# How do you model a physical robot (diff drive) in python?

I am using PyGame to simulate the motion of a differential drive robot. So far, I have used the standard diff drive kinematics, multiplying by delta time, and I have the motion working as expected.

What I want to do next is build a PID controller to control the robot in trajectory planning, but I feel that I should build a more realistic model (or rather, I would like to). I understand PID control well, but I am a little unsure about how to build the actual robot model and what I should control as I am new to mobile robotics.

From my understanding, on a real robot (diff drive) you send a control signal to each wheel setting it to a certain velocity. The velocity is read by the wheel encoders, position is estimated using that value in the kinematic equation, and the feedback loop (or whatever control style being used) checks the error on the position to the desired position and adjusts the control signal.

As I stated, in my simulation so far I just have the kinematic models to get the robot position, and a set_velocity function to set the velocity of each wheel. I am confused as to how I can incorporate the other aspects of the real-world examples into my simulation, ie setting the velocities and checking them with the encoders (I think I have an idea of how to simulate an encoder in general since I know how to convert ticks to velocity).

Right now, my velocities are simply a class variable of the robot, so I am having trouble understanding how I might do this...

import pygame
import math
class Robot:
def __init__(self, image_of_robot, starting_location, width, initial_velocity, max_velocity):
# Define the width between wheels
self.width = width

# Define the intial conditions of the robot and define states
self._x_position = starting_location[0]
self._y_position = starting_location[1]
self._direction_angle = starting_location[2]
self._left_wheel_velocity = initial_velocity
self._right_wheel_velocity = initial_velocity

# Define the limits on the velocity of the robot
self.max_velocity = max_velocity

# Create the robot image, the image to be edited, and a bounding rectangle for position
self.rotated_image = self.original_image
self.bounding_rectangle = self.rotated_image.get_rect(center=(self.x_position, self.y_position))

# Get the current robot image and bounding rectangle for blit
def get_image(self):
return (self.rotated_image, self.bounding_rectangle)

# Set the velocity of the wheels.
# Right_wheel = 1 -> right wheel velocity to be set
# Right_wheel = 0 -> left wheel velocity to be set
def set_velocity(self, right_wheel=1, value=0):
if right_wheel:
if abs(self.right_wheel_velocity + value) <= self.max_velocity:
self.right_wheel_velocity += value
else:
if abs(self.left_wheel_velocity + value) <= self.max_velocity:
self.left_wheel_velocity += value

# Update the states (position, direction) of the robot using the equations of motion
def update_states(self, time):
self.x_position += ((self.left_wheel_velocity + self.right_wheel_velocity)/2
*math.cos(self.direction_angle)*time)

self.y_position -= ((self.left_wheel_velocity + self.right_wheel_velocity)/2
*math.sin(self.direction_angle)*time)

self.direction_angle += ((self.right_wheel_velocity - self. left_wheel_velocity)
/ self.width * time)

# Update the images based on current states
def update_image(self):
self.rotated_image = pygame.transform.rotozoom(self.original_image, math.degrees(self.direction_angle),1)
self.bounding_rectangle = self.rotated_image.get_rect(center=(self.x_position, self.y_position))

# Update the robot by updating the states and images to reflect current status
def update(self, time):
self.update_states(time)
self.update_image()


If you want to have a model that will behave closely to a real-life one, I would definitely believe that you should also have a look at the dynamic modeling of a differential drive robot, as Kinematic equations do not include the forces and torques acting on your robot. In fact during the simulation part of a real-world control problem, you should use the dynamic equations of your plant to have reliable results. As far as I believe, there are some online sources, such as articles, on this one.

If you want to keep it simpler however, you can use your kinematic model as well. It might cause heavy amount of work to have the dynamic equations since it requires you to evaluate some coefficients of your vehicle. Kinematic models yield more realistic results for environments with no disturbance (such as friction).

Also what you think about your input signal is correct, you send velocity commands to each wheel. However, if you want to add more reality you may create a function to map the wheel velocity to a PWM signal pulse-width. Because in real-life you would probably use PWM signals to control your motor (therefore, wheel) speeds.

You can also add some noise to your encoder data, since the sensors do not measure with perfect accuracy, which is a very important phenomena for Controller Design process. A controller which yields good results with perfect sensor readings (on simulation) may perform badly in real-life because of sensor noises.

For what you want to control, it is a bit up to you. You may control the position and orientation of the robot, or you may do position and velocity control. It depends on your application and what you want to achieve.

These are what straightly jumped to my mind when I saw your question. I might add some other suggestions, if anything comes to my mind.

• Thank you for the suggestions! What kind of noise would most accurately model the noise seen in encoders normally? I usually use Gaussian because I am familiar with it from schooling but I am not sure that is the type of noise actually present in most real-world sensors. Sep 24 at 16:01
• I am afraid I do not have sufficient knowledge about that. However, as far as I know, some real life sensors' noise characteristics may behave close to a Gaussian distribution (such as a GPS sensor). However some people use different distrubtions to model noise for their sensors (such as Allan Variance). I am however not sure what would be a typical noise model for a wheel encoder, since they have a specific working style. Sep 26 at 14:05