Last week DeepMind, an artificial intelligence company part of Google, released a sort of “video game” meant to be played by artificial intelligence systems. The game is called “DeepMind Lab” and was created by researchers to train AI systems how to interact with virtual environments. DeepMind released the code for the virtual AI testing environment on GitHub, free for anybody to view and contribute to. They hope that other AI researchers will be able to use the training environment to train new AI agents, and to contribute back to the project.
Training machines to play video games may seem silly, but DeepMind thinks it is an important step for the advancement of robotics. Training an AI system to interact with a simulated environment can pave the way to teaching robots how to interact with a physical one. Typically, AI agents would be fed special information about a virtual environment – more information than a human could possibly process in real time. In this training environment though, the agent only has access to the pixels displayed on a screen as well as access to the controller – just like a human would when playing a video game, or a robot walking through the real world. The challenge is to navigate and even interact with this environment by learning from just the pixels on the screen. DeepMind describes the project on their GitHub page:
“DeepMind Lab provides a suite of challenging 3D navigation and puzzle-solving tasks for learning agents. Its primary purpose is to act as a testbed for research in artificial intelligence, especially deep reinforcement learning.”
DeepMind is using a form of machine learning called reinforcement learning where the agent receives rewards for performing an action successfully. DeepMind stated in a research video, in Lab, “the player can either go for a quick reward (apples) or can take a small negative reward (lemon) in order to get the highly prized melon!”
The release of Lab follows DeepMind’s success with AlphaGo, a program that beat world-class Go player Lee Se-dol at 5 rounds of Go. Lab is the next step for DeepMind to experiment with reinforcement learning. “Just like you or I would learn about the world as a child, it’s a very fundamental approach to this learning and generality problem,” said Shane Legg, machine learning researcher and founder of DeepMind. Eventually the research done with Lab could help robots navigate the real world.
To learn more:
“DeepMind Lab” GitHub page: https://github.com/deepmind/lab
DeepMind video showing reinforcement learning with Lab: https://www.youtube.com/watch?v=gC_e8AHzvOw