👾Multi-Agent 🎮 FPS Gym Environment with 🏋️ bevy_rl

app.2022-09-27.04-13-58.mp4

Project Description

This is a basic multi-agent gym environment for bevy_rl. It is a deathmatch free-for-all environment where agents spawn as red spheres and get +10 reward on kill. The environment is implemented in Rust using bevy game engine and bevy_rl plugin.

It implements very basics of a multi-agent environment. It is a good starting point for creating more complex environments.

• Random initialization of agents
• REST API for controlling agents (including state, reward and camera pixels)
• REST API to reset an environment

You can wrap the environment with a python wrapper and use it with OpenAI Gym interface. (example in python/bevy_rl_rest_api.ipynb)

Environment Description

• 16 agents spawn in a random position
• Agents can move and rotate
• Environment pauses every 0.1 second to fetch control commands from REST API
• Reward: +10 on kill

Usage

• follow bevy's setup guide to set up Rust
• build an environment with cargo build +nightly --release;
• run environment with ./target/release/bevy_rl_shooter --mode train
• python/env.py implements a python wrapper for an environment
• python/bevy_rl_rest_api.ipynb illustrates how to use the wrapper