Abstract by Max Robinson
Holodeck: a High-Fidelity Simulator for Deep Reinforcement Learning
Holodeck is a virtual environment powered developed for robotic control training that allows agents to act and learn in a visually rich artificial world. We created Holodeck with numerous realisitic worlds, robotic agents, complex tasks, sensors that provide a testbed for reinforcement learning algorithms. Reinforcement learning is a method by which agents learn how to interact in an environment by receiving rewards for their actions. With current methods, this requires large amounts of experience, which is impractical and expensive in the real world. Virtual environments give agents the ability to quickly gain large amounts of experience without potentially damaging hardware. Robot control tasks focus on learning control policies that move a robotic agent through a world. Most popular control environments are visually simple and rely on joint and location sensors as opposed to Holodeck’s vision based sensors. Our work on Holodeck creates a unique framework for deep reinforcement learning with a focus on a visual state space.