Datasets:
json dict | __key__ stringlengths 25 30 | __url__ stringclasses 2
values |
|---|---|---|
{
"episode_id": 165,
"instruction": "Navigate to the 木製引き出し, then orbit it once while recording.",
"steps": [
{
"step_id": 8958,
"episode_id": 165,
"image_path": "steps/step_008958.jpg",
"instruction": "Navigate to the 木製引き出し, then orbit it once while recording.",
"action_vector"... | dataset/episodes/episode_00165 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 20,
"instruction": "Approach the 木製引き出し, fly around it, and take photos.",
"steps": [
{
"step_id": 1066,
"episode_id": 20,
"image_path": "steps/step_001066.jpg",
"instruction": "Approach the 木製引き出し, fly around it, and take photos.",
"action_vector": [
-0.2... | dataset/episodes/episode_00020 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 98,
"instruction": "Navigate to the アームチェア, then orbit it once while recording.",
"steps": [
{
"step_id": 5317,
"episode_id": 98,
"image_path": "steps/step_005317.jpg",
"instruction": "Navigate to the アームチェア, then orbit it once while recording.",
"action_vector": ... | dataset/episodes/episode_00098 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 132,
"instruction": "Go close to the ソファ and circle around it to observe.",
"steps": [
{
"step_id": 7152,
"episode_id": 132,
"image_path": "steps/step_007152.jpg",
"instruction": "Go close to the ソファ and circle around it to observe.",
"action_vector": [
0.... | dataset/episodes/episode_00132 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 77,
"instruction": "Fly toward the アームチェア and do a full loop around it.",
"steps": [
{
"step_id": 4203,
"episode_id": 77,
"image_path": "steps/step_004203.jpg",
"instruction": "Fly toward the アームチェア and do a full loop around it.",
"action_vector": [
-0.833... | dataset/episodes/episode_00077 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 124,
"instruction": "Approach the 木製引き出し, fly around it, and take photos.",
"steps": [
{
"step_id": 6723,
"episode_id": 124,
"image_path": "steps/step_006723.jpg",
"instruction": "Approach the 木製引き出し, fly around it, and take photos.",
"action_vector": [
0.... | dataset/episodes/episode_00124 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 61,
"instruction": "Navigate to the アームチェア, then orbit it once while recording.",
"steps": [
{
"step_id": 3336,
"episode_id": 61,
"image_path": "steps/step_003336.jpg",
"instruction": "Navigate to the アームチェア, then orbit it once while recording.",
"action_vector": ... | dataset/episodes/episode_00061 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 173,
"instruction": "Approach the ソファ, fly around it, and take photos.",
"steps": [
{
"step_id": 9377,
"episode_id": 173,
"image_path": "steps/step_009377.jpg",
"instruction": "Approach the ソファ, fly around it, and take photos.",
"action_vector": [
-0.49845... | dataset/episodes/episode_00173 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 36,
"instruction": "Navigate to the アームチェア, then orbit it once while recording.",
"steps": [
{
"step_id": 1950,
"episode_id": 36,
"image_path": "steps/step_001950.jpg",
"instruction": "Navigate to the アームチェア, then orbit it once while recording.",
"action_vector": ... | dataset/episodes/episode_00036 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 221,
"instruction": "Navigate to the ソファ, then orbit it once while recording.",
"steps": [
{
"step_id": 11983,
"episode_id": 221,
"image_path": "steps/step_011983.jpg",
"instruction": "Navigate to the ソファ, then orbit it once while recording.",
"action_vector": [
... | dataset/episodes/episode_00221 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
{
"episode_id": 108,
"instruction": "Fly toward the 木製引き出し and do a full loop around it.",
"steps": [
{
"step_id": 5881,
"episode_id": 108,
"image_path": "steps/step_005881.jpg",
"instruction": "Fly toward the 木製引き出し and do a full loop around it.",
"action_vector": [
0.92... | dataset/episodes/episode_00108 | hf://datasets/furuharu/drone-furniture-fly-around-dataset-v2@600e296afb78df74c8537376fbda5caf8ca239fa/episodes.tar.gz |
End of preview. Expand in Data Studio
drone-furniture-fly-around-dataset
本データセットは、関西電力KOI×VOLTMIND 生成AIハッカソンで使用しました。
Genesis 物理シミュレーターで自動生成した、ドローン VLA(Vision-Language-Action)学習用合成データセット。 FPV カメラ画像・自然言語命令・ドローン制御アクションの三つ組みで構成される。
データセット概要
室内に配置された3種類の家具(ソファ・アームチェア・木製引き出し)に対して、 PID コントローラーで自動生成した「接近 → 周回 → ホバー」軌跡を記録したデータセット。 OpenVLA 7B の LoRA ファインチューニング用途に設計されている。
データの内容
規模
| 項目 | 値 |
|---|---|
| エピソード数 | 300 |
| 総ステップ数 | 約 16000 |
| 記録周波数 | 10 Hz(シミュレーターは 100 Hz、10 ステップに 1 回記録) |
| 画像解像度 | 224 × 224 px (RGB) |
| アクション次元 | 4(学習時に 7D へゼロ拡張) |
フォーマット
dataset/
episodes/
episode_00000.json ← エピソードメタデータ
episode_00001.json
...
steps/
step_000000.jpg ← FPV カメラ画像(JPEG, quality=95)
step_000001.jpg
...
dataset_info.json ← データセット全体の統計情報
カラム説明
episodes/episode_XXXXX.json
| フィールド | 型 | 説明 |
|---|---|---|
episode_id |
int | エピソード連番(0-indexed) |
instruction |
str | 自然言語命令 |
target_name |
str | ターゲットオブジェクト名 |
target_pos |
list[float] | ターゲットの3D座標 [x, y, z](メートル) |
success |
bool | エピソードが正常完了したか |
steps |
list[dict] | 各ステップのデータ(下記参照) |
steps(各エピソード内)
| フィールド | 型 | 説明 |
|---|---|---|
step_id |
int | 全エピソード通しのステップ連番 |
episode_id |
int | 所属エピソードID |
image_path |
str | 対応する JPEG 画像のパス(steps/step_XXXXXX.jpg) |
instruction |
str | 自然言語命令(エピソード内で全ステップ共通) |
action_vector |
list[float] | 4次元アクション [vx_body, vy_body, vz_body, yaw_rate] |
phase |
str | フライトフェーズ(approach / orbit / hover / done) |
is_first |
bool | エピソードの最初のステップか |
is_last |
bool | エピソードの最後のステップか |
is_terminal |
bool | 正常終了ステップか |
アクション形式(4次元・ドローンボディフレーム)
| インデックス | フィールド | 内容 | 単位 |
|---|---|---|---|
| 0 | vx_body |
機首方向の速度(前進) | m/s |
| 1 | vy_body |
機体左方向の速度 | m/s |
| 2 | vz_body |
上方向の速度 | m/s |
| 3 | yaw_rate |
ヨー角速度(回転) | rad/s |
学習スクリプト(
train.py)ではゼロ埋めして OpenVLA 互換の 7 次元に拡張されます。
dataset_info.json
| フィールド | 型 | 説明 |
|---|---|---|
n_episodes |
int | 総エピソード数 |
n_steps |
int | 総ステップ数 |
img_size |
int | 画像解像度(正方形) |
action_dim |
int | アクション次元数(4) |
action_format |
str | アクション形式の説明文字列 |
success_rate |
float | エピソード成功率(0〜1) |
データの取得元(出典)
すべてのデータは Genesis 物理シミュレーター で自動生成した合成データです。実世界データは含みません。
| 素材 | 出典 | ライセンス |
|---|---|---|
| シミュレーション環境 | Genesis-Embodied-AI/Genesis | Apache 2.0 |
| アームチェア 3D モデル | modern_arm_chair_01_4k.glb(Poly Haven) | CC0 1.0 |
| ソファ 3D モデル | sofa_02_4k.glb(Poly Haven) | CC0 1.0 |
| 木製引き出し 3D モデル | vintage_wooden_drawer_01_4k.glb(Poly Haven) | CC0 1.0 |
作成方法
収集方法
scripts/collect_v2.py によって以下のパイプラインで自動生成:
1. オブジェクト選択(3種からランダム)
2. オブジェクトをランダム位置に配置
3. ドローンをランダム位置にスポーン
4. 命令文を 4 テンプレートからランダム選択
5. PID コントローラーでターゲットへ自動接近
6. ターゲット到達後、周囲を一周回る
7. 周回完了後、数ステップホバー
8. FPV 画像・アクション・メタデータを 10 Hz で記録
命令文テンプレート
| テンプレート |
|---|
"Approach the {name}, fly around it, and take photos." |
"Go close to the {name} and circle around it to observe." |
"Navigate to the {name}, then orbit it once while recording." |
"Fly toward the {name} and do a full loop around it." |
{name} には arm chair / sofa / wooden drawer 等のオブジェクト名が入ります。
※ 現状、日本語でオブジェクト名が入っているため、修正が必要です。
アノテーション方法
アノテーションは人手なし。アクションは PID コントローラーが算出したワールド座標系速度を、ドローンのヨー角でボディフレームに変換して自動生成しています。
シミュレーション設定
| 設定 | 値 |
|---|---|
| シミュレーター dt | 0.01 s(100 Hz) |
| 記録周波数 | 10 Hz |
| 重力 | −9.81 m/s² |
| ドローンモデル | Crazyflie 2x(cf2x.urdf) |
| カメラ FOV | 90° |
| 最大速度(接近) | 1.5 m/s |
| 周回角速度 | 2.0 rad/s |
前処理内容
- FPV 画像を JPEG(quality=95)で保存
- アクションをワールド座標系 → ドローンボディフレームに変換
- 学習時に 4 次元アクションをゼロ埋めで 7 次元(OpenVLA 互換)に拡張
想定ユースケース
- VLA モデルのファインチューニング(特に OpenVLA 7B + LoRA)
- ドローンナビゲーションの模倣学習研究
- シミュレーターベースのロボット学習のベースライン構築
- Vision-Language-Action アーキテクチャの実験・検証
注意点・制約
バイアス
| バイアスの種類 | 内容 |
|---|---|
| オブジェクトバイアス | 3種類の家具のみ。他オブジェクトへの汎化は保証されない |
| 軌跡バイアス | すべて「接近→周回→ホバー」の固定パターン。多様な行動は含まない |
| 環境バイアス | 単一オブジェクト × 平坦な床面のシンプルなシーンのみ |
| 照明バイアス | Genesis のデフォルト照明のみ。実世界の照明変動は未考慮 |
| 命令バイアス | 4 テンプレートのみ。自由文・複雑な命令は含まない |
欠損・品質
- エピソードによっては
success=false(タイムアウト)が含まれる場合がある - 画像は JPEG 圧縮のため微細なテクスチャ情報に損失がある
- アクション値は PID 制御の近似値であり、物理的に最適な軌跡ではない
利用上の注意
- 本データセットは Genesis シミュレーター内専用の合成データです
- 実機ドローン制御への直接適用は Sim-to-Real ギャップがあるため非推奨です
- 安全性が求められる用途(実機飛行・人の近くでの運用など)には使用しないでください
ライセンス / 利用条件
MIT
使用している 3D モデル(Poly Haven)はすべて CC0 1.0 です。 Genesis シミュレーターは Apache 2.0 です。
関連リンク
- ファインチューニング済みモデル: furuharu/openvla-7b-finetuned-drone-furniture-fly-around
- ベースモデル: openvla/openvla-7b
- Genesis シミュレーター: Genesis-Embodied-AI/Genesis
- Downloads last month
- 59