Download
The universe-starter-agent repository is an archived OpenAI codebase designed as a starter reinforcement-learning agent that can interact with and solve tasks in OpenAI’s Universe environment platform. Its purpose is to serve as a baseline or reference implementation so researchers or developers can see how to build agents that operate in real-time, visual environments (e.g., games, browser apps) via pixel observations and keyboard/mouse actions. Under the hood, this starter agent implements a version of the A3C (Asynchronous Advantage Actor-Critic) algorithm, adapted for the specific challenges of Universe environments (e.g., network latency, VNC streaming, asynchronous observations). The repo includes modules like train.py, worker.py, model.py, a3c.py, and envs.py to support training, parallel worker management, policy/critics, and environment wrappers.
Features
- Asynchronous A3C implementation adapted for remote, visual (pixel-based) environments
- Wrappers to interface with VNC, remote desktops, and Universe’s streaming interface
- Support for multi-worker training with parameter server architecture
- Environment abstraction via envs.py to manage different Universe tasks (Atari, Flash, etc.)
- Configuration and monitoring via TensorBoard, tmux session orchestration for parallel processes
- Instructions and example commands for training agents (e.g. on PongDeterministic-v3)
Project Samples
