Luanti navigation episodes

Agent-navigation episodes recorded in Luanti (formerly Minetest) through the Craftium environment interface. Every episode pairs a first-person RGB video with the voxel grid around the player, the camera pose and the actions, all aligned per step.

Files

File	Environment	Episodes	Size	Unpacked
motworld-s1nav-10k.tar.zst	OpenWorldCreative-v0	9,987	66.3 GB	109 GiB
motworld-hudnav-20k.tar.zst	OpenWorldCreativeHud-v0 (HUD rendered in the frame)	19,977	114.6 GB	211 GiB
motworld-s1nav-10k.splits.json	split labels for the archive above	9,987	1.2 MB	-
motworld-hudnav-20k.splits.json	split labels for the archive above	19,977	2.5 MB	-

Both are plain tar archives compressed with zstd (level 10, no long-window mode, so no special flags are needed to decompress).

tar --zstd -xf motworld-s1nav-10k.tar.zst      # -> OpenWorldCreative-v0/<seed>/...
tar --zstd -xf motworld-hudnav-20k.tar.zst     # -> OpenWorldCreativeHud-v0/<seed>/...

# or stream a few episodes without unpacking everything
zstd -dc motworld-s1nav-10k.tar.zst | tar -xf - --wildcards 'OpenWorldCreative-v0/10001*'

SHA-256 of the archives:

5885f8410f33ed35cf7ae5f9a9558eea7dcfa1330e95f063f0242bb82e0f9e83  motworld-s1nav-10k.tar.zst
e8de6598189ac9d992f2360f6e406b229e82f63675afdc75292a7a9ebc01e0b7  motworld-hudnav-20k.tar.zst

Train / eval splits

The archives contain every recorded episode and no split information; the split lives in the two *.splits.json files, one entry per episode directory:

{"name": "1000149598", "sha256": "8009d1a8f239...", "split": "train"}

name is the episode directory inside the archive, sha256 is its unique id (also stored in the episode's sha256.txt), and split is one of:

Split	s1nav-10k	hudnav-20k	Meaning
train	9,463	19,545	training episodes
eval	200	200	held-out evaluation episodes (the ones that also ship a voxel_t0.glb)
unused	324	232	recorded but excluded from both splits

import json
s = json.load(open("motworld-s1nav-10k.splits.json"))
eval_dirs = [f"{s['environment']}/{e['name']}" for e in s["episodes"] if e["split"] == "eval"]

The two archives are generated from overlapping world seeds, and each has its own split: 103 of the s1nav-10k eval seeds are train seeds in hudnav-20k (same world, different trajectory). Filter on name if you train on one archive and evaluate on the other.

Episode layout

Each archive holds one directory per episode, named by the world seed:

OpenWorldCreative-v0/
  1000149598/
    rgb.mp4               # H.264, 640x360, 24 fps, 600 frames (25 s)
    data.npz              # per-step arrays, see below (numpy, compressed)
    level_metadata.json   # seed, fov_x, fov_y, spawn_pos, minetest_conf
    sha256.txt            # the episode's unique id (same value as `sha256` in the split files)
    voxel_t0.glb          # mesh export of the step-0 voxel grid (only in 200 episodes per archive)

Every episode has T = 600 steps; frame t of rgb.mp4 corresponds to index t of every array.

data.npz

Key	dtype	Shape	Meaning
obs_voxel_mt	int16	(T, 49, 49, 49, 2)	voxel grid around the player (radius 24 nodes per axis); last axis = node content id, param2
obs_voxel_center	int64	(T, 3)	world coordinate of the grid centre
action	bool	(T, 23)	Craftium key/mouse action vector
player_pos, player_vel	float64	(T, 3)	player position / velocity
player_pitch, player_yaw	float64	(T,)	player orientation, degrees
cam_pos, cam_dir	float64	(T, 3)	camera position (world) / view direction
cam_pos_local	float64	(T, 3)	camera position in the grid-local frame (normalised units)
fov_x, fov_y	float64	(T,)	field of view, degrees
intrinsics	float32	(T, 3, 3)	pinhole camera matrix
extrinsics_global, extrinsics_local	float32	(T, 4, 4)	camera pose in world / grid-local coordinates
timestep_craftium	int64	(T,)	environment step counter
dt_minetest	float64	(T,)	engine time delta per step
termination_flag, truncation_flag	bool	(T,)	episode-end flags

Node content ids in obs_voxel_mt are assigned by the engine at runtime, so they are consistent within this dataset but are not a fixed public enumeration.

import numpy as np
ep = np.load("OpenWorldCreative-v0/1000149598/data.npz")
voxels = ep["obs_voxel_mt"]        # (600, 49, 49, 49, 2)
pose   = ep["extrinsics_global"]   # (600, 4, 4)