README.md

metadata

license: mit
task_categories:
  - other
tags:
  - gesture-control
  - minecraft
  - body-pose
  - crowdsourced
pretty_name: MCGestureControl Crowdsourced Data

MCGestureControl Crowdsourced Data

Training data for the MCGestureControl project — an AI system that plays Minecraft using body gestures captured by a phone camera.

Dataset Structure

minecraft/    — Gameplay telemetry NPZ files (game_state, controls, timestamps)
body/         — Body skeleton NPY files (260-dim: pose + hands per frame)
paired/       — Paired gameplay + body recordings

Data Formats

Minecraft Gameplay (NPZ)

• controls — (T, 28) float32: resolved game controls per frame
• game_state — (T, 46) float32: game state vectors per frame
• timestamps — (T,) float64: unix timestamps
• frame_indices — (T,) int32: corresponding video frame indices
• fps — float32: recording framerate (typically 30)
• version — int32: recorder version
• recorder_id — string: anonymous machine identifier

Body Skeleton (NPY)

• Shape: (T, 260) float32
• Layout: pose_world(99) + left_hand_3d(63) + right_hand_3d(63) + left_present(1) + right_present(1) + pose_visibility(33)

Controls Layout (28 dims)

Movement (0-6), Attack/Use (7-8), Look yaw/pitch (9-10), Utility (11-13), Hotbar one-hot (14-22), Cursor (23-24), Inventory clicks (25-27)

Contributing

Data is collected via the MCGestureControl crowdsourcing portal. Contributors register, accept a data agreement, and upload recordings.

Privacy

• Raw body video is auto-processed into skeleton data and deleted
• Skeleton data contains no personally identifiable information
• Contributor IPs are stored as irreversible SHA-256 hashes