The goal is to develop reinforcement learning for robotics systems by using Aloha Stationary Robot Arms from Trossen Robotics. We developed the reinforcement learning for our robot arm, making it accomplish simple tasks autonomously, such as moving a cup or taking items off a shelf. An operator develops a training dataset for a task through manual operation, which is processed using AI to adapt the variations between datasets to create a flexible method of accomplishing the task without intervention. This is already possible in industry, so we are developing this for Purdue to use as a demonstration for guests.
Reinforced Learning dates back to the early 1960s with computer scientist Donald Michie using it to train a computer to play Tic-Tac-Toe through positive and negative reinforcement. Other developments included the use of evaluation functions, which were used to develop chess algorithms to be able to improve without needing to understand every possibility. Despite years of development in this field, training robotics still had limitations compared to software mediums which slowed its development for direct application. Other problems at the time included the need to control the motion of the robots, not just a list of inputs. This was eventually overcome with neural networks, which would predict which combination of actions would lead to the most reward.
Some current industry issues include high startup costs, varying operating systems and system requirements, and equipment failure. Our robot developed by Trossen Robotics costs $30,000. There are many different robotics companies, each using their own systems which must be paid for separately. Sensory equipment, including our own, can be at risk of failing or miscommunication with the control system, leading to undesirable results such as limited motion or changing results based on tertiary data not related to the task.
This project is a continuation of a project started last year to implement a reinforcement learning policy on a set of robot arms. We have continued the research and setup a GUI system, hugging face repository, simulation trainings, and episode replay ability. We are wrapping up this semester with enough of a framework that hopefully the third year team will be set up for success and be able to complete this project.
Despite engineers' best efforts, simulated and real world environments are different. Reinforcement Learning (RL) robots can be trained in both environments but have difficulty bridging the gap between the two. These types of training for robots have been deployed in the industry. However, most are unsuccessful at performing their assigned tasks, even with training and simulating done as best as possible. If a robot fails while operating in the industrial setting, the consequences can range from a simple disruption all the way to something dangerous, depending on its assigned task and how serious the failure is. This can incur costs, some being resources to fix the robot’s operating system, damage to other components, delayed production, or lost efficiency. As a result, this could make the robot completely useless in the industry. To find the best way to train the robot through Reinforcement Learning (RL), we need to learn its fundamentals. We need to start with the most basic way of training, and then progress to more advanced training methods