SKU RLE-1503 · Sold by External
NetHack Learning Environment (NLE)
Product specifications
| SKU | RLE-1503 |
|---|---|
| Data type | RL environments |
| Volume | 9 registered Gymnasium environments (base NetHack + 8 task variants) |
| Size on disk | Not published by source |
| Format | Gymnasium (Gym) reinforcement-learning environment (Python package with C/NetHack backend) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license |
| Quality score | — |
| License | NetHack General Public License (NGPL) |
The NetHack Learning Environment (NLE) is an open-source RL environment, first presented at NeurIPS 2020, that turns the procedurally-generated dungeon game NetHack into a standard Gymnasium interface. This version is built on NetHack 3.6.7 and ships nine registered environments: a base 'NetHack-v0' plus eight goal-oriented task variants (Score, Staircase, StaircasePet, Oracle, Gold, Eat, Scout, and the full-keyboard Challenge). Observations are symbolic arrays of the game state - glyph/character/color grids of the 21x79 dungeon map, a 27-value bottom-line stats vector, the current 256-byte message, inventory arrays, and the raw 24x80 terminal (tty) view. The task environments use a 23-action discrete space (message dismissal, 8 compass directions, 8 'longer/run' compass directions, up/down/wait, and kick/eat/search), while the base and Challenge envs use the full NetHack keyboard action set. Each task defines its own programmatic reward function (dense score/gold/exploration/hunger deltas or a sparse +1 goal signal) plus an optional time-step penalty for state-unchanging steps. NLE and NetHack are distributed under the NetHack General Public License.