DSLE: A Learning Environment for Dark Souls Boss En...

DSLE: A Learning Environment for Dark Souls Boss En... | AI Research

Key Takeaways

The Dark Souls Learning Environment (DSLE) is a containerized platform designed to turn Dark Souls: Remastered into a standardized benchmark for artificial i...
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.
On DSLE-5 we evaluate a random policy, an expert system, an evolutionary baseline, and PPO and DQN agents trained from visual input.
A broader study running the evolutionary baseline across all 22 encounters under advantaged all level-50 stats yields wins on only a handful of additional early-game bosses and leaves the rest unwon.

Paper AbstractExpand

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 visual input. The expert system and the evolutionary baseline each defeat the Asylum Demon, the game's tutorial boss (63% and 43% peak win rates), but none of the five methods defeats the other four DSLE-5 bosses; PPO and DQN show no measurable learning (at most 0.33% win rate on the tutorial boss, 0% elsewhere) within a budget that already costs tens of wall-clock hours per run. A broader study running the evolutionary baseline across all 22 encounters under advantaged all level-50 stats yields wins on only a handful of additional early-game bosses and leaves the rest unwon. The failure cases range from sub-10-second deaths in cramped, multi-target encounters to minute-long stalemates that inflict almost no damage, and we report them through survival time and damage dealt rather than win rate alone.

The Dark Souls Learning Environment (DSLE) is a containerized platform designed to turn Dark Souls: Remastered into a standardized benchmark for artificial intelligence. By providing a Gymnasium-style interface for all 22 boss encounters in the game, the researchers aim to create a reproducible way to evaluate how agents handle complex, real-time combat, high-dimensional visual inputs, and sparse rewards.

Standardizing Boss Encounters

Previous attempts to use Dark Souls for AI research relied on visual estimation of game states and were often limited to single-boss experiments. DSLE improves on this by using direct memory access to reliably track player and boss health, automated save-state swapping for consistent resets, and headless execution to support distributed training. Each of the 22 boss encounters is treated as an independent task, allowing researchers to test agents across a variety of challenges, such as multi-target fights, environmental hazards, and spatial constraints.

The DSLE-5 Benchmark

To provide a starting point for researchers, the authors defined "DSLE-5," a representative subset of five boss encounters. This suite includes the Asylum Demon (a tutorial boss), the Capra Demon (a spatially constrained fight), Chaos Witch Quelaag (an environmental-hazard fight), Ornstein and Smough (a multi-target fight), and a fast-paced final boss encounter. The authors recommend this subset as a baseline for comparing different agent architectures.

Performance and Limitations

In evaluations using the DSLE-5 suite, current methods struggled significantly. While an expert system and an evolutionary baseline were able to defeat the tutorial boss (the Asylum Demon) with peak win rates of 63% and 43% respectively, none of the five tested methods—including PPO and DQN agents trained from visual input—could defeat the other four bosses.
The PPO and DQN agents showed no measurable learning on the harder encounters and failed to exceed a 0.33% win rate on the tutorial boss, despite training budgets costing tens of wall-clock hours per run. Broader testing with an evolutionary baseline across all 22 bosses, even with advantaged character stats, resulted in wins on only a few early-game encounters. The authors note that failure cases range from rapid deaths in cramped arenas to long stalemates where the agent fails to deal meaningful damage.

Franklin Analysis

The results suggest that Dark Souls: Remastered remains a formidable challenge for current reinforcement learning agents. The evidence—specifically the failure of PPO and DQN agents to learn beyond the tutorial boss—indicates that the game’s combination of high-dimensional visual input and complex, reactive combat dynamics creates a significant barrier for standard learning algorithms. The authors’ decision to report results through survival time and damage dealt, rather than just win rates, provides a more granular view of agent progress in an environment where total victory is currently elusive.