Gym-microrts's V2 Benchmark

In this report, we hope to provide a preliminary benchmark on gym-microrts's V2 environments, which add the following features compared to the V0 environments:

1. Support for full action mask, meaning the mask on action parameters is provided as well in V2 in addition to the just the source unit mask avaiable in V0
2. Support for no-frame skipping.
3. Support for issuing actions on all units at the same game tick.

Feature 2 & 3 actually suggests the bot trained by our RL algorithm is now ready to compete against the built-in bots of $\mu$RTS, which is not possible for the V0 environments.

We now present our benchmark on some popular tasks in $\mu$RTS such as Mining resources, producing workers and on defeating some built-in bots.

Play against existing bots