moincraft

๐Ÿค MineCollab

Collaborative embodied-agent benchmark covering cooking, construction, crafting, evaluation, and parallel worlds.

MineCollab is the collaborative-task benchmark for embodied LLM agents in Minecraft.

Task families#

Task Challenge Evaluation
Cooking Shared recipe planning, ingredient gathering, and coordination. Binary success/failure.
Construction Procedural blueprint execution with incomplete knowledge/resources. Edit-distance style score.
Crafting Multi-step recipe chains with partial resource/recipe distribution. Binary success/failure.

Benchmark posture#

  • no single agent has complete information or resources;
  • communication is part of the task, not an optional extra;
  • environments and goals can be randomized between episodes;
  • harder variants split recipe/build knowledge across agents.

Task file shape#

Task JSON captures a goal, inventory, agent count, target item/block, blocked actions, crafting-table requirements, depth, and timeout.

Running tasks#

  • node main.js --task_path ... --task_id ... for basic single-agent runs;
  • Python evaluation flows for multi-agent crafting/cooking/construction;
  • --num_parallel <N> for parallel worlds when capacity allows.

Operational cautions#

  • construction can require insecure coding, so only enable it intentionally;
  • parallel worlds multiply server/process load quickly;
  • evaluation outputs are experimental artifacts, not public wiki content.

Sources: wiki/content/docs/minecollab.md.