = invdyn2d(rcm[:, (4, 5)], grf[:, (2, 2)]*[1, 0], kin[:, (4, 5)], acm[:, (4, 5)], \
aang[:, 2], mass[2], Icm[2], grf[:, (0, 1)], 0)
# knee
[Fk, Mk] = invdyn2d(rcm[:,(2, 3)], kin[:, (4, 5)], kin[:, (2, 3)], acm[:, (2, 3)], \
aang[:,1], mass[1], Icm[1], -Fa, -Ma)
# hip
[Fh, Mh] = invdyn2d(rcm[:, (0, 1)], kin[:, (2, 3)], kin[:, (0, 1)], acm[:, (0, 1)], \
aang[:, 0], mass[0], Icm[0], -Fk, -Mk)
# magnitude of the calculated hip, knee, and ankle resultant joint force
Fam = np.sqrt(np.sum(np.abs(Fa)**2, axis=-1))
Fkm = np.sqrt(np.sum(np.abs(Fk)**2, axis=-1))
Fhm = np.sqrt(np.sum(np.abs(Fh)**2, axis=-1))
```

In [12]:

```
# load file with true joint forces and moments of force
forces = np.loadtxt('./../data/all.fmg') # [Hip, knee, ankle]
moments = np.loadtxt('./../data/all.mom') # [Hip, knee, ankle]
#heel strike occurs at sample 3001
forces, moments = forces[3001-int(freq/40):-int(freq/20), :], moments[3001-int(freq/40):-int(freq/20), :]
```

Let's plot these data but because later we will need to plot similar plots, let's create a function for the plot to avoid repetition of code:

In [15]:

```
def plotdata(time, Fh, Fk, Fa, Mh, Mk, Ma, forces, moments, stitle):
# plot hip, knee, and ankle moments of force
hfig, hax = plt.subplots(3, 2, sharex = True, squeeze=True, figsize=(9, 5))
# forces
hax[0, 0].plot(time, Fh, label='invdyn'), hax[0, 0].set_title('Hip')
hax[1, 0].plot(time, Fk), hax[1, 0].set_title('Knee')
hax[2, 0].plot(time, Fa), hax[2, 0].set_title('Ankle')
hax[1, 0].set_ylabel('Joint force [N]')
hax[2, 0].set_xlabel('Time [s]')
# moments of force
hax[0, 1].plot(time, Mh), hax[0, 1].set_title('Hip')
hax[1, 1].plot(time, Mk), hax[1, 1].set_title('Knee')
hax[2, 1].plot(time, Ma), hax[2, 1].set_title('Ankle')
hax[1, 1].set_ylabel('Moment of Force [Nm]')
hax[2, 1].set_xlabel('Time [s]')
# true joint forces and moments of force
hax[0, 0].plot(time, forces[:, 0], 'r--', label='True'), hax[0, 0].legend(frameon=False)
hax[1, 0].plot(time, forces[:, 1], 'r--')
hax[2, 0].plot(time, forces[:, 2], 'r--')
hax[0, 1].plot(time, moments[:, 0], 'r--')
hax[1, 1].plot(time, moments[:, 1], 'r--')
hax[2, 1].plot(time, moments[:, 2], 'r--')
plt.suptitle(stitle, fontsize=16)
for x in hax.flat:
x.locator_params(nbins=5); x.grid()
plt.show()
plotdata(time, Fhm, Fkm, Fam, Mh, Mk, Ma, forces, moments,
'Inverse dynamics: estimated versus true values')
```

The results are very similar; only a small part of the moments of force is different because of some noise.

Let's see what happens with the joint forces and moments of force when we neglect the contribution of some terms in the inverse dynamics analysis of these data.

Consider the case where the segment acceleration is neglected:

In [16]:

```
# ankle
[Fast, Mast] = invdyn2d(rcm[:, (4, 5)], grf[:, (2, 2)]*[1, 0], kin[:,(4, 5)], acm[:, (4, 5)]*0, \
aang[:,2]*0, mass[2], Icm[2], grf[:, (0, 1)], 0)
# knee
[Fkst, Mkst] = invdyn2d(rcm[:, (2, 3)], kin[:,(4, 5)], kin[:, (2, 3)], acm[:, (2, 3)]*0, \
aang[:, 1]*0, mass[1], Icm[1], -Fast, -Mast)
# hip
[Fhst, Mhst] = invdyn2d(rcm[:, (0, 1)], kin[:,(2, 3)], kin[:, (0, 1)], acm[:, (0, 1)]*0, \
aang[:,0]*0, mass[0], Icm[0], -Fkst, -Mkst)
# magnitude of the calculated hip, knee, and ankle resultant joint force
Fastm = np.sqrt(np.sum(np.abs(Fast)**2, axis=-1))
Fkstm = np.sqrt(np.sum(np.abs(Fkst)**2, axis=-1))
Fhstm = np.sqrt(np.sum(np.abs(Fhst)**2, axis=-1))
plotdata(time, Fhstm, Fkstm, Fastm, Mhst, Mkst, Mast, forces, moments,
'Inverse dynamics: quasis-static approach versus true values')
```