The NetHackers community, launched by the AIRI Institute, aims to develop an autonomous AI capable of consistently completing the roguelike game NetHack. The project focuses on creating a unified benchmarking system for AI agents, as the game remains a significant challenge due to procedural generation, incomplete information, and long-term planning requirements.
The community uses NetHack version 3.6.6 and the NLE environment for testing. Evaluation is based on 73 starting character classes, each tested across 15 public and 15 private dungeons to prevent overfitting. Results are tracked via a leaderboard that measures progress toward specific game milestones rather than just win rates.
The baseline for the project is AutoAscend, a symbolic bot that won the 2021 NetHack Challenge. While AutoAscend achieved approximately 7.8% progress, recent developments in reinforcement learning and language models have pushed the current record to 32%. Despite these improvements, no autonomous program has yet achieved a consistent ascension.
Participants are encouraged to contribute by improving specific game rules, developing coding agents, or refining existing architectures. To ensure security and reproducibility, all submitted code is executed in isolated containers with resource limits, and results must be linked to public source code repositories.
