Introduction to the inverse kinematics of serial

manipulators
Sandipan Bandyopadhyay
Department of Engineering Design
Indian Institute of Technology Madras, Chennai 600 036
Email: [email protected]
Web: http://www.ed.iitm.ac.in/sandipan

Introduction

Position analysis is an essential step in the design, analysis and control of robots. In this article, a basic introduction to the position analysis of serial manipulators is given. This topic
is invariably covered in all the textbooks on this subject. Therefore, instead of repeating
the standard details of forward kinematics, such as, the designation of the reference frames,
determination of the Denavit-Hartenberg (DH) parameters, multiplication of the 4 4 transformation matrices to get the end-effector position and orientation etc., more emphasis is
laid on the inverse problem, which is relatively more complicated in such manipulators. Simple examples, such as a planar 2-R and a spatial 3-R serial robot are discussed in detail.
Related concepts, including the numbers of solutions to the inverse kinematic problem, condition(s) for coincidence of the solution branches and their equivalence with singularities, the
workspace boundaries as the loci of singular configurations etc. are presented in a manner
that once the reader follows these clearly, he/she would find it very easy to apply these
to another manipulator. Multiple approaches to the same problem are included to provide
just a fleeting glance at the richness of this fascinating subject, and to suit the potentially
different tastes of the readers.
As the reader is expected to be familiar with the basic concepts, the rest of the article
would focus more on the actual solution of two representative examples, one planar and
1

another spatial. For further details, the reader is directed to the standard texts in this
domain [1, 2, 3].

Example 1: 2-R planar manipulator

Let us consider one of the simplest possible manipulators in this section, namely, the 2-R
planar serial robot. The robot is shown in Fig. 1. The designation 2-R derives from the

Figure 1: Schematic of the 2-R planar serial robot

fact that the robot has one rotary actuator (i.e., motor) at each of its joints. The problem
of position kinematics (also known as zeroth-order kinematics) can be further divided in two
subproblems: forward, and inverse kinematics. These are discussed in the following.

2.1

Forward kinematics of the planar 2-R manipulator

Forward kinematics refers to the problem of finding the position of the end-effector (in this
case, represented by the point p = (x, y)T in Fig. 1), given the link lengths l1 , l2 , and the
inputs = (1 , 2 )T . For this manipulator, the forward kinematics problem is trivially solved
by observing Fig. 1:
x = l1 cos 1 + l2 cos 12 ,
y = l1 sin 1 + l2 sin 12 ,

(1)
where 12 = 1 + 2 .

(2)

Eqs. (1,2) are also referred to as the forward kinematic map1 , as they map the joint angles (1 , 2 )T to the tip coordinates (x, y)T . As this map is nonlinear in nature, the inverse
of it, i.e., finding the inverse kinematic map, is generally more complicated.

2.2

Inverse kinematics of the planar 2-R manipulator

As explained above, the task in this case is to find 1 , 2 , given x, y. There are multiple ways
of solving this problem, and some of these are noted below.
2.2.1

Trigonometric method

Note that Eqs. (1,2) are nonlinear in 1 , 2 , but these are linear in the sines and cosines
of 1 , 12 . Taking advantage of this, we solve for cos 1 , sin 1 :
cos 1 = (x l2 cos 12 )/l1
sin 1 = (y l2 sin 12 )/l1 ;

(3)
l1 > 0.

(4)

In the above, the sine and cosine of 1 have been taken to be two distinct unknown variables.
Obviously, to complete the solution meaningfully, we need to state the inherent dependence
of these two explicitly:
cos2 1 + sin2 1 1 = 0.

(5)

After performing some simplifications, Eq. (5) reduces to the following form:
a cos 12 + b sin 12 + c = 0,

where

a = 2l2 x, b = 2l2 y, c = l12 l22 x2 y 2 .

(6)
(7)

The manipulations stated above serve one key purpose: they reduce a pair of simultaneous
equations in 1 , 12 to a single equation (6) in a single unknown 12 . However, this equation
requires further processing, so that it can be solved trigonometrically. Recall that it is possible to directly solve only three classes of equations (and their reciprocals) in trigonometry:
cos = d,

(8)

sin = e,

(9)

tan = f,
1

where d, e [1, 1]; f, R.

(10)

It is interesting to note here that the said map is not from R2 to R2 , as it might seem on the surface. If it

is assumed that both the joints allow full-cycle mobility, i.e., there is no physical limit on the joint motions,
then 1 , 2 S1 , and thus (1 , 2 )T T2 S1 S1 .

Therefore, Eq. (6) needs to be transformed to one of these solvable forms. Let us start with
the following step:

a
b
c
cos 12 +
sin 12 +
= 0,
a2 + b2
a2 + b2
a2 + b 2

Further, as seen in Fig. 2,

a
a2 +b2

= cos and

b
a2 +b2

assuming a2 + b2 > 0.

(11)

= sin , for some angle , determined

Figure 2: Geometric interpretation of the angle

uniquely by:
= arctan(b, a),

(12)

where arctan(sin(), cos()) represents the two-argument arctangent function2 . With these,
Eq. (11) reduces to:
cos cos 12 + sin sin 12 =
cos(12 ) =

c
,
a2 + b 2

c
+ b2

a2

assuming a2 + b2 > 0.

(13)
(14)

For the above equation to have a real solution in the remaining unknown 12 , we must ensure:


 c 


 a2 + b2  [0, 1]
a2 + b2 c2 0

(15)

If this condition holds good, then we get:




c
12 = arccos
a2 + b 2


c
12 = arccos
+ arctan(b, a).
a2 + b 2
2

(16)
(17)

The need for such a function as opposed to the common single-argument version of it is explained in

Appendix A.

Thus, there are two solutions for 12 for a given x, y, under the above assumptions.For each
of these solutions, the corresponding values of 1 can be computed from Eqs. (3,4) as:
1 = arctan(y l2 sin 12 , x l2 cos 12 )

(18)

These two distinct pairs of solutions are also known as the branches of inverse kinematics.
This completes the solution of the inverse kinematic problem, in the general case. Let us
now take a look at the special cases, in particular, when a2 + b2 c2 = 0.
After substituting the expressions for a, b, c from Eq. (7) into Eq. (15) and performing
some simplifications, this condition reduces to:
(x2 + y 2 (l1 l2 )2 )(x2 + y 2 (l1 + l2 )2 ) = 0

(19)

It can be seen that condition (19) holds true on the two solid circles shown in Fig. 3. These

Figure 3: Singularities and workspace of the planar 2-R manipulator

circles delimit the set of points in the plane that the manipulator can reach. In other
words, these circles define the boundaries of the workspace (or more precisely, the reachable
workspace) of the manipulator. The reader can readily verify from Eq. (16) the collapse of the
two branches of the solutions of inverse kinematics to one in this case a phenomenon also
called as a singularity. Much more on this can be found in the standard textbooks [1, 2, 3].
Finally, it should be obvious as to what happens when condition (15) is violated. If not, the
reader can find more about it below.

2.2.2

Algebraic method

Methods from algebraic geometry are very popular in the study of position kinematics of
robot manipulators. There is a reasonable introduction to the topic in [2], and the references therein. While these methods can be very powerful, they generally require symbolic
manipulations, i.e., the use of computer algebra systems (CAS) such as Maple, Mathematica,
Singular, etc. As the scope of discussion in this article is very limited, only the treatment
of a single equation would be shown here. The reader can refer to [4] for acquiring the theoretical background in solving systems of multi-variate equations, and find some applications
of these in robotics in [2].
Let us pick up the analysis again at Eq. (6), i.e., after the variable 1 has been eliminated,
leaving behind a univariate equation in 12 . It is well-known, that such an equation can be
converted into a polynomial equation in a single algebraic variable. This is achieved by the
use of two trigonometric identities:
cos 12 =

1 t2
2t
, sin 12 =
,
2
1+t
1 + t2

where t = tan(12 /2).

(20)

Substituting these in Eq. (6), we get after some rationalisation and collection of like terms:
(c a)t2 + 2bt + (c + a) = 0

(21)

Some of the observations made in Section 2.2.1 are obvious from Eq. (21). For instance, it is
obvious that there are two solutions to the problem in the general case. To investigate the
special cases, let us now study the discriminant of the quadratic equation:
= 4(a2 + b2 c2 ).

(22)

The condition for having two distinct real roots is that > 0. It should not be surprising that
this is identical to the condition for the general case described in Section 2.2.1. Likewise, =
0 implies a repeated real root, and < 0 results in a pair of complex conjugate roots,
indicating that the commanded position for the end-effector p(x, y) is beyond the reach
of the manipulator, i.e., outside its workspace. Thus, physically meaningful solutions are
obtained only in the cases 0. From the solutions of Eq. (21) the unknown 12 can be
solved for as:
12 = 2 arctan(t).

(23)

Why the single argument version of arc-tangent function works in this case as opposed to
the two-argument version of it is explained in Appendix B.
6

2.2.3

Geometric method

Consider posing the inverse kinematic problem in a different way, which does not use the
forward kinematic equations as a starting point. In Fig. 1, there are two known points:

p'
1

O
p
1

C1

C2

Figure 4: Geometric interpretation of the two branches of inverse kinematics of the 2-R serial
manipulator
the origin O, and the tip point p. The point p1 , being the tip of link 1, is constrained to
lie on the circle C1 . Likewise, p1 also lies on the circle C2 , being an end point of link 2.
Therefore, the locus of the point p1 (x1 , y1 ) is confined to the intersections of the circles C1 , C2 .
This is depicted in Fig. 4, where the two branches of inverse kinematics are shown as Op1 -p and O-p0 1 -p respectively. As the 2-R is similar to the human hand in architecture,
these configurations are generally termed as elbow-down and elbow-up respectively, the
point p1 (equivalently, p01 ) being the elbow. The mathematical steps for finding the two
configurations follow. The equations of the circles C1 , C2 are given by, respectively:
x21 + y12 l12 = 0

(24)

x21 + y12 + x2 + y 2 2xx1 2yy1 l22 = 0

(25)

Deducting Eq. (25) from Eq. (24), we get a linear equation in x, y, which describes the
common chord L of the two circles passing through the points p1 , p0 1 :
2xx1 + 2yy1 x2 y 2 + l22 l12 = 0.
7

(26)

Solving Eq. (26) for y1 , we get:

y1 =

l12 l22 + x2 + y 2 2xx1

,
2y

assuming y 6= 0.

(27)

Substituting the expression for y in the equation of one of the circles, say, C1 , we get a
quadratic equation in x1 :
a1 x21 + b1 x1 + c1 = 0,

where

(28)

a1 = 4(x2 + y 2 ), b1 = 4x(l12 l22 + x2 + y 2 ),

c1 = (x2 + y 2 )2 + l14 2l12 l22 + l24 + 2l12 x2 2l22 x2 2l12 y 2 2l22 y 2 .
All the results obtained in the other methods, i.e., number of solutions, singularity conditions
etc., can also be derived in this method as well. In addition, another geometric condition for
singularity emerges from this analysis. At a singularity, the circles C1 , C2 are tangential to
each other, the point of tangency being p1 , (which coincides with p01 in this configuration).
Equivalently, the common chord becomes the common tangent. The reader is encouraged to
draw the singular configurations to validate the above surmises.

Example 2: 3-R spatial manipulator

Figure 5: Schematic of the 3-R spatial manipulator

Let us now consider a simple spatial manipulator, namely, the spatial 3-R. The robot is
shown schematically in Fig. 5. In spite of the simplicity, the architecture of this manipulator
8

is important for one reason: it captures the general design of the first three joints and links
of many wrist-decoupled 6-degrees-of-freedom (6-DoF) industrial manipulators, such as the
PUMA, Kuka, Fanuc etc3 . In other words, such 6-DoF manipulators can be designed by
adding a 3-DoF wrist at the point p of the 3-R spatial.

3.1

Forward kinematics of the 3-R spatial manipulator

As in the case of the planar 2-R manipulator, the forward kinematics is fairly trivial. The
reader can find out the DH parameters of the robot, and perform the forward kinematics in
a systematic manner to verify the final results given below:
x = cos 1 (l2 cos 2 + l3 cos 23 )

(29)

y = sin 1 (l2 cos 2 + l3 cos 23 )

(30)

z = l1 + l2 sin 2 + l3 sin 23 ,

3.2

where 23 = 2 + 3 .

(31)

Inverse kinematics of the 3-R spatial manipulator

In this case, there are three unknown inputs: = (1 , 2 , 3 )T . The position vector of the
end-effector, p = (x, y, z)T , is known.
In Eqs. (29,30,31), it can be observed easily that the variable 1 appears only in the
first two, and that these equations are, once again, linear in the sine and cosine of the
angle 1 . Therefore, the same procedure is adopted to solve for the variable 1 (in terms
of the other variables) and to eliminate it from the system of equations, resulting in the
following equations:
1 = arctan(y/r, x/r),

where r =

p
x2 + y 2 is assumed to be non-zero;

l2 cos 2 + l3 cos 23 r = 0.

(32)
(33)

The two equations, (31, 33) are simultaneous linear equations in the sines and cosines of the
angles 2 , 23 . As such, we treat these equations just as above, to solve for, and eliminate
one of the two variables, to finally arrive at a univariate equation which is our eventual
objective.
3

Readers interested in the inverse kinematics of these 6-DoF manipulators may refer to a standard text-

book, e.g., [2].

Choosing to solve for 2 first, we get:

2 = arctan(z l1 l3 sin 23 ), r l3 cos 23 ),

and

2l3 r cos 23 + 2l3 (z l1 ) sin 23 r2 + l22 l32 (z l1 )2 = 0

(34)
(35)

We have already seen how to solve equations such as (35) in Section 2.2.1. Solving it, we
get real and distinct solutions when the point p is inside the workspace, a repeated real
root when the point is on the boundary of the workspace (i.e., when the manipulator is
singular), and a pair of complex conjugates when p is outside the workspace. Following the
2-R example, the reader can easily derive the conditions for the singularities and boundaries
of the workspace. It is obvious that one of these conditions is already known, since the
links 2, 3 form a planar 2-R in the vertical plane. Naturally, this condition is independent of
the angle 1 . The reader can verify easily that the other condition is also independent of 1 .
Understanding why it is so is left as an exercise for the reader. The reader is also encouraged
to draw the workspace boundaries, as well as the singular configurations of the manipulator
for greater understanding of these concepts.

Summary

In this brief article, emphasis is laid on the actual computation of forward and more importantly, inverse kinematics of serial robots. Detailed theoretical treatment of these topics
are available in many textbooks, some of which have been mentioned in the references. It
is hoped that the illustrative examples and the focus on the underlying trigonometric, algebraic, and geometric concepts would nicely complement the basic theories offered in these
books. Treatment of more complicated manipulators, e.g., 6-DoF robots such as the PUMA
560, is beyond the scope of the present discussion. However, understanding the concepts
described in this article should help the reader to follow the analysis of such manipulators.

References
[1] J. J. Craig, Introduction to Robotics: Mechanics and Control. California: Addison Wesley,
second ed., 1986.
[2] A. Ghosal, Robotics: Fundamental Concepts and Analysis. New Delhi: Oxford University
Press, 2006.
10

[3] S. K. Saha, Introduction to Robotics. New Delhi: Tata McGraw-Hill, 2008.

[4] D. Cox, J. Little, and D. OShea, Ideals, Varieties, and Algorithms: An Introduction
to Computational Algebraic Geometry and Commutative Algebra. New York: SpringerVerlag, 1991.

The two-argument arctangent function

Figure 6: The utility of the arctan(sin(), cos()) function

Fig. 6 illustrates why arcsin(), arccos(), or even arctan() were not used in this article
instead of arctan(sin(), cos()). Clearly, arcsin fails to distinguish between points p1 , p2 ,
arccos between p1 , p4 , and arctan between p1 , p3 , while arctan(sin(), cos()) alone identifies
each point distinctly.

Bijectivity of the tangent-half-angle function

As seen in Fig. 7, the function tan(/2) is bijective in the range [, ]. Thus, if t =

tan(/2), then = 2 arctan(t ) always returns an unambiguous result.

11

tan J

10
0
-5
-10
- 3 - 2 -1 0
1
Hin radian sL

Figure 7: The plot of tan(/2) vs. [, ]

12