This project explored various machine learning algorithms for handwritten character recognition. We use an OCR framework based on several machine learning techniques, both supervised and unsupervised. First, individual characters are extracted from an input image using the unsupervised clustering algorithm DB SCAN. For unlabeled characters, we explore the use of K-Means and a Gaussian Mixture model to cluster unlabeled characters for convenient manual labeling. Finally, we implement several supervised models to predict characters from images: Decision Trees, Naïve Bayes, a linear Support Vector Machine (SVM), and two convolutional neural networks trained using Stochastic Gradient Descent. More details here.

For this project, a simplified PUMA robot arm is modeled and controlled using Simulink. The robot must complete a drawing task in under 10 seconds with a maximum positional error less than 3 cm. The dynamic equations of motion are defined and a PID controller is implemented to minimize error along the desired path. At each time step, the joint angles are input into the Jacobian transpose and multiplied by the PID controller previously calculated to determine how much torque the motors should apply based on the error in the system. The result is a robot that can draw custom shapes, or in this case, spell words using a single input path.