The grand challenge of this project is to use video games and virtual environments to train agents for real-world robotics use. This project is important because it aims to improve the ability and efficiency of training robots. Using virtual environments allows for fast and cheap training and testing. We used a deep reinforcement learning algorithm to train our agents. We started with 2D games, such as Atari Space Invaders, then moved on to 3D environments in Unity. In our simulated search and rescue environment, which consisted of a robot moving through a pseudorandom environment to retrieve a person and take them out of a building, we achieved an 89.8% success rate after 60,000 episodes, or about 2 hours of training. Without time and hardware restraints, this number should approach 100%.

View the full poster here