Tencent AI Lab PySC2 Extension

Note: the original Deepmind PySC2 README can be found here.

Note: Current commit (>= 5042919a 2020/11/19) works for TStarBot-X; To run with the old TStarBot1 and TStarBot2, please revert to the commit 4f790218 2019/5/15

Besides the "feature_layer" observations/actions interface, this Tencent AI Lab fork also exposes the "raw" interface of s2client-proto to enable a per-unit-control.

It supports a hybrid use of the two intefaces. For example, consider a two-player game and the code below

timesteps = env.step(actions)

For player_id = 0, all the uints in pb format can be accessed via timesteps[player_id].observation['units], while the original Deepmind PySC2 features can still be accessed via timesteps[player_id].observation['feat_name'].

For the actions passed in, acionts[player_id] can be either a list of pb actions or a single Deepmind PySC2 action. (TODO: support a list of hybrid action when necessary).

It goes similar for the other player player_id = 1.

Installation

git clone the repo, cd to the folder, and run

pip install -e .

Note: the in-place -e . installation is REQUIRED, as we have binaries (i.e., the tech_tree data) shipped with the fork and the -e . in-place installation makes life easier.

Note also that you need pip uninstall the original Deempind PySC2 before installing/using our fork. Doning so would not be a problem, as this fork is compatible with the original Deepmind PySC2.