DeepMind, the Google-owned startup behind AlphaGo, has announced that it is open-sourcing its flagship platform, DeepMind Lab, to spur further research and innovation in artificial intelligence. The long-term goal is to eventually develop sophisticated cognitive systems that can learn to solve complex problems without guidance. To create AI agents that are innovative, adaptable, and able to operate across a wide variety of tasks, complex training and evaluative environments must also be designed and implemented. To that end, DeepMind Lab was created as a fully-3D game-like platform for AI research and training.
This environment is meant to foster and evaluate AI learning by emphasizing factors such as navigation, memory, 3D vision from a first person viewpoint, motor control, planning, strategy, time, and autonomy. The platform, comes with rich and detailed visuals and is observed from the first-person viewpoint of the agent– just like a video game. The agent is rendered as a floating orb that looks around using a rotating, ball-in-socket camera. It navigates its 3D environment with the aid of propulsive thrusters that enable it to jump, crouch, move forward and back, and strafe left and right. It can perform an array of tasks which include collecting fruit, navigating in mazes, traversing dangerous passages while avoiding falling off cliffs, bouncing through space using launch pads to move between platforms, and playing laser tag.
DeepMind Lab is also easily customizable and extendable. New levels can be designed using off-the-shelf software. These levels can be further customized with gameplay logic, item pickups, custom observations, level restarts, reward schemes, in-game messages, etc. DeepMind Lab’s level-creation interface can even be used to design maps that extend and alter themselves, in order to test how agents deal with unfamiliar environments and uncertainty. By opening the platform to the broader research community, DeepMind hopes to empower researchers to both train and test their agents and push the frontiers of AI.
Why turn to 3D environments that can be observed from a first-person POV in order to train and evaluate AI agents? The answer, of course, is that this is meant to approximate the natural world, which is the only environment we know of (thus far) that has developed intelligent beings. It is believed that general intelligence developed as a direct consequence of the complex, 3D environment furnished in the natural world. In other words, animal and human intelligences most likely could not have evolved to their current states without being shaped by the perceptually and physically rich environments– an idea that traces it philosophical lineage back to John Locke and David Hume. To put it differently, Counter-Strike gameplay in 3D requires and promotes intelligence more than 2D Space Invaders.