For our project, we wanted to find the best way to quickly train robots to learn and adapt to complete various tasks as efficiently as possible with minimal training. Some types of training can take up to days or weeks. However, imitation learning, which is a type of supervised learning, can take only a couple minutes to train. It involves going through multiple tests which include collecting data, preprocessing data, finding the optimal hyperparameters for the model, testing and simulating the model, and accomplishing the task in the real world with a robotic arm.

For this project, we used Python and Anaconda for scripting and environment management. We used Jupyter Notebook to create and manage our scripts. We used PyTorch and their neural network library to build our machine learning model. For our simulation environment, we used MuJoCo, along with a miniaturized version of our arm for collecting data and verifying prediction outputs.

When building a machine learning model, there are a few configuration variables - called hyperparameters - which adjust how our neural network algorithm learns and uses the given training data. The hyperparameters critical to ensuring accuracy are number of epochs, batch sizes, and learning rates.

• Epoch: Represents 1 execution of the training dataset and the prediction that is made as a result of it

• Batch Size: Represents how much data should be passed into the model in each iteration through the training data

• Learning Rate: Represents how much the algorithm should "explore" vs. "exploit" the training data

An accurate model means the calculated loss is as low and "smooth" as possible. The graphs you see below represent loss vs epochs. Ideally, we want a model that reaches low loss values quickly, and stays constant.