OpenAI was created to democratize and decentralize artificial intelligence. The nonprofit came closer to that goal with the release of Universe, a platform that will allow anyone to create advanced AI agents as part of the group's ongoing research to make AI more human-like. Alphabet's DeepMind also released its own open source 3D virtual lab to help researchers train AI agents in real world-like environments.
Universe allows AI agents to do all sorts of tasks a human would do with a keyboard and a computer while watching a screen. The platform comes with over 1,000 environments--including Flash games, browser tasks, and games like Grand Theft Auto V--to fill that purpose. OpenAI's ultimate goal is to create an AI agent that draws from the experience of the many other agents using the platform to learn new skills within all the available environments.
Games will be an area of significant focus for the Universe platform. AI researchers, or anyone willing to help, can let the AI agents train on games they play. Soon, the AI agents will be able to play those games as well as human-like bots that learn by visualizing what’s happening on the screen, rather than taking actions based on pre-programmed algorithms.
So far, Universe includes over 1,000 Flash games, Atari 2600 games, and environments where the AI agent interacts with websites’ user interfaces. OpenAI hopes to expand this list to include more games from partners such as EA, Valve, and others, as well as HTML5 games, Unity games, online educational games, and possibly other types of environments as well.
Although the DeepMind AI has also been trained on a multitude of games, the company has also created its own virtual lab for building 3D environments in which AI agents could train. The environment is seen from the first person view of the AI agent, which has a “floating orb” for a body. It can look and move around in these 3D environments, and it can also do things like collect fruit, navigate mazes, traverse dangerous passages, play laser tag, and learn and remember procedurally generated environments.
DeepMind said its virtual lab emphasizes navigation, memory, 3D vision, motor control, planning, strategy, and time. Fully autonomous agents also have to learn on their own what tasks to perform in a given environment.
According to the company, the DeepMind Lab is highly customizable, and researchers can build their own levels for AI training. The levels can be modified with gameplay logic, item pickups, custom observations, level restarts, reward schemes, in-game messages, and more. All code, maps, and level scripts will be available on DeepMind’s Github page.
The company said that so far it has only scratched the surface on many of these tasks, and there is still much to be discovered through research in the areas of navigation, memory, and exploration. However, it hopes that other researchers using the DeepMind Lab can speed up the process of achieving a general-purpose artificial intelligence.
The DeepMind team believes that it’s fundamentally easier to train artificial general intelligence in 3D environments seen from a first person point of view. The team thinks that we, humans, wouldn’t have developed too much general-purpose intelligence if we were born in a Pac-Man environment. Following the same logic, it should be easier to create a general-purpose AI agent by training it in a 3D world.