shinohararin/minecraft-raytrace

Airicraft Vision Dataset

Egocentric Minecraft frames with per-pixel raycast ground truth, captured by a Fabric mod on Minecraft 1.21.8. Built to train spatially-aware vision models: every label cell carries the exact block/entity hit plus metric depth and egocentric (camera-space) offsets, so a model can learn where things are, not just what they are.

Scale

• 10222 captures across 47 biomes, with day/night and clear/rain/thunder coverage (clear 6885 / rain 1812 / thunder 1525).
• Label grid: 107x60 cells per 854x480 frame (stride 8 px), every cell raycast-verified.
• Voxel region dump per capture (~139k cells) with viewVisible line-of-sight mask — the "privileged modality" for LLaVA-3D-style training.
• Raw labels only — no precomputed task set. Generate your own spatial QA/tasks from the label/region payloads; the reference generator is scripts/generate-spatial-qa in the airicraft repo.

Layout

Everything lives in data/captures-*.parquet (18 shards, ~330 MB each, zstd) — one row per capture:

column	content
capture_id, label	capture dir name / collector label
frame_png	rendered frame bytes (854x480, HUD stripped)
meta_json	camera pose, basis vectors, projectionMatrixRowMajor, biome, weather, time, lighting tag (natural | nightvision | torch)
labels_gz	gzip'd label grid JSON. Block cells: blockId, stateKey, blockX/Y/Z, depth, egoForward/egoRight/egoUp, hitLight. Alpha-aware (fern/leaf/glass texels pass through); unrendered sections count as air, so labels always match pixels
region_gz	gzip'd voxel dump around camera with viewVisible LOS mask
entities_json	entities with projected screen rects + hit cells
biome, lighting, world_time, raining, mean_luminance, fov	duplicated scalar columns for cheap filtering without parsing JSON

• captures.jsonl — same index rows (id, label, stats incl. settleWaitMs)
• parquet/captures.parquet — flattened index, no payloads

import pyarrow.parquet as pq, gzip, json, io
from PIL import Image
t = pq.read_table("data/captures-000.parquet")
row = t.slice(0, 1).to_pylist()[0]
img = Image.open(io.BytesIO(row["frame_png"]))
labels = json.loads(gzip.decompress(row["labels_gz"]))

Caveats

• Captures with camera inside a block or submerged are skipped at capture time; lighting torch frames place glowstone blocks (tagged, removable).
• dark natural frames are intentionally kept — filter by lighting or mean_luminance when training.
• Egocentric offsets are in meters, Minecraft convention (y-up, forward along view).