Demonstrating TRAinAR

摘要

We demonstrate TrainAR, an augmented reality (AR)-based tool that is designed to improve sim2real reinforcement learning (RL) for robots. Users with TRAinAR can tailor a virtual training environment with constraints to match the real-world, visualize training data to gain insights into an agent's learning process, and animate a robot's future actions before execution. The system described here enabled a robotic arm manipulator to learn how to navigate its end-effector toward a target object. The novelty of our AR application will hopefully enable users to better train a robot by quickly prototyping complex environments that are difficult to model.