AZH04/embodied-memory-minecraft-demos

Embodied Memory: Minecraft Expert Demonstrations

First-person Minecraft demonstrations for training an action model: a 640x360 RGB frame goes in and a short text action comes out. A scripted expert played 13,780 episodes and made 346,918 decisions. Every episode has a short text goal.

The expert reads privileged game state to choose its actions. Frames and goal text are the intended policy input. teacher_json is the game state after each decision, kept for analysis and labeling only.

Quick start

Stream rows with datasets:

from datasets import load_dataset

rows = load_dataset("AZH04/embodied-memory-minecraft-demos", "skills_click", split="train", streaming=True)
row = next(iter(rows))
row["png"]              # the frame the expert saw
row["json"]["goal"]     # "I find a nearby tree and collect 2 logs from it."
row["json"]["label"]    # "turn_right 90"

Or download and rebuild episode directories, which is the layout our trainer reads. unpack.py needs only the Python standard library:

hf download AZH04/embodied-memory-minecraft-demos --repo-type dataset --local-dir demos \
  --include "data/skills_click/*" --include "metadata/*" --include "unpack.py"
python demos/unpack.py --data demos --configs skills_click --output episodes

This gives episodes/skills_click/<seed>/<episode>/ with observations/NNNNNN.png, samples.jsonl, last_frame.png and episode.json. An episode the expert failed has episode_failed.json instead.

Configs

Config	Episodes	Expert failures	Decisions	Size
skills_click	4,300	0	46,322	3.8 GB
arena_click	600	0	8,796	0.7 GB
natural_click	562	139	38,395	4.2 GB
mine_blocks	360	0	3,952	0.3 GB
choice_tree_block	600	0	7,354	0.8 GB
choice_two_blocks	800	0	6,987	0.6 GB
choice_item_build	1,600	0	14,647	1.2 GB
choice_more_blocks	360	0	3,039	0.3 GB
chain_two_objects	300	0	7,098	0.7 GB
arena_pointer	1,451	0	30,121	2.7 GB
arena_pointer_dagger	450	0	7,756	0.7 GB
natural_pointer	2,241	540	157,601	17.4 GB
natural_pointer_dagger	156	0	14,850	2.0 GB
identify_qa	2,000 scenes		24,571 text pairs	0.9 GB

The _click and scene configs share one action format, so they can be combined. The _pointer configs are older: they craft by moving the pointer with camera actions instead of click X Y. Their movement and mining labels match the newer format, but their crafting labels do not.

Config	What it shows
skills_click	Single skills in a flat arena: look, walk, mine, climb, gather logs, planks, table, wood, table chain, mine a block, pick up an item, stairs, wall, pillar, plus crafting from a recipe catalog.
arena_click	Arena planks, table and table chain.
natural_click	Natural worlds: find a tree, collect wood, craft a table.
mine_blocks	Break a block of a named material.
choice_tree_block	A tree and a block in one scene. The goal says which to collect.
choice_two_blocks	Two different blocks in one scene. The goal names the target.
choice_item_build	An item and a block, or blocks in hand to build a row or a pillar. The goal decides.
choice_more_blocks	More tree-and-block scenes, always collecting the block, for four materials.
chain_two_objects	Both objects in a named order: "I do A, then I do B."
arena_pointer, arena_pointer_dagger	Older arena demonstrations, and expert labels on states a model reached (DAgger).
natural_pointer, natural_pointer_dagger	Older natural-world demonstrations (wild_table, wild_wood), and DAgger.
identify_qa	Text questions about scene screenshots, such as "Where is the pig?". Not actions, and not a load_dataset config: see below.

To read identify_qa, load the questions as JSON. Each row's images are member names in the tar shards under extras/identify/.

qa = load_dataset(
    "json",
    data_files="hf://datasets/AZH04/embodied-memory-minecraft-demos/metadata/identify_qa.jsonl",
    split="train",
    streaming=True,
)

Row fields

Each decision is one row of a WebDataset shard: png is the frame and json holds the following.

Field	Meaning
config, episode, step	Config name, episode path and decision number. A row with a null step is the episode's last frame, with no action.
frame	The frame's path inside the episode directory.
task, goal	Task name and the episode's goal text.
expert_success	Whether the expert finished the task. Failed episodes are included and flagged.
label	What the expert chose: the training target.
executed	What ran. It differs from label when noise replaced the action, or in _dagger configs, where a model acted and label is what the expert would have done.
noise, source	Whether this decision was a random action, and who acted: expert, policy (a model) or null in older data.
subgoal, note	The expert's own description of the step. Not a policy input.
teacher_json	Game state after the decision, as a JSON string. Fields vary by task.

Episode-level details are in metadata/<config>.jsonl, and collection settings in metadata/<config>_collections.jsonl. Counts per task are in metadata/stats.json.

Action format

Each row is one decision that can last several game ticks. Only the frame before each decision is saved, so these labels can't be turned into per-tick MineRL actions without losing the frames in between.

Label	Effect
turn_left 30, turn_right 30, look_up 20, look_down 20	Rotate by degrees. Can be combined with and.
forward; ticks 10, back, left, right	Move for up to 20 game ticks (default 10).
jump and forward, jump and use	Buttons pressed together.
attack; ticks 20, use; ticks 1	Break a block, or place a block.
inventory, hotbar.3	Open or close the inventory, or select a hotbar slot.
click 248 247, right_click 338 123	Click a pixel of the 640x360 frame while a screen is open.
no_op	Do nothing. Older configs only.

How it was collected

• Expert: a script that knows object positions and the inventory. Crafting uses fixed recipes.
• Noise: each skills_click episode draws a noise level of 0, 0.15 or 0.3, the chance that a decision is replaced by a random action so that demonstrations include recovery. Noise is off while a screen is open, so there are no crafting mistakes or recoveries. Other sets have their own settings in metadata/<config>_collections.jsonl.
• Worlds: arena scenes are built by commands, and natural episodes use generated Minecraft terrain. No task involves combat.
• Held-out test pool: in skills_click, the craft_heldout task holds recipes kept out of our training. Exclude it if you want the same split.

Limitations

• This is scripted play, not human play, so behavior is narrow and direct.
• Play stops at a crafting table. There is no mining with tools, no stone progression and no combat, and building is limited to rows and pillars.
• The recipe catalog hands the expert its ingredients and has fewer than 2 episodes per recipe.
• About a third of successful natural episodes start with a tree more than 10 blocks away.
• The expert fails some natural episodes, more often when the tree is far. Those episodes are flagged and our trainer skips them.

License and notice

Labels, metadata and code in this repository are released under CC BY-NC 4.0. Frames are Minecraft screenshots, and their game content remains subject to Mojang's and Microsoft's terms.

Not an official Minecraft product. Not approved by or associated with Mojang or Microsoft.