Definition
Plain language
A small simulated crafting environment in the style of Minecraft, used to test agents on hierarchical tasks.
As stated in the literature
A text-based agent benchmark involving recipe-driven crafting tasks, used as both an evaluation and a tunable testbed for hierarchical agent training.
Also called: TextCraft-Synth
Why it matters: It is a cheap, controllable testbed for hierarchical planning that doesn't require running a heavy game engine.
For example, an agent in TextCraft might need to gather wood, craft a pickaxe, mine stone, and then craft a furnace to complete a recipe.
Heard on the show
“ALFWorld, ScienceWorld, TextCraft — these are PDDL-style simulators where the system internally knows exactly what objects exist, what rooms exist, what actions are valid.”Episode 052 — An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents