How I Made A Deep Learning Robot

While AI is often associated with digital applications, its role in hardware and physical environments is just as exciting! To explore AI past software, I built and trained a robotic arm capable of learning and replicating tasks using a machine learning model called an Action Chunking Transformer (ACT). To do this, I 3D printed and wired both a "leader" and a "follower" arm, allowing me to teleoperate the follower and record demonstrations such as placing blocks into different buckets. Using this dataset, I trained the ACT model to learn from my examples, enabling the robotic arm to autonomously perform the task and even generalize to fixing its own mistakes.