Paper Overview
Research area: Machine Learning Authors: Derin Gezgin, Jim O'Connor, Tanner Goodwin Published: 2026-08-12 arXiv: 2508.05142
Introduction
This paper presents the Dark Souls Learning Environment (DSLE), a containerized platform that turns all 22 boss encounters of *Dark Souls: Remastered* into game-playing agent benchmarks through a Gymnasium-style interface.
Key properties of DSLE
- Real-time combat combined with high-dimensional visual input
- Sparse, terminal-only rewards
- Every environment step is a real action executed against the running game
- Expert systems and evolutionary baselines can each defeat the tutorial boss, the Asylum Demon, with peak win rates of 63% and 43% respectively
- None of the five methods can defeat the other four DSLE-5 bosses
- PPO and DQN agents show no measurable learning even with budgets of tens of wall-clock hours per run (at most 0.33% win rate on the tutorial boss, 0% elsewhere)
DSLE-5 benchmark subset
To support controlled comparison, the authors define DSLE-5, a representative five-boss subset recommended as the starting suite for agents built on DSLE:
1. A melee fight 2. A spatially constrained arena 3. An environmental-hazard fight 4. A multi-target fight 5. A fast final-boss fight
Evaluation Results on DSLE-5
The authors evaluate a random policy, an expert system, an evolutionary baseline, and PPO and DQN agents trained from visual input:
Failure modes range from dying within 10 seconds in narrow multi-target fights, to stalemates lasting nearly a minute while dealing almost no damage.
Reporting Metrics
Because win rate alone is too coarse, the paper also reports results in terms of survival time and damage dealt.
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...
Paper link: <https://arxiv.org/abs/2508.05142>
--- *Auto-collected on 2026-08-12*