Project Description
This project focuses on automating the labor-intensive process of making handmade tortillas using
a 6-Degrees of Freedom (6-DoF) robot arm. By integrating advanced tracking and motion planning
algorithms, the robot can efficiently and accurately handle different positions of tortilla
dough balls, enhancing productivity and uniformity in the production process.
Core Features
- Automatic Doughball Handling: Use a robotic arm to automate the repetitive task
of moving doughballs from one location to another.
- Improved Efficiency and Uniformity: Implement off-line motion planning to optimize
the robot's movements and adapt to changes in the doughball layout.
- Visual Tracking: A camera positioned above the workspace provides real-time tracking
(using a Fiducial Marker) and allows for adaptation to changes in doughball placement.
- Machine Learning: Integration of machine learning libraries using the sklearn library
to improve the robot's decision-making process and use the best angle of approach to grab doughballs.
Determining the Angle of Approach (AoA) using Machine Learning
A Random Forest Regressor was chosen as the algorithm that provided the most robust performance in varied scenarios.
- Traning the Model:
-
Data Collection: Data was gathered from multiple test runs, capturing various scenarios
of doughball placements.
-
Feature Selection: Key features include x, y, z coordinates of the doughballs and the angle of
approach.
-
Model Traning: A dataset of 1000 different possible doughball placements was hand annotated with the best
angle of approach (angle increments of 15 degrees - 0, 15, 45, 60, 90, etc..). The Random Forest Regressor model was
trained using this annotated dataset to predict the optimal angle of approach.
-----Place image of trained dataset------
- AoA Determination:
-
Input Processing: Trained model processes the input data (current coordinates) to determine the best
AoA for each doughball.
-
AoA Decision: The AoA function uses the model output to decide which angle to approach (in increments of 15 degress)
the doughball, considering potential obstructions and kinematic constraints.
State Machine
----- Place image of state diagram -----
- Initialize: Robot will calibrate its position using a fiducial marker, and a reference
doughball is placed the the center of the designated work area. This establishes baseline coordinates and any
necessary offsets for accurate operation.
- Visual Scanning (continuous): An Intel Realsense is positioned approximatley 1 meter above the doughball
to provide a birds eye view.
The camera continuously scans the workspace to detect the presence and positions of doughballs in real-time.
- Angle of Approach Determination: Based on the detected positions, the angle of approach (AoA) for
picking up each doughball is determined using a machine learning algorithm. This decision considers whether the
doughball is close to the center of the edges of the workspace.
- Motion Planning: The robot arm's movement is planned using the determined AoA, ensuring it approaches
doughballs at the correct orientation.
- Doughball pickup: The robot arm moves to the correct location, and the gripper gently grabs
the doughball without causing deformation or damage.
- Resting: Once all doughballs are handled, the robot returns to its home position, ready for
more doughballs to be placed.