日本フィジカルAI新聞

世界のフィジカルAIを、日本語で。

週刊ニュースレター購読
VLAarXiv:2610.19746

人間の動画1本でVLAモデルを新規タスクに適応させるCapPro

One-shot Human Video-Guided Adaptation for Vision-Language-Action Models

シェア:XThreadsFacebookLINEはてブBluesky

ロボット実演データや追加学習なしに、人間の実演動画1本だけで視覚言語行動モデルを未知のマニピュレーションタスクへ適応させるフレームワークを提案。

詳しい要約

1. どんなもの？

- 対象: Vision-language-action (VLA) モデルの新規タスク適応 - 提案: CapPro というフレームワーク - 特徴: 単一の human demonstration video のみで未見の manipulation タスクに適応 - 構成: phase-aware human video representation module と video-conditioned adaptation module - 目的: target-task robot data や fine-tuning への依存を削減

2. 先行研究と比べてどこがすごい？

- 従来: タスク固有の robot demonstrations と追加学習が必要で labor/compute-intensive - 従来の限界: 訓練分布外の新規タスクへの迅速な適応が困難 - 提案の優位: 単一の human video のみで適応可能 - 提案の優位: target-task robot demonstrations や target-task fine-tuning が不要 - 提案の優位: 元の VLA action-generation backbone をほぼ保持

3. 技術・手法の肝は？

- 第一: phase-aware human video representation module - learned visual token selection で compact video representation を構築 - phase-caption supervision の下で task-relevant query representations を学習 - 第二: video-conditioned adaptation module - compact video と query-derived representations を pretrained VLA policy の action-conditioning context に統合 - 設計意図: human demonstration を embodiment-aligned trajectory の模倣ではなく task-specific execution guidance として利用

4. どうやって有効だと検証した？

- 実験設定: source tasks と unseen target tasks で評価 - 結果: unseen target groups で平均性能が向上 - 結果: source tasks では comparable performance を維持 - 結論: human-video guidance が cross-task VLA adaptation に有効であることを示す

5. 議論はある？

- 主張: human-video guidance が cross-task VLA adaptation に有効 - 利点: target-task robot data と fine-tuning への依存を削減し、open-world scenarios への scalability に寄与 - 限界や課題: 要旨からは不明

6. 次に読むべき論文は？

- 要旨で参照/比較されている研究: 明示的な個別研究名は記載なし - 関連手法: Vision-language-action (VLA) models, robot datasets, foundation models - 同分野の定番: 大規模ロボットデータセットと基盤モデルを活用した general-purpose robot manipulation の研究

※ AIが要旨から生成した要約です。正確性は原文をご確認ください。

著者: Qi Chen, Yuwei Wu, Quan Chen, Kunpeng Gao, Jin Chen, Che Sun, Zhi Gao, Yunde Jia

分類: cs.RO

原文アブストラクト

Vision-language-action (VLA) models have recently achieved remarkable progress by leveraging large-scale robot datasets and foundation models for general-purpose robot manipulation. However, their ability to rapidly adapt to novel manipulation tasks beyond the training distribution remains limited. Existing approaches often require task-specific robot demonstrations and additional training, which are labor- and compute-intensive, limiting the scalability of VLA adaptation to open-world scenarios. To reduce this dependence on target-task robot data, we propose CapPro, a framework that enables a VLA policy to adapt to previously unseen manipulation tasks using only a single human demonstration video. CapPro consists of two components. First, a phase-aware human video representation module constructs a compact video representation through learned visual token selection and learns task-relevant query representations under phase-caption supervision. Second, a video-conditioned adaptation module jointly incorporates the compact video and query-derived representations into the action-conditioning context of the pretrained VLA policy. % This design introduces task-specific guidance while largely preserving the original VLA action-generation backbone. In this way, the human demonstration provides task-specific execution guidance rather than an embodiment-aligned trajectory to imitate, without requiring target-task robot demonstrations or target-task fine-tuning. Experiments across source and unseen target tasks show that CapPro improves average performance across unseen target groups while maintaining comparable performance on the source tasks. These results demonstrate the effectiveness of human-video guidance for cross-task VLA adaptation.

arXivで原文を読む →arXiv概要ページ

関連論文

← 論文一覧
PR本紙発行元 EmplifAI