強化學習／評測
DSLE Packages 22 Dark Souls Boss Fights as Gymnasium Environments, but the Public Repository Is Still an Empty Shell
Using Docker, Wine, virtual displays, and process-memory reads, DSLE turns real-time combat in a commercial game into a reinforcement learning benchmark with parallel resets. Five baselines consistently conquer only the tutorial boss, while the code the paper claims is public has yet to appear in the repository.
Dark Souls Learning Environment (DSLE) wraps 22 boss fights from *Dark Souls: Remastered* in a Gymnasium-style `reset`/`step` interface. By default, agents receive 600×800 grayscale frames and choose from 14 discrete actions covering movement, attacks, healing, rolling, and blocking; the game executes each action for 250 milliseconds. Rewards account for damage dealt to the boss, damage taken by the player, a per-step cost, victory, and death. Termination is determined from health values and event flags in the game process's memory, but the policy itself is not given access to these structured states.
The execution layer is more noteworthy than that of a typical simulator. DSLE runs the unmodified Windows game through Wine inside Ubuntu Docker containers, uses DXVK to translate DirectX to Vulkan, and assigns each instance its own X11 display, Wine prefix, save file, and VNC connection. During resets, it copies in a designated save file, navigates menus using screen templates, handles cutscenes, and walks into the boss arena. Reads from `/proc/<pid>/mem` provide reward signals, diagnostics, and termination detection. On a host with an RTX 5090 and 128GB of RAM, the authors ran as many as 30 instances, consuming roughly 10GB of VRAM and 80GB of system memory in total.
The results show that this is not a game wrapper already solved by conventional RL. Each PPO and DQN configuration was limited to 100,000 steps, yet a single run still took about 9 to 41 hours. Neither algorithm exceeded a 0.33% win rate against the tutorial boss, Asylum Demon, and both recorded zero wins against the other four DSLE-5 bosses. A rule-based controller and the SCOPE evolutionary strategy reached peak tutorial-stage win rates of 63% and 43%, respectively, but also failed to defeat the harder scenarios.
The biggest release risk is reproducibility. The paper explicitly states that the repository contains Docker files, save files, baselines, and evaluation scripts, but at the time of inspection, the GitHub repository was still empty. Users must also provide their own legally obtained copy of the game, along with Linux, an NVIDIA GPU, and Docker. For now, DSLE is better viewed as a thoroughly described benchmark proposal. Only once the implementation is actually released will it be possible to verify whether its memory offsets work across different game versions and whether parallel resets are reliable.