The game this site plays: Crafter's ideas, NetHack's depth, written in JAX so agents train on billions of steps.
ZzzHold to sleep
Affects model creativity.
| W A S D | walk |
| Shift + W A S D | turn without stepping |
| ← → ↑ ↓ | look around |
| Space | act on the tile ahead |
| 1–0, − | pick a hotbar slot |
| E | inventory and crafting |
| Space or Q | climb a ladder |
| Z | sleep (energy not full) |
| R | rest (health not full) |
| X | place block ahead |
| C V B | 2D, 1P/3P, bird's-eye |
| M | color, phosphor, laser |
| Shift + M | the same, backward |
| F5 | First or third person |
| T | the next theme |
| Shift + T | the same, backward |
| P | agent plays / take over |
| K | sound on or off |
| G | graphics quality |
| H | these controls |
| Esc | pause |
| move | look (mouse captured) |
| drag | look (mouse free) |
| click | hit, mine, drink, open |
| Shift + click | break a table or furnace |
| click, item held | place, shoot or cast |
| wheel | pick a hotbar slot |
| U | capture or free mouse |
| Alt | hold: free the mouse |
| Esc | pause, free the mouse |
| the arrow pad | walk (slide across) |
| Use | act ahead; hold: repeat |
| the pad's middle | Use (phone on its side) |
| Sleep / Zzz | sleep (hold or tap) |
| tap the view | act on what you tap |
| drag the view | look (3/4: swing view) |
| hold on water | keep drinking |
| Full screen | hide the browser bars |
The game this site plays: Crafter's ideas, NetHack's depth, written in JAX so agents train on billions of steps.
Crafter, the 2D survival benchmark Craftax extends: one world, many achievements, scored by how many an agent unlocks.
The roguelike benchmark Craftax borrows its floors, monsters and potions from.
Language and vision-language models playing Crafter, NetHack and others: how far reasoning gets without RL training.
Craftax for teams of agents: the rules we plan to follow for co-op rooms.
Fast RL tooling, and the C port of Craftax our engine is forked from; its released agent, PufferLib 162, plays here.
PPO: how all three agents here were trained.
The gated Transformer-XL: the memory of the TransformerXL PPO agent.
The minGRU: the recurrent layer PufferLib 162 remembers with.
PPO with minGRU memory and a convolutional board encoder, from random weights: beats the final boss in 44% of games, 73% achievement score at 20B steps. The authors published the training code, not the weights; the agent here is a retraining of their recipe.