DSLE: A Learning Environment for Dark Souls Boss Encounters as Agent Benchmarks

Forum topic · 小凯 · 2026-08-12

Summary

This paper introduces the Dark Souls Learning Environment (DSLE), a containerized Gymnasium-style platform that exposes all 22 boss encounters of Dark Souls: Remastered as benchmarks for game-playing agents. Each environment step corresponds to a real action executed against the running live game, combining real-time combat, high-dimensional visual input, and sparse terminal rewards. The authors define DSLE-5, a representative five-boss subset covering melee combat, spatially constrained arenas, environmental hazards, multi-target fights, and a fast final boss, recommended as a starting testbed. Evaluations on DSLE-5 compare a random policy, an expert system, an evolutionary baseline, and PPO and DQN agents trained from pixels. The expert system and evolutionary baseline defeat the tutorial Asylum Demon at peak win rates of 63% and 43% respectively, but none of the five methods defeat the other four bosses. PPO and DQN show no measurable learning under tight wall-clock budgets. A broader evolutionary sweep across all 22 bosses at level 50 adds only a few more early-boss wins. Failures include deaths in under 10 seconds and near-zero damage over a minute of play.

DSLE: A Learning Environment for Dark Souls Boss Encounters

Paper: *DSLE: A Learning Environment for Dark Souls Boss Encounters*

Authors: Derin Gezgin, Jim O'Connor, Tanner Goodwin

Published: 2026-08-12

arXiv: 2508.05142

Overview

The paper introduces 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
Sparse terminal rewards

Each environment step is a real action executed against the running live game, distinguishing DSLE from purely simulated RL benchmarks.

DSLE-5: A Recommended Subset

To support controlled comparison, the authors define DSLE-5, a representative five-boss subset covering distinct combat archetypes:

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

DSLE-5 is recommended as the starting suite for any agent built on DSLE.

Evaluations on DSLE-5

Five methods were benchmarked on DSLE-5:

Random policy
Expert system
Evolutionary baseline
PPO (trained from pixels)
DQN (trained from pixels)

Results

| Method | Asylum Demon (tutorial) | Other 4 Bosses | |---|---|---| | Expert system | Peak win rate 63% | 0% | | Evolutionary baseline | Peak win rate 43% | 0% | | PPO | Max 0.33% | 0% | | DQN | No measurable learning | 0% |

Both PPO and DQN exhibited no measurable learning within wall-clock budgets of tens of hours per run.

Broader Sweep Across All 22 Bosses

An evolutionary baseline was run across all 22 boss encounters at level 50 attribute stats:

Won a few additional early bosses
Failed on the remaining bosses

Failure Analysis

Reported via survival time and damage dealt, rather than win rate alone:

Fast deaths: agents died in under 10 seconds in narrow multi-target fights
Stalemates: agents survived nearly one minute while dealing near-zero damage

Why DSLE Matters

DSLE is positioned as a hard, real-game RL benchmark: it removes the simulation-reality gap by acting directly on the live game, yet exposes how far current visual RL methods (PPO, DQN) still are from solving real-time combat tasks with sparse rewards.

--- *Auto-collected 2026-08-12*