Autonomous Balancing of 2-wheeled segway robot

BITS Pilani
Pilani Campus
BITS Pilani
Pilani Campus
NAME(ID) : Rishabh Jain - ( 2018A8PS0430P )
Akshit Patel - ( 2018A8PS0094P )
Jash Shah - ( 2018A8PS0507P )
Balancing a Segway robot using LQR controller
based on genetic and bacteria foraging
optimization algorithms
Date: 14 April 2021

BITS Pilani, Pilani Campus
Aim
Our aim of the project is to design optimal control
of the segway robot using lqr and soft computing
optimization techniques like genetic algorithm
and bacteria foraging algorithm.

Introduction
The Segway robot is an electric, two-wheeled self-balancing human transporter with a
computer-controlled gyroscopic stabilization and control system.The mathematical model of the
system dynamics is derived and then state-space formulation for the system is presented to enable
the design of a state feedback controller scheme. In this research, an optimal control system based
on the linear quadratic regulator (LQR) technique is proposed to stabilize the mobile robot.Two
tuning methods, genetic algorithm (GA) and bacteria foraging optimization algorithm (BFOA) are
used to obtain optimal values for controller parameters.

Description of the Robot
The Segway is an extraordinary robot
with very little input constraint and a
high degree of freedom. This makes
it very convenient and fast for use
even for even law enforcement. A
good amount of control is required to
balance the vehicle as it is a
combination of a Unicycle and an
Inverted Pendulum. It now
becomes very intuitive that the
inverted pendulum needs to be
balanced

Modelling of Segway Robot
State Space Equation
where,
where,

Methodology

The Linear Quadratic Regulator (LQR) is
a well-known method that provides
optimally controlled feedback gains to
enable the closed-loop stable and high
performance design of systems.The LQR
algorithm reduces the amount of work
done by the control systems engineer to
optimize the controller.The LQR algorithm
is essentially an automated way of finding
an appropriate state-feedback controller.
LQR

Genetic algorithm is a method
based on reproduction of DNA in
various generations in living
beings. We try to simulate the
process for finding an optimized
solution.The pseudo code for the
process is shown in the figure.
Genetic Algorithm

BFOA
Forging strategies are methods for
locating, handling, and ingesting
food. This method favours those
bacterias that have successful
foraging strategies or the bacterias
with enough food s.t they can
reproduce.After many generations,
poor foraging strategies are either
eliminated or shaped into good
ones i.e. redesigned.

Fitness Function
Fitness function in the paper has not been defined properly, first, we took fitness function to be
variation from desired values which is mentioned in the paper but as rising time and settling time
were not varying much with each population, only small changes in overshoot we directly added
all three values which seemed to provide better results. So our fitness function was the addition of
rising time, settling time, and overshoot of position vs time graph in each iteration.
Fitness Function = |Rise Time| + |Settling Time| + |Overshoot |

MATLAB Simulation
Model Used for Simulation

Population Size = 15
Generation = 10
Best Q Obtained =
q11=7.056311630519287
q22=0.040608399663108
q33=0.150562417774016
q44=0.030452791998497
Best R obtained = 0.000016648572929
K = 1.0e+02 *[-6.510290527654438 -3.928285136135124 -5.775516991619111 -0.891476241249279]
Genetic LQR Parameters

GA graphs(step response vs amplitude)
Position vs Time for GA LQR Angle vs Time for GA LQR

GA graph(Voltage and Time)

GA graph(position and angle vs time
from SimuLink)

GA graph(q values vs iteration)

GA graph (r vs iteration)

Fitness Function = Rise Time + Settling Time + Overshoot
Number of elimination and dispersal steps (Ne)=10;
Reproductive Steps (Nr)=10;
Chemotactic Steps (Nc)=10;
Population Size (Np)=5;
Swim Steps (Ns)=8;
Swarm Size (D)=5;
Probability of elimination (Ped)=0.9;
The run-length unit i.e., the chemotactic step size during each run or tumble (C)=0.01;
Fitness Function Value = 2.170198213995942
Value of Q and R obtained =
[q11 q22 q33 q44 R] = [ q = [72471.87066219141 -0490.898006878771 -04750.45905426314 -05.18027662623499 0.401824162185799];]
Value of K obtained from it =
1.0e+02 * [-4.246847946535302 -2.057666744004326 -2.923004678200982 -0.430108120238977 ]
BFO Parameters

BFO Graphs(voltage vs time)

BFO graph (position vs graph)
Position vs Time for BFO LQR Angle vs Time for BFO LQR

BFO graph(Position and angle vs time)

Some points on original results
a) Value of Q,R,K mentioned in the paper for GA optimization
[q11 q22 q33 q44 R] = [8.969 0.308 0.121 0.0085 2.123*10−5]
K = [-6.4988 -3.8592 -5.1982 -0.7348]
But according to the values of Q and R following K should be obtained which provides correct results
K = 1.0e+02 * [-6.499755421319758 -3.859528575379667 -5.198431774768298 -0.734822500368363]
b) Value of Q and R mentioned in paper for BFO algorithm is following
[q11 q22 q33 q44 R] = [289.104 -5.15e-5 -5e-5 -5.1e-5 0.00028]
K = [-1022 -260.2 886.72 77.7]
But according to the values mentioned of Q and R value of K should be,
K = 1.0e03 *[-1.016127101161195 -0.462096381004273 -0.527358592608302 -0.072462682693461]
In this magnitude is same but signs are different
And according to the values of K found by us results come this which are wrong according to us. Hence BFO is
not providing providing better results according to the values mentioned in the paper

Cont.
Fig. Variation of Position and Angle according to Q and R values calculated from BFO

Observations and Discussions
A. GA LQR seems to perform better and requires less computational power. It took around 1
sec for execution of GA and 4min for complete execution of BFO. Therefore the
conclusion obtained by us is different from that mentioned in paper.
B. Also the value of the fitness function for the most optimized result is the same for GA and
BFO whereas that should have been different for both. This is due to the fact that steady
state is not an error and is not included in the fitness function.
C. According to our modelling we simulated the results using both Steady state and Transfer
function models for a state feedback control scheme. This was different from the standard
controller because of the fact that modelling a TF model for such a controller needs heavy
mathematical manipulations.

Learnings
A. For our model, we incorporated a deviational set-point scheme ( not present in paper ) and
provided it with a step input to see the results. The graphs were obtained as expected after
mathematical manipulations and tunings.
B. We also tried using an animation model, made by Columbia University professors, into our model
to see the results of the segway model. But, we could not get the running animation, but only the
final figure :

Conclusions.
● In this work, an optimal linear control system was adopted to balance a
Segway two-wheeled mobile robot.
● First state space model of the plant was derived and the model was simulated both in
MATLAB and SIMULINK for a step input to position.
● To improve the results for LQR two biologically inspired algorithms were
implemented, GA and BFO, to find optimal value of Q and R.
● According to the obtained results it was found that GA LQR had better performance
than BFO. Though the mentioned conclusion is different from one that is mentioned in
the paper but values provided by them also provides simulation results similar to that
done by us.

Future Scope
● The developed model is a baseline for many other research works. The use of a simple inverted pendulum controller scheme has been
extended to our model. We have not incorporated the turning scenario which would involve non-uniform voltage inputs to both the wheels.
● We have only used a state-feedback LQR for tuning the controller gains. Instead many other controllers like PID and its variants( like Fuzzy
and FOPID) could be used. Even a simple Q-learning based approach could also be incorporated.
● Similarly, for tuning the controller parameters, we could have used any other data-based( DL) and data-free methods. We had planned to use
PSO and ACO along with BFO and GA, but were not able to complete it on time.
● Further, the distance-speed graphs obtained can be sampled and provided to any navigation algorithm using Artificial Potential field and be
used for autonomous navigation of the segway. This hardware model can be easily applied to any microcontroller.
● State estimation Kalman filters could have been used to find the accurate values of the output variables, and this could have made the model
more robust and real-world. In our output graphs, we do not see ripples, which would have been manifested by using state estimation filter.

Thank
You

Autonomous Balancing of 2-wheeled segway robot

More Related Content

Similar to Autonomous Balancing of 2-wheeled segway robot (20)

More from Jash Shah (6)

Recently uploaded (20)

Autonomous Balancing of 2-wheeled segway robot