Overview
- Field: Machine Learning
- Authors: Derin Gezgin, Jim O'Connor, Tanner Goodwin
- Published: 2026-08-11
- arXiv: 2508.03798
- DSLE platform: A containerized learning environment exposing all 22 boss encounters of *Dark Souls: Remastered* as game-playing agent benchmarks through a Gymnasium-style interface. It combines real-time combat, high-dimensional visual input, and sparse terminal rewards — each environment step is a real action executed against the running game.
- DSLE-5 benchmark subset: A representative five-boss suite recommended as the starting point for agents built on DSLE, spanning:
- a melee fight
- a spatially constrained arena
- an environmental-hazard fight
- a multi-target fight
- a fast final-boss fight
- Baselines evaluated on DSLE-5: a random policy, an expert system, an evolutionary baseline, and PPO and DQN agents trained from visual input.
- The expert system and the evolutionary baseline each defeated the tutorial boss, the Asylum Demon, with peak win rates of 63% and 43% respectively.
- None of the five methods could defeat the other four DSLE-5 bosses.
- PPO and DQN showed no measurable learning within evaluation budgets (up to 0.33% win rate on the tutorial boss, 0% elsewhere), where each run already required tens of hours of wall-clock time.
- A broader study ran the evolutionary baseline across all 22 encounters; even with all 50 levels invested in each advantageous attribute, it won only a few additional early-game bosses, losing the rest.
- Failure cases ranged from deaths in under 10 seconds in narrow, multi-target encounters to stalemates lasting nearly a minute with almost no damage dealt. The authors report these via survival time and damage dealt rather than win rate alone.
Key Points
Findings
Original Abstract (excerpt)
> We introduce the Dark Souls Learning Environment (DSLE), a containerized platform that presents all 22 boss encounters of Dark Souls: Remastered as game-playing agent benchmarks through a Gymnasium-style interface. DSLE combines real-time combat, high-dimensional visual input, and sparse terminal rewards, with each environment step being a real action executed against the running game. To support controlled comparison, we define DSLE-5, a representative five-boss subset, spanning a melee fight, a spatially constrained arena, an environmental-hazard fight, a multi-target fight, and a fast final-boss fight, that we recommend as the starting suite for agents built on DSLE. On DSLE-5 we evaluate a random policy, an expert system, an evolutionary baseline, and PPO and DQN agents trained from vi...
---
*Auto-collected on 2026-08-12*