MiniHack the Planet: A Sandbox for Open-Ended Reinforcement Learning Research

TL;DR

MiniHack introduces a flexible sandbox for open-ended RL research built on NetHack's rich assets and its domain-specific language (des-file). It provides two main interfaces—MiniHackNavigation and MiniHackSkill—along with a LevelGenerator and RewardManager to rapidly design and extend tasks, including the porting of existing benchmarks. The authors validate MiniHack through navigation and skill acquisition experiments, unsupervised environment design with PAIRED, architecture comparisons, and RLlib integrations, revealing both the potential and the current limits of standard baselines on more complex, realistic tasks. The framework enables controlled scaling of difficulty, multi-modal observations, and systematic generalization testing, offering a practical platform for advancing RL research and benchmarking.

Abstract

Progress in deep reinforcement learning (RL) is heavily driven by the availability of challenging benchmarks used for training agents. However, benchmarks that are widely adopted by the community are not explicitly designed for evaluating specific capabilities of RL methods. While there exist environments for assessing particular open problems in RL (such as exploration, transfer learning, unsupervised environment design, or even language-assisted RL), it is generally difficult to extend these to richer, more complex environments once research goes beyond proof-of-concept results. We present MiniHack, a powerful sandbox framework for easily designing novel RL environments. MiniHack is a one-stop shop for RL experiments with environments ranging from small rooms to complex, procedurally generated worlds. By leveraging the full set of entities and environment dynamics from NetHack, one of the richest grid-based video games, MiniHack allows designing custom RL testbeds that are fast and convenient to use. With this sandbox framework, novel environments can be designed easily, either using a human-readable description language or a simple Python interface. In addition to a variety of RL tasks and baselines, MiniHack can wrap existing RL benchmarks and provide ways to seamlessly add additional complexity.

Figures (32)

• Figure 1: Examples of procedurally generated environments using the des-file format. (Top): MAZEWALK command applied on a 15x15 grid, (Middle) corridors generated via RANDOM_CORRIDOR, (Bottom): environments generated using the code snippet from \ref{['code:des_procgen']}.
• Figure 2: A sample code snippet in des-file format language. The $river variable is used to sample a terrain feature ('L' for lava, 'W' for water and 'I' for ice). The LOOP block draws two rivers via the randline command and places two random monsters at random locations. The REPLACE_TERRAIN commands replaces 5% of floors ('.') with trees ('T'). A stair down is added at random locations.
• Figure 3: A des-file example for a simple NetHack level.
• Figure 4: Different forms of agent-centred observations of the grid of the map in MiniHack.
• Figure 5: A sample code snippet for creating a custom MiniHack task using the LevelGenerator and RewardManager.