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:2610.36013

OmniBrain-0: 身体性知能のための統合エージェントモデル

OmniBrain-0: A Unified Agentic Model for Embodied Intelligence

シェア:XThreadsFacebookLINEはてブBluesky

操作とナビゲーションを単一の軽量な意思決定モデルで統合し、段階的学習と自己改善の仕組みで長期的タスク遂行と失敗からの回復を実現した研究。

詳しい要約

1. どんなもの?

- OmniBrain-0 は manipulation と navigation を単一の high-level decision model で統括する軽量な unified embodied agent。 - task understanding、planning、progress tracking、feedback-driven recovery を統合する。 - 長い horizon での多技能調整、進捗追跡、失敗からの回復を狙う。 - 学習は embodied pretraining、domain-specialized agent distillation、multi-teacher on-policy distillation (MOPD) を段階的に組み合わせる。

2. 先行研究と比べてどこがすごい?

- 単独の specialist policy では task-level reliability を担保できない点を出発点とする。 - 強力な multimodal agent は costly reasoning と external experience に依存し、compact model での再現が難しいと指摘。 - 提案法は統合を tool interface 上ではなく学習された decision model 内で実現する点を特徴とする。 - closed-source teacher より inference と execution の overhead を大幅に低減しつつ cross-domain 性能を両立。

3. 技術・手法の肝は?

- 共有 high-level decision model が manipulation と navigation を統括する構成。 - progressive training: embodied pretraining → domain-specialized agent distillation → MOPD。 - MOPD で teacher の interaction experience を manipulation/navigation expert に蒸留し、その decision 能力を OmniBrain-0 に統合。 - Rubric-Guided Recursive Self-Improvement (RSI): Meta-Agent が execution evidence を refined rubrics、step-level reminders、verified trajectories に変換し、domain-expert 更新後に MOPD で再統合する data flywheel。

4. どうやって有効だと検証した?

- LIBERO-Pro、LIBERO-Plus、RoboCasa365、RoboTwin 2.0、R2R-CE Val-Unseen で評価。 - 複数の real-world manipulation tasks でも評価。 - cross-domain task performance と inference/execution overhead の低減を closed-source teacher と比較して示す。 - 具体的な数値や被験者数は要旨からは不明。

5. 議論はある?

- unification を learned decision model 内で行う設計の有効性と限界は要旨からは不明。 - RSI の data flywheel が性能向上に寄与する度合いや収束性は要旨からは不明。 - teacher 依存の distillation が compact model の汎化や failure recovery に与える影響は要旨からは不明。 - real-world 評価の規模・多様性・安全性に関する議論は要旨からは不明。

6. 次に読むべき論文は?

- multi-teacher on-policy distillation (MOPD) の原論文・関連手法。 - LIBERO-Pro、LIBERO-Plus、RoboCasa365、RoboTwin 2.0、R2R-CE の各 benchmark 論文。 - embodied agent の distillation や on-policy distillation に関する研究。 - rubric-guided self-improvement / data flywheel 関連の LLM agent 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: Xiaomi Embodied Intelligence Team, Yuechen Luo, Chong He, Fang Li, Shaoqing Xu, Jingkai Wang, Menglun Su, Jiaxin Liu, Shengyin Jiang, Zhiyi Lai, Long Chen, Longlong Wang, Liang Hu, Haochen Liu, Yuncheng Jiang, Hongwei Xie, Jieming Cui, Hanbing Li, Junjie Huang, Zhixiang Duan, Hexin Zhang, Weixiang Liang, Ying Gong, Yuhan Wang, Yong Pan, Pengna Li, Kangyi Wu, Ken Deng, Kun Ma, Fuxi Wen, Zhi-xin Yang, Hangjun Ye

分類: cs.CL, cs.CV

原文アブストラクト

General-purpose embodied intelligence requires robots to coordinate diverse skills over long horizons, track task progress, and recover from failures. Specialist policies alone do not ensure task-level reliability, while powerful multimodal agents rely on costly reasoning and external experience that compact models struggle to reproduce. We introduce OmniBrain-0, a lightweight unified embodied agent that coordinates manipulation and navigation through a shared high-level decision model, integrating task understanding, planning, progress tracking, and feedback-driven recovery. Our progressive training framework combines embodied pretraining, domain-specialized agent distillation, and multi-teacher on-policy distillation (MOPD). It distills powerful teachers' interaction experience into manipulation and navigation experts, then consolidates their decision capabilities into OmniBrain-0. Unification therefore occurs within the learned decision model rather than only at the tool interface. We further introduce Rubric-Guided Recursive Self-Improvement (RSI), a data flywheel in which a Meta-Agent turns execution evidence into refined rubrics, step-level reminders, and verified trajectories for domain-expert updates before their reconsolidation into OmniBrain-0 through MOPD. We evaluate OmniBrain-0 on LIBERO-Pro, LIBERO-Plus, RoboCasa365, RoboTwin 2.0, and R2R-CE Val-Unseen, as well as multiple real-world manipulation tasks. OmniBrain-0 combines strong cross-domain task performance with substantially lower inference and execution overhead than its closed-source teacher, demonstrating efficient transfer and consolidation of embodied capabilities into a compact unified agent.

関連論文

PR本紙発行元 EmplifAI