- Update code to work with new Gym and numpy.
- Include Toribash binaries in the repo to make installation easier.
- Added state-machine to make sure actions/states are handled in right order.
- Added (experimental) code for going through replay files
- Added new experiment codes (stable_baselines and rlkit experiments)
- Added possibility to "remote control" in multiplayer games
- Done via "manual_torille.py", where human player must manually launch the script in Toribash, which then connects to the remote controller
- Refactoring code to avoid one huge file
- Added match length, current frame and frames next turn to state
- Added possibility to change game mod
- Fixed off-by-one error with turnframes which prevented from using turnframes=1
- Updated pypi-package to use Toribash 5.4
-
Fixed Gym environments changing actions list in-place during call to
step -
Added info on game winner to ending state
-
Added rotation matrix of players' groins to game state
-
Added velocities of body parts to game state
-
Added "get_normalized_locations" function to game state
-
Added normalization to Gym environments
-
Added new duo-combat environment with winning reward (rather than injury reward)
-
Torille now attempts to make stderr.txt file read-only to avoid huge files (on Linux/Darwin)
-
Added code of new experiments to
experimentsfolder
- Added support for Mac/OSX (with Windows binaries, durr)
- Fixed Gym environments on Windows
- Restructured Gym environments and added some more
- Added limit for turnframes [2,matchframes]. Later turnframes=1 might be made possible
- Fixed "static" frames at the start of learning: First few states were always the same
- Added sanity check for available displays on Linux (for $DISPLAY="")
- FPS on Windows 10 improved by ~10%
- Removed resolution changes. This caused crashing on some Linux systems with Wine and windows
- Fixed Wine version checking
Initial public release.