Robot Trajectory Planning Using Multi-objective

Genetic Algorithm Optimization

E.J. Solteiro Pires1 , J.A. Tenreiro Machado2 , and P.B. de Moura Oliveira1
1
Universidade de Trás-os-Montes e Alto Douro, Dep. de Engenharia Electrotécnica,
Quinta de Prados, 5000–911 Vila Real, Portugal,
{epires,oliveira}@utad.pt, http://www.utad.pt/˜epires
http://www.utad.pt/˜oliveira
2
Instituto Superior de Engenharia do Porto, Dep. de Engenharia Electrotécnica,
Rua Dr. António Bernadino de Almeida, 4200-072 Porto, Portugal
[email protected], http://www.dee.isep.ipp.pt/˜jtm

Abstract. Generating manipulator trajectories considering multiple objectives

and obstacle avoidance is a non trivial optimization problem. In this paper a multi-
objective genetic algorithm is proposed to address this problem. Multiple criteria
are optimized up to five simultaneous objectives. Simulations results are presen-
ted for robots with two and three degrees of freedom, considering two and five
objectives optimization. A subsequent analysis of the solutions distribution along
the converged non-dominated Pareto front is carried out, in terms of the achieved
diversity.

1 Introduction
In the last twenty years genetic algorithms (GAs) have been applied in a plethora of
fields such as: control, system identification, robotics, planning and scheduling, image
processing, pattern recognition and speech recognition [1]. This paper addresses the
planning of trajectories, meaning the development of an algorithm to find a continuous
motion that takes the manipulator from a given starting configuration to a desired end
position in the workspace without colliding with any obstacle.
Several single-objective methods for trajectory planning, collision avoidance and
manipulator structure definition have been proposed. A possible approach consists in
adopting the differential inverse kinematics, using the Jacobian matrix, for generating
the manipulator trajectories [2,3]. However, this algorithm must take into account the
kinematic singularities that may be hard to tackle. To avoid this problem, other algorithms
for the trajectory generation are based on the direct kinematics [4,5,6,7,8].
Chen and Zalzala [2] propose a GA method to generate the position and the configu-
ration of a mobile manipulator. In this report the inverse kinematics scheme is applied
to optimize the least torque norm, the manipulability, the torque distribution and the
obstacle avoidance. Davidor [3] also applies GAs to the trajectory generation by sear-
ching the inverse kinematics solutions to pre-defined end effector robot paths. Kubota
et al. [4] study a hierarchical trajectory planning method for a redundant manipulator
with a virus-evolutionary GA, running simultaneously two processes. One process cal-
culates some manipulator collision-free positions and the other generates a collision free

K. Deb et al. (Eds.): GECCO 2004, LNCS 3102, pp. 615–626, 2004.


c Springer-Verlag Berlin Heidelberg 2004
616 E.J. Solteiro Pires, J.A. Tenreiro Machado, and P.B. de Moura Oliveira

trajectory by combining these intermediate positions. Rana and Zalzala [5] develope a
method to plan a near time-optimal, collision-free, motion in the case of multi-arm ma-
nipulators. The planning is carried out in the joint space and the path is represented as a
string of via-points connected through cubic splines. Chocron and Bidaud [9] propose an
evolutionary algorithm to perform a task-based design of modular robotic systems. The
system consists in a mobile base and an arm that may be built with serially assembled
links and joints modules. The optimization design is evaluated with geometric and kine-
matic performance measures. Kim and Khosha [10] present the design of a manipulator
that is best suited for a given task. The design consists of determining the trajectory and
the length of a three degrees of freedom (dof ) manipulator. Han et al. [11] describe a
design method of a modular manipulator that uses the kinematic equations to determine
the robot configuration and, in a second phase, adopts a GA to find the optimal length.
Gacôgne [12] presents a problem involving obstacle avoidance. He looks for an
emergence of rules system for a mobile robot to have a good road-holding behavior in
different playgrounds. A multi-objective genetic algorithm is used to find a short and
readable solutions for every concrete problem.
Multi-objective techniques using GAs have been increasing in relevance as a rese-
arch area. In 1989, Goldberg [13] suggested the use of a GA to solve multi-objective
problems and since then other investigators have been developing new methods, such
as multi-objective genetic algorithm (MOGA) [14], non-dominated sorted genetic al-
gorithm (NSGA) [15] and niched Pareto genetic algorithm (NPGA) [16], among many
other variants [17].
In this line of thought, this paper proposes the use of a multi-objective method to
optimize a manipulator trajectory. This method is based on a GA adopting direct kine-
matics. The optimal manipulator front is the one that minimizes the objectives without
any collision with the obstacles in the workspace. Following this introduction, the paper
is organized as follows: section 2 formulates the problem and the GA-based method for
its resolution. Section 3 presents several simulations results involving different robots,
objectives and workspace settings. Finally, section 4 outlines the main conclusions.

2 Problem and Algorithm Formulation

This study considers robotic manipulators that are required to move from an initial point
up to a given final configuration. Two and three dof planar manipulators (i.e. 2R and 3R
robots) are used in the experiments with link lengths of one meter and rotational joints
which are free to rotate 2π rad. To test a possible manipulator/obstacle collision, the
arm structure is analyzed in order to verify if it is inside of any obstacle. The trajectory
consists in a set of strings representing the joint positions between the initial and final
robot configurations.

2.1 Representation

The path for a iR manipulator (i = 2, 3), at generation T , is directly encoded as vectors

in the joint space to be used by the GA as:
 Robot Trajectory Planning Using Multi-objective Genetic Algorithm Optimization 617

(∆t,T ) (∆t,T ) (2∆t,T ) (2∆t,T ) ((n−2)∆t,T ) ((n−2)∆t,T )

[{q1 , .., qi }, {q1 , .., qi }, .., {q1 , .., qi }] (1)

where i is the number of dof and ∆t the sampling time between two consecutive confi-
gurations.
(j∆t,0)
The joints values ql (j = 1, . . . , n − 2; l = 1, . . . , i) are randomly initialized
in the range ] − π, +π] rad. It should be noted that the initial and final configurations
have not been encoded into the string because they remain unchanged throughout the
trajectory search. Without losing generality, for simplicity, it is adopted a normalized
time of ∆t = 0.1 sec, because it is always possible to perform a time re-scaling.

2.2 Operators in the Multi-objective Genetic Algorithm

The initial population of strings is randomly generated. The search is then carried out
among this population. Three different operators are used in the genetic planning: sel-
ection, crossover and mutation, as described in the sequel.
In what concerns the selection operator, the successive generations of new strings
are reproduced on the basis of a Pareto ranking [13] with σshare = 0.01 and α = 2.
To promote population diversity a metric count is used. This metric uses all solutions
in the population independently of their rank to evaluate every fitness function. For the
crossover operator it is used the simulated binary crossover (SBX)[15]. After crossover,
the best solutions (among both parents and children) are chosen to form the next popu-
lation. The mutation operator replaces one gene value with a given probability using the
equation:

(j∆t,T +1) (j∆t,T )

√
qi = qi + N (0, 1/ 2π) (2)

at generation T , where N (µ, σ) is the normal distribution function with average µ and
standard deviation σ.

2.3 Evolution Criteria

Five indices {q, q̇, p, ṗ, Ea } (3) are used to qualify the evolving trajectory robotic ma-
nipulators. These criteria are minimized by the planner to find the optimal Pareto front.
((j+1)∆t,T ) (j∆t,T )
Before evaluating any solution all the values such that |qi − qi | > π are
readjusted, adding or removing a multiple value of 2π, in the strings.

i 
n


2
(j∆t,T )
q= q̇l (3a)
j=1 l=1

i 
n


2
(j∆t,T )
q̇ = q̈l (3b)
j=1 l=1
618 E.J. Solteiro Pires, J.A. Tenreiro Machado, and P.B. de Moura Oliveira

Table 1. Fronts parameters statistics

Pareto front Local front

κ α β κ α β
Median 13.46 −8.32 −10.77 19.23 49.28 −13.02
Average 13.45 −7.40 −9.95 19.18 49.48 −13.19
Standard deviation 0.37 2.71 1.82 0.30 3.65 0.76

n

2
p= d (pj , pj−1 ) (3c)
j=2

n

2
ṗ = {d (pj , pj−1 ) − d (pj−1 , pj−2 )} (3d)
j=3

n


i
(j∆t,T )
Ea = (n − 1)∆t Pa = |τl .∆ql | (3e)
j=1 l=1

The joint distance q (3a) is used to minimize the manipulator joints travelling di-
stance. For a function y = g(x) the curve length is defined by:

∫ [1 + (dg/dx)2 ]dx (4)

and, consequently, to minimize the curve length distance is adopted the simplified ex-
pression:

∫ (dg/dx)2 dx = ∫ ġ 2 dx. (5)

The joint velocity q̇ is used to minimize the ripple in time evolution. The cartesian
distance p (3c) minimizes the total arm trajectory length, from the initial point up to
the final point, where pj is the robot j intermediate arm cartesian position and d(·, ·)
is a function that gives the distance between two arguments. The cartesian velocity is
responsible for reducing the ripple in arm time evolution. Finally, the energy Ea in
expression (3e), where τl are the robot joint torques, is computed assuming that power
regeneration is not available by motors doing negative work, that is, by taking the absolute
value of the power [18].

3 Simulation Results
In this section results of various experiments are presented. In this line of thought,
subsections 3.1 and 3.2 present the trajectory optimization for the 2R and 3R robots
respectively, for two objectives (2D). Finally, subsection 3.3 shows the results of a five
dimensional (5D) optimization for a 2R robot.
 Robot Trajectory Planning Using Multi-objective Genetic Algorithm Optimization 619

21 a 80

20 a
70
19
60
18
50
ṗ

ṗ
17
40
16

15 30
b b
14 20
10 20 30 40 50 60 70 0 50 100 150 200 250 300 350
q̇ q̇

(a) Pareto optimal front (b) local optimal front

Fig. 1. Optimal fronts for the 2R robot

ṗ(q̇)
r1
r2
ṗ
r3
rM

q̇

Fig. 2. Normal straight lines to the front obtained with p(q) function

3.1 2R Robot Trajectory with 2D Optimization

The experiments consist on moving a 2R robotic arm from the starting configuration,
defined by the joint coordinates A ≡ {−1.149, 1.808} rad, up to the final configuration,
defined by B ≡ {1.181, 1.466} rad, in a workspace without obstacles. The objectives
used in this section to optimize are the joint velocity q̇ (3b) and the cartesian velocity ṗ
(3d).
The simulations results achieved by the algorithm, with n = 9 configurations,
Tt = 15000 generations and popsize=300 , converge to two optimal fronts. One of the
fronts (fig. 1(a)) corresponds to the movement of the manipulator around its base in the
counterclockwise direction. The other front (fig. 1(b)) is obtained when the manipulator
moves in the clockwise direction. The solutions a and b, shown in fig. 1 represent the
best solution found for the q̇ and ṗ objectives, respectively.
In 66.6% of the 21 total number of runs, the Pareto optimal front was found. In all
simulations for both cases the solutions converged to a front type which can be modelled
by the following equation (κ, α, β ∈ R):
620 E.J. Solteiro Pires, J.A. Tenreiro Machado, and P.B. de Moura Oliveira

60

50 Non Dominated Solutions

Dominated Solutions
40 Standard deviation
Solutions number

30

20

10

I1 I2 I3 I4 I5 I6 I7 I8 I9 I10 I11 I12 I13 I14 I15 I16 I17 I18 I19 I20
Intervals

Fig. 3. Solution distribution statistics for the 2R robot

B B
1 1

0.5 0.5
y [m]

y [m]

0 0
A A
−0.5 −0.5

−1 −0.5 0 0.5 1 1.5 −1 −0.5 0 0.5 1 1.5

x [m] x [m]

(a) q̇ optimization (solution a) (b) ṗ optimization (solution b)

Fig. 4. Successive 2R robot configurations

Table 2. Range objectives in the 5D optimization

q q̇ Ea p ṗ
min 79.8 18.2 1056.7 83.5 15.0
max 182.3 101.7 4602.7 121.8 56.4

q̇ + α
ṗ(q̇) = κ (6)
q̇ + β

The achieved median, average and standard deviation for the parameters κ, α and β
of (6) are shown in table 1, both for the Pareto optimal and local fronts.
 Robot Trajectory Planning Using Multi-objective Genetic Algorithm Optimization 621

5 8

4 6
q̇1 q̇1
3
4
q̇i [rad/s]

q̇i [rad/s]
2
2
1
q̇2
0
0
q̇2
−1 −2

−2 −4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t [s] t [s]

(a) q̇ optimization (solution a) (b) ṗ optimization (solution b)

Fig. 5. Joint time evolution versus time for the 2R robot

400
a
380 f2

360 f1
f2
340

320 c
p

300
f1
280

260 r
240
d
220 b
50 100 150 200 250 300 350 400
q

–
Fig. 6. Pareto optimal fronts, angular distance vs. cartesian distance optimization: f1 = ab
workspace without obstacles; f2 = cd
– workspace with one obstacle

To study the solution front diversity, the approximated front was split into several
intervals, limited by normal straight lines rm (fig. 2), such that the front curve length
is equal for all intervals. For any two consecutive normal straight lines an interval Im
(m = 1, . . . , 19) is associated, and the solutions located between these lines are counted.
Figure 3 shows the solution distribution statistics achieved by all simulation runs. In this
chart non-dominated and dominated solutions are represented, namely its average and
its standard deviation. From this chart, it can be seen that the solutions are distributed
by all intervals. However, the distribution is not uniform. This is due to the use of a
sharing function in the attribute domain in spite of the objective domain. Moreover,
the algorithm does not incorporate any mechanism to promote the development of well
distributed solutions in the objective domain.
The results obtained for solutions a and b, of the Pareto optimal front in figure 1(a),
are presented in figures 4 and 5. Comparing figures 4(a) and 5(a) with figures 4(b) and
622 E.J. Solteiro Pires, J.A. Tenreiro Machado, and P.B. de Moura Oliveira

2.5
1.5
B 2

1
1.5 B
y [m]

y [m]
0.5 1

0.5
0
0
A A
-0.5
-0.5

-1.5 -1 -0.5 0 0.5 1 1.5 2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3

x [m] x [m]

(a) Solution a (b) Solution c

Fig. 7. Successive 3R robot configurations

2 2 5
q2 (t)
q2 (t) 2
1 5

1 5
1
q1 (t) 1
q(t) [rad]

q(t) [rad]

0 5
0 5
0 q3 (t)
0
q3 (t)
0 5
0 5
q1 (t)
1 1

1 05 0 5 1 05 0 5
0 1 0 2 0 3 0 4 0 6 0 7 0 1 0 2 0 3 0 4 0 6 0 7
t [s] t [s]

(a) Solution a (b) Solution c

Fig. 8. Joint position of trajectory vs. time for the 3R robot

5(b) it is clear that the joint/cartesian time evolution for the optimal solutions a and b,
respectively, is significantly different due to the objective considered. Between these
extreme optimal solutions several others were found, that have a intermediate behavior,
and which can be selected according with the importance of each objective.

3.2 3R Robot Trajectory with 2D Optimization

In this subsection a 3R robot trajectory is optimized using the objectives q (3a) and p (3c)
in a workspace which may include a circle obstacle with center at (x, y) = (2, 2) and
radius ρ = 1. The initial and final configurations are A ≡ {−1.15, 1.81, −0.50} rad
and B ≡ {1.18, 1.47, 0.50} rad, respectively. The Tt and popsize parameters used are
identical to those adopted in the previous subsection. The trajectories which collide with
the obstacle are assigned a very high fitness value.
 Robot Trajectory Planning Using Multi-objective Genetic Algorithm Optimization 623

1 1
s1
Normalized objective value

Normalized objective value

s3
0.8 0.8

0.6 0.6
s2
0.4 0.4

s5
0.2 0.2
s4
0 0
q q̇ Ea p ṗ q q̇ Ea p ṗ
Objectives Objectives

(a) Population tradeoffs (b) Best solutions tradeoffs

Fig. 9. Normalized tradeoffs between q, q̇, Ea , p and ṗ

For an optimization without any obstacle in the workspace the f2 = ab  front (fig. 6)

is obtained. However, when the obstacle is introduced the front is reduced to the f1 = cd.
Thus, only the objective q if affected by the introduction of the obstacle (figures 7 and
8). The solutions {a, b} and {c, d} represent the best solution found for the objectives
{q, p} for the 3R manipulator without and with obstacles, respectively.

3.3 2R Robot Trajectory with 5D Optimization

Here, the 2R manipulator trajectory is optimized considering the five objectives described
by equations (3). Figure 9 and 10 show the optimization results achieved with Tt = 50000
generations and popsize = 1000.
Table 2 contains the objectives range values archived in one simulation. Although, the
5D algorithms can’t obtain so goods solutions as the 2D algorithm due to the significantly
increase in the search complexity, the solutions have a good distribution (figure 9(a)) with
values near to the ones for the 2D respective simulation. Figure 9 shows the normalized
tradeoffs for the entire population and best solutions si = {s1 , s2 , s3 , s4 , s5 } for each
objective Oi = {q, q̇, Ea , p, ṗ}. From figure 9(b) it can be concluded that q and q̇ or
p and ṗ are conflicting objectives with a relative low tradeoff between them. On the
other hand, the Ea objective presents the highest tradeoff among the others objectives.
Figures 10 and 11 show the best solutions obtained for the objectives si . For the studied
trajectory, the results indicate that as the manipulator moves near to its basis the energy
consumed is lower (figures 11(a), 11(c) and 11(d)).

4 Summary and Conclusions

A multi-objective genetic algorithm robot trajectory planner, based on the kinematics
approach, was proposed. The multi-objective genetic algorithm is able to reach optimal
624 E.J. Solteiro Pires, J.A. Tenreiro Machado, and P.B. de Moura Oliveira

1.5 s1 3
s2
s3 2.8
1 s4 s3
s5 2.6
0.5 s4
2.4
q1 (t) [rad]

q2 (t) [rad]
0 2.2 s5

2
−0.5 s2
1.8
−1
1.6 s1
−1.5 1.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t [s] t[s]

(a) Joint 1 position versus time (b) Joint 2 position versus time

6 4
3
5
2 s5
1 s3
4
s2
q̇1 (t) [rad/s]

q̇2 (t) [rad/s]

s1 0
s1
3 −1
s2 s3
−2
2
−3
s5 s4
−4
1
s4 −5
0 −6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t [s] t [s]

(c) Joint 1 time evolution versus time (d) Joint 2 time evolution versus time

Fig. 10. Behavior of the best solutions

solutions regarding the optimization of multiple objectives. Simulation results were

presented considering the optimization of two and five simultaneous objectives. The
results obtained indicate that obstacles in the workspace may reduce the Pareto front
length and for the case studied the single obstacle considered do not represent a difficulty
for the algorithm to reach optimal solutions. Furthermore, the algorithm determines the
non-dominated front maintaining a good distribution of solutions along the Pareto front.

Acknowledgment. This paper is partially supported by the grant Prodep III (2/5.3/2001)
from FSE.
 Robot Trajectory Planning Using Multi-objective Genetic Algorithm Optimization 625

1.4 4
1.2
3 s2
1 s5
0.8 s3 S4 s1
2 s5

vy [m/s]
s4
y [m]

0.6 s2
0.4 1
0.2
0 s1
0
−0.2 s3
−0.5 0 0.5 1 −1 −5 −4 −3 −2 −1 0 1
x [m] vx [m/s]

(a) Cartesian movement (b) Cartesian time evolution

3500 160
s1
3000 140
s1
120
2500
100 s3
2000
Ea

Ea

80
1500
60
s2
1000 s5 s5
40 s2
500 s4
20
s4
0 s3 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t [s] t [s]

(c) Joint 1 energy required (d) Joint 2 energy required

Fig. 11. Behavior of the best solutions (cont.)

References

1. Bäck, T., Hammel, U., Schwefel, H.P.: Evolutionary computation: Comments on the history
and current state. IEEE Trans. on Evolutionary Computation 1 (1997) 3–17
2. Chen, M., Zalzala, A.M.S.: A genetic approach to motion planning of redundant mobile
manipulator systems considering safety and configuration. Journal Robotic Systems 14 (1997)
529–544
3. Davidor, Y.: Genetic Algorithms and Robotics, a Heuristic Strategy for Optimization. Num-
ber 1 in Series in Robotics and Automated Systems. World Scientific Publishing Co. Pte Ltd
(1991)
4. Kubota, N., Arakawa, T., Fukuda, T.: Trajectory generation for redundant manipulator using
virus evolutionary genetic algorithm. In: IEEE International Conference on Robotics and
Automation, Albuquerque, New Mexico (1997) 205–210
626 E.J. Solteiro Pires, J.A. Tenreiro Machado, and P.B. de Moura Oliveira

5. Rana, A., Zalzala, A.: An evolutionary planner for near time-optimal collision-free motion of
multi-arm robotic manipulators. In: UKACC International Conference on Control. Volume 1.
(1996) 29–35
6. Wang, Q., Zalzala, A.M.S.: Genetic control of near time-optimal motion for an industrial
robot arm. In: IEEE International Conference on Robotics and Automation, Minneapolis,
Minnesota (1996) 2592–2597
7. Pires, E.S., Machado, J.T.: Trajectory optimization for redundant robots using genetic algo-
rithms. In Whitley, D., Goldberg, D., Cantu-Paz, E., Spector, L., Parmee, I., Beyer, H.G., eds.:
Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2000), Las
Vegas, Nevada, USA, Morgan Kaufmann (2000) 967
8. Pires, E.J.S., Machado, J.A.T.: A GA perspective of the energy requirements for manipulators
maneuvering in a workspace with obstacles. In: Proc. of the 2000 Congress on Evolutionary
Computation, Piscataway, NJ, IEEE Service Center (2000) 1110–1116
9. Chocron, O., Bidaud, P.: Evolutionary algorithms in kinematic design of robotic system. In:
IEEE/RSJ International Conference on Intelligent Robotics and Systems, Grenoble, France
(1997) 279–286
10. Kim, J.O., Khosla, P.K.: A multi-population genetic algorithm and its application to design of
manipulators. In: IEEE/RSJ Int. Conf. on Intelligent Robotics and Systems, Raleight, North
Caroline (1992) 279–286
11. Han, J., Chung, W.K., Youm, Y., Kim, S.H.: Task based design of modular robotic mani-
pulator using efficient genetic algorithm. In: IEEE Int. Conf. on Robotics and Automation,
Albuquerque, New Mexico (1997) 507–512
12. Gacôgne, L.: Multiple objective optimization of fuzzy rules for obstacles avoiding by an
evolution algorithm with adaptative operators. In: In Proceedings of the Fifth International
Mendel Conference on Soft Computing (Mendel’99), Brno, Czech Republic (1999) 236–242
13. Goldberg, D.E.: GeneticAlgorithms in Search, Optimization, and Machine Learning. Addison
– Wesley (1989)
14. Fonseca, C.M., Fleming, P.J.: An overview of evolutionary algorithms in multi-objective
optimization. Evolutionary Computation Journal 3 (1995) 1–16
15. Deb, K.: Multi-Objective Optimization using Evolutionary Algorithms. Wiley-Interscience
Series in Systems and Optimization. (2001)
16. Horn, J., Nafploitis, N., Goldberg, D.: A niched pareto genetic algorithm for multi-objective
optimization, Proceedings of the First IEEE Conference on Evolutionary Computation (1994)
82–87
17. Coello, C., Carlos, A.: A comprehensive survey of evolutionary-based multiobjective opti-
mization techniques. Knowledge and Information Systems 1 (1999) 269–308
18. Silva, F., Tenreiro Machado, J.: Energy analysis during biped walking. In: Proc. IEEE Int.
Conf. Robotics and Automation, Detroit, Michigan (1999) 59–64