DeepMind Technologies, an artificial intelligence company owned by Google parent Alphabet plans to open-source all the code for the central platform it uses to train virtual software agents to perform complex tasks autonomously.
The code for DeepMind Lab will be hosted on GitHub along with scripts and all other relevant assets related to the technology. The idea is to give AI researchers at other organizations and even the general public a way to test and train their own artificial intelligence agents using the DeepMind environment.
DeepMind grabbed considerable attention earlier this year when its AI system beat the top ranked player in the world at the ancient Chinese board game of Go.
The company’s research is focused both on designing intelligent AI agents like the one that played Go and also on building increasingly sophisticated environments where such agents can learn how to do things on their own using raw inputs. Google has used DeepMind internally for some time and has said the technology is critical to its efforts to make its applications and services smarter and more intuitive.
DeepMind describes Lab as a first-person 3D game platform in which an AI agent autonomously learns and remembers how to perform different tasks like navigating a maze, bouncing through space, collecting apples and traversing narrow pathways without falling off cliffs.
“Artificial general intelligence research in DeepMind Lab emphasizes navigation, memory, 3D vision from a first person viewpoint,” DeepMind researchers said in a blog this week. The platform also provides an environment to study and refine how autonomous agents acquire motor control, learn how to plan and strategize as well as assess what tasks to perform by exploring their 3D game environments.
The software that DeepMind will make public the week of Dec. 5 is both extendable and customizable. Researchers can add their own levels to the 3D game platform using readily available tools or customize it in a variety of others ways. For example, researchers could add their own gameplay logic, set their own thresholds for attaining new levels of gameplay or add new reward signals when an agent learns or acquires a new skill.
AI Researchers also can create new game environments in DeepMind Lab on the fly so agents know how to navigate through and learn in unfamiliar environments, the DeepMind researchers said. “Our hope is that the community will help us shape and develop the platform going forward.”
Though DeepMind’s technology has played a critical role in moving AI research forward, a lot still remains undiscovered in the realm, according to the researchers. There are still multiple opportunities for major contributions in relatively untouched AI research domains, such as autonomous exploration, navigation and memory.
Meanwhile, in a separate but related development, Elon Musk’s OpenAI artificial intelligence company also said on Dec. 5 that it would open source the code for its Universe platform for training and measuring an AI agent’s intelligence capabilities.
The software, according to OpenAI, is designed to let an AI agent use a computer much like a human being does by operating a virtual key board and mouse and looking at screen pixels to determine what’s on the screen. Universe provides an environment where researchers can train a single AI agent to perform any task that can humans can do with a computer.
OpenAI’s release includes about 1,000 game environments that researchers can use to train their AI agents on various tasks.