English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

DSLE: A Gymnasium-Style Reinforcement Learning Benchmark Built on Dark Souls Boss Fights

Forum topic · 小凯 · 2026-08-11

Summary

This paper introduces the Dark Souls Learning Environment (DSLE), a containerized reinforcement learning platform that exposes all 22 boss encounters of Dark Souls: Remastered through a Gymnasium-style interface. Each environment step is a real action executed against the running game, combining real-time combat, high-dimensional visual input, and sparse terminal rewards. To enable controlled comparison, the authors define DSLE-5, a five-boss subset covering melee combat, spatially constrained arenas, environmental hazards, multi-target fights, and fast final-boss encounters. They evaluate random, expert-system, evolutionary, PPO, and DQN agents. Expert and evolutionary baselines defeat the Asylum Demon tutorial boss 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 within tens of hours of wall-clock training. A broader evolutionary sweep across all 22 bosses wins only a few early-game encounters. Failure modes are reported via survival time and damage dealt, not just win rate.

Overview

  • Field: Machine Learning / Reinforcement Learning Benchmarks
  • Authors: Derin Gezgin, Jim O'Connor, Tanner Goodwin
  • Released: 2026-08-11
  • arXiv: 2508.03798
  • Summary

    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:

    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

    We recommend DSLE-5 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 evolutionary baseline each defeat the tutorial boss *Asylum Demon* (peak win rates of 63% and 43%, respectively), but none of the five methods defeat the other four bosses in DSLE-5. PPO and DQN show no measurable learning within budgets that already require tens of hours of wall-clock time per run (at most 0.33% win rate on the tutorial boss and 0% elsewhere).

    A broader study runs the evolutionary baseline across all 22 encounters; with an advantageous full level-50 stat build it wins only a handful of additional early-game bosses, with the rest unwon. Failure cases range from deaths in under ten seconds in narrow, multi-target encounters to near-minute-long stalemates with negligible damage dealt. We report these results through survival time and damage dealt rather than win rate alone.

    Key Findings

  • New benchmark: DSLE exposes 22 real-time, high-dimensional, sparsely-rewarded boss fights as a Gymnasium-compatible RL benchmark running against the live game.
  • Recommended starter suite: DSLE-5 covers five distinct combat archetypes for controlled comparison.
  • Baselines tested: random policy, expert system, evolutionary baseline, PPO, and DQN from pixels.
  • Only the tutorial boss is solvable: expert and evolutionary methods reach 63% and 43% peak win rates on *Asylum Demon*; all other four DSLE-5 bosses remain unbeaten.
  • Deep RL does not learn: PPO and DQN achieve at most 0.33% win rate on the tutorial boss and 0% elsewhere, despite tens of hours of wall-clock training.
  • Full sweep is largely unsolved: the evolutionary baseline wins only a few early-game bosses out of all 22 encounters, even with favorable stats.
  • Evaluation beyond win rate: failure is characterized by survival time and damage dealt; some agents survive nearly a minute while dealing almost no damage.
  • Resources

  • arXiv: https://arxiv.org/abs/2508.03798

Tags

#reinforcement-learning#benchmark#dark-souls#gymnasium#ppo#dqn#evolutionary-baselines#game-ai

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633364