Google open sources DeepMind Lab, aims to make general AI a reality

There are two types of artificial intelligence. Specific, or Narrow AI implementations uses machine learning and neural networks to train artificial intelligences for particular tasks. General AI can handle new circumstances and novel situations that it has not been explicitly trained for. The latter better approximates the results of human like intelligence, and can go beyond the capabilities of humans. General purpose intelligence evolved in the natural world because of rich environments with a variety of challenges. DeepMind Lab hopes to simulate these rich environments to better train AI. The Lab was being used internally by DeepMind researchers. The Lab is a platform similar to a 3D game engine, which can be used to train agents for various tasks. These include navigating procedurally generated environments, remembering paths through mazes, collecting rewards, avoiding hazards, navigating past obstacles or having to evaluate a decision between risk and reward.

Motor control, planning, strategy, navigation and 3D vision are some of the tasks that agents can learn in the environment. There are built in tools for level creation. The rewards, gameplay logic, items, and in-game messages can be customised. This can be used for example, to test out how a trained agent behaves in a novel environment. The assets, along with the code, map and level scripts will be hosted on GitHub. Google hopes that researchers will contribute towards shaping the platform. Here is DeepMind playing laser tag.

Google DeepMind research is focused on building better AI for more demanding cognitive tasks, and building superior environments that can train these advanced AI better. DeepMind has opened sourced its research platform, DeepMind Lab, for broader use by the research community. A GitHub repository for Google DeepMind Lab is expected to go live later this week.