DSLE: A New Benchmark for AI Agents in Dark Souls Boss Fights
A new research initiative has unveiled the Dark Souls Learning Environment (DSLE), which incorporates all 22 challenging boss encounters from Dark Souls: Remastered, designed for AI gameplay agents. This innovative platform mimics a Gymnasium format, facilitating real-time interactions while handling intricate visuals and delivering limited rewards. Among its components, the DSLE-5 focuses on five specific boss fights, offering diverse combat scenarios ideal for training AI systems. Researchers assessed several baseline strategies on DSLE-5, finding that the expert system achieved a 63% success rate against the Asylum Demon, while the evolutionary approach performed at 43%. The findings are published on arXiv under ID 2608.09902.
Key facts
- DSLE is a containerized platform for Dark Souls: Remastered boss encounters.
- It includes all 22 boss encounters as benchmarks.
- Uses a Gymnasium-style interface for agent interaction.
- DSLE-5 is a subset of five representative boss fights.
- Evaluated agents: random policy, expert system, evolutionary baseline, PPO, and DQN.
- Expert system achieved 63% peak win rate against Asylum Demon.
- Evolutionary baseline achieved 43% peak win rate against Asylum Demon.
- Paper available on arXiv (2608.09902).
Entities
Institutions
- arXiv