Measurement and training for artificial intelligence.

Read blog post View on GitHub

Bring your own agent.

Use existing numerical computation libraries, such as TensorFlow or Theano.

Run any program.

Universe allows an AI agent to easily interact with any computer program.

Perceive like a human.

Observe the program the same way a person does, by looking at the pixels.

React like a human.

Act using the same controls as a person, by using keyboard and mouse.

Learn. Grow. Get Smarter.

Solve successively harder environments to develop general problem solving ability.

1000+ environments today.
Tens of thousands tomorrow.

Solve and learn 1000s of environments to develop common sense — the ability to rapidly solve a new problem.

Measure intelligence

A software platform for evaluating and training intelligent agents across the world’s supply of games, websites and other applications.

1000+ environments

Universe contains an ever-expanding catalog of environments. Environments will soon include programs from EA, Microsoft Studios, Valve, Zachtronics, and others.

View on GitHub

Human-like interface

Agents use the same senses and controls as humans: seeing pixels and using a keyboard and mouse. Universe makes it possible to train a single agent on any task a human can complete with a computer.

Read blog post Join forum

Diverse AI training

Develop general problem solving skills by training on hundreds of diverse tasks.

Most AI agents are trained to solve individual tasks one at a time. Create sophisticated, flexible agents by training them on a diverse suite of tasks.

  • 1000+ environments available in the catalog to learn from.
  • Diverse set of tasks: games, scientific software, and more.
  • Integrate more environments to accelerate the entire community.

1000+ Difficulties

Continous difficulty gradient.

Shared AI benchmarks for transfer learning.

Compare and share benchmarks with the wider community.

Coming Soon

Measure and train agents across a set of standardized tasks.

Transfer and lifetime learning

Universe supports agents which learn across tasks, developing general-purpose knowledge about the world.

Vast diversity of tasks

Turn any program into an environment, or train across existing ones. Universe doesn't need access to program internals, source code, or bot APIs.

Getting Started

Run your AI agent on 1000+ environments. All it takes is nine lines of code. 

import gym
import universe # register the universe environments

env = gym.make('flashgames.DuskDrive-v0')
env.configure(remotes=1) # create one flashgames Docker container
observation_n = env.reset()

while True:
  # your agent generates action_n at 60 frames per second
  action_n = [[('KeyEvent', 'ArrowUp', True)] for ob in observation_n]
  observation_n, reward_n, done_n, info = env.step(action_n)
  env.render()

View the Universe library on GitHub. Start training on any environment in minutes.

We have many more environments waiting to be integrated.

Learn more

Join our community of researchers and engineers on our open forum.

Join forum