Home

Awesome

SMAClite - Starcraft Multi-Agent Challenge lite

This is a repository for the SMAClite environment. It is a (nearly) pure Python reimplementation of the Starcraft Multi-Agent Challenge, using Numpy and OpenAI Gym.

[!CAUTION] The SMAClite environment was updated to support the new Gymnasium interface in replacement of the deprecated gym=0.25/0.26 dependency. For backwards compatibility, please see Gymnasium compatibility documentation or use version v1.0.0 of the repository.

Features

The main features of this environment include:

Available units

The following units are available in this environment:

Available scenarios

The following scenarios are available in this environment:

Note that further scenarios can easily be added by modifying or creating a scenario JSON file.

Installation

Run

pip install .

In the SMAClite directory

Running

As far as we are aware, this project fully adheres to the OpenAI Gym API, so it can be used with any framework capable of interfacing with Gym-capable environments. We recommend the ePyMARL framework, made available in our repository. EPyMARL uses yaml files to specify run configurations. To train a model in the MMM2 scenario using the MAPPO algorithm, you can use this example command:

python3 src/main.py --config=mappo --env-config=gymma with seed=1 env_args.time_limit=120 env_args.key="smaclite:smaclite/MMM2-v0

Note that to use the C++ version of the collision avoidance algorithm, you will have to add the line use_cpp_rvo2: true to the yaml config file you're referencing, since Sacred does not allow defining new config entries in the command itself.