Vous êtes sur la page 1sur 24

ROBOTIC

GRASPING
OF
OBJECTS
USING
ARTIFICIAL
INTELLIGENCE

Deepika Udupa
K
USN No: 1MS11TE015

Sindhu S Babu
USN No: 1MS11TE053

B.E- Telecommunication Engg


Guide :
Dr. K. R. Shobha
Associate Professor
MSRIT, Bangalore

M. S. Ramaiah Institute of Technology,

BLOCK DIAGRAM

Vision
Minoru 3D webcam
Stereo vision
Two images captured from two
cameras
Standard 640x480 image
20 frames/sec

Camera calibration
Checkerboard method is used for
calibration

M. S. Ramaiah Institute of Technology,

Triangulation
Depth is calculated by
triangulation, now we have the 3D
point of the grasp point in camera
frame

Predicting grasp point on


the image
Dataset of synthetically
generated objects and
its corresponding grasp
points.
Different classes of
image: mugs, cereal
bowls, erasers, books,
staplers, martini glass,
thick pencil

Feature extraction
Feature detection- Harris corner
detector and GFTT(Good features
to track)
Feature discription- SIFT128 mx128 matrix where each row
represents the feature
description of that corner
Each row is given a label y
six object classes

Machine learning
Logistic regression:
Probability that a
particular sample
belongs to a positive
class lies on the
sigmoid curve.
Threshold is chosen
above which the
positive class is
defined.
Confidence of
predicting the grasp
point was low.
threshold had to be
0.4.

SVM with RBF Kernel:


Tries to find a
boundary that
separates the classes.
Linear kernel gave an
accuracy of 4%, hence
not suitable
Radial Basis
Functiob(RBF) maps
features into higher
dimensional spaces
Since the training set
was much skewed with
only % positive class,
it did not perfrom
well on this

J48 Decision trees


Quinlans algorithm
for generating a
pruned or unpruned
C4.5 decision tree.
Splits the data into
smaller subsets
offers better
classification
Performed well, gave
stunning results.

Mechanical design of the Arm

Katana Architecture
5 DoF : base roll,shoulder, elbow,
wrist pitch, wrist roll.
Parallel plates as gripper
Aluminium flats, 3mm thick, 25mm
wide, for links.
Steel spacers
Ball bearing 6006-ZZ used at the
base roll and 6003-ZZ at wrist roll.

W1=20gm

Servo torque Calulations

l1= 3cm
W2=100gm
l2=17cm
W3=200gm
l3=10cm
W4=100gm
l4=6.5cm
W5= 140gm
l5=16.5cm
W6=56gm
W7=172gm
W8=80gm
W9=160gm
W10=200gm

Gripper servo = l5*W10= 200*16.5= 3.3Kg/cm

Wrist roll= (l5/2)* W9=

(16.5/2)* 160=1.32Kg/cm

Wrist pitch= (l4+l5/2)*W9+ l4*W8+(l4/2)*W7


=(6.5+16.5/2)*160+ 6.5*80+ (6.5/2)*172
=3.439Kg/cm
Elbow=(l3+l4+l5/2)*W9+

(l3+l4)*W8+(l3+l4/2)*W7+l3*W6+

(l3/2)*W5
= 10.139Kg/cm

Shoulder=(l2+l3+l4+l5/2)*W9 + (l2+l3+l4)*W8 + (l2+l3+l4/2)*W7


+ (l2+l3)*W6 + (l2+l3/2)*W5 +
=20.175Kg/cm

l2*W4 + (l2/2)*W3

Base

6 Kg-cm

Shoulde 19 Kgr

cm

Elbow

16 Kgcm

Wrist

12 Kg-

Pitch

cm

Wrist

6 Kg-cm

Roll
Paralle 6 Kg-cm
l Plate (2
Gripper No.s)

Electronic Design

Servos draw a lot of current i.e, up


to 2A, Hence we used SPMS(switched
mode power supply)
All the servos operate at 5V
Servos are controlled through PWM
waves supplied through arduino Mega
PWM pins.
Servos are moved in steps of 2 degrees
Force sensing resistors are used at
the tips of the gripper to classify
the kind of grasp required.
Medium grasp of objects is achieved
with a threshold of 500.

Kinematic solution for the joint


angles
Forward kinematics to calculate the
initial position
Calculated using DH parameters

The transformation from base frame to


the end effector frame T0-4= T0-1*T1-2*T23

*T3-4

Transformation from camera frame to end effector


frame.
Pure Translation

Tcan-4= Tcan-base*T0-1*T1-2*T2-3*T3-4

Frame diagram of the arm is shown


below

DH Parameter table

ai

i-1

di

90

17

10

10

16

Inverse Kinematics
1:

base roll

2 : shoulder
3 : elbow
4: wrist pitch
5 : wrist roll

1 = atan2(y, x)
2 = atan2(y, x)
atan2(k2 , k1 )
where,
k1= l1+ l2 cos 3
k2= l2 sin 3

and

3 = atan2(sin3 , cos3 )
where,

4 = - 2 - 3

The calculated theta values are passed to the arduino


through serial port.

Vous aimerez peut-être aussi