3D基盤モデル/SLAM2026/9/18arXiv:2609.21502

適応的世界記憶3D基盤モデル：スケーラブルな3Dマッピング・自己位置推定・レンダリング

Adaptive World Memory 3D Foundation Model for Scalable 3D Mapping, Localization, and Rendering

Transformerベースのゲート付き更新と時空間調整を組み合わせた適応的世界記憶機構により、長期記憶・大規模マッピング・ガウシアンレンダリングを単一モデルで統合した3D基盤モデルを提案。

詳しい要約

1. どんなもの？

- RGB画像から汎化可能な幾何推論を行う3D foundation model - 永続メモリ・スケーラビリティ・レンダリング可能なscene modelingが課題 - 提案はmemory-centricな3D foundation model - scalableなrobotic localization - reconstruction - Gaussian rendering - 単一モデルでcamera pose estimation・dense point-cloud reconstruction・photorealistic renderingを統合

2. 先行研究と比べてどこがすごい？

- 既存3D foundation modelはpersistent memory・scalability・renderable scene modelingが限定的 - 提案はadaptive world memoryで長期画像列の更新・忘却を制御 - local submapsとprogressive mapping/tracking・loop closure・SL(4)-based global refinementを統合 - 局所精度と大域一貫性を両立 - 公開benchmarkと多様なrobotic platformの自収集データで - trajectory accuracy - reconstruction completeness - rendering quality が既存3D foundation reconstruction・SLAM baselineを上回る

3. 技術・手法の肝は？

- 中核はadaptive world memory mechanism - transformer-based gated updates - test-time temporal-spatial regulation - learned gatesがrecurrent memory propagationを制御 - temporal state evolutionとspatial observation-state consistencyが - token-wise updatesとforgettingを長期画像列で調整 - 大規模mappingのためmemoryをlocal submapsに組織 - progressive mapping and tracking・loop closure・SL(4)-based global refinementを統合 - Gaussian reconstruction headがmemory-enhanced featuresをrenderable primitivesへ復号

4. どうやって有効だと検証した？

- 公開benchmarkと多様なrobotic platformからの自収集データで実験 - 評価指標はtrajectory accuracy・reconstruction completeness・rendering quality - 既存の3D foundation reconstructionおよびSLAM baselineと比較し改善を確認 - datasetとcodeは公開予定(URL記載)

5. 議論はある？

- 結果はadaptive memoryがpersistent robotic world modelingの基盤となることを支持 - 限界・失敗事例・計算コスト・メモリ容量などの詳細は要旨からは不明 - 倫理面・安全性・実運用上の議論は要旨からは不明

6. 次に読むべき論文は？

- 要旨で参照/比較されている既存研究 - 3D foundation reconstruction - SLAM baselines - 関連手法 - transformer-based gated updates - Gaussian rendering - SL(4)-based global refinement - 同分野の定番としてSLAM・visual odometry・NeRF/Gaussian Splatting系の基盤研究

※ 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: Tianchen Deng, Guole Shen, Yilin Shen, Wenhua Wu, Yilin Fang, Ziqi Ma, Tianjun Zhang, Shenghai Yuan, Wolfram Burgard, Hesheng Wang

分類: cs.CV, cs.RO

原文アブストラクト

Recent 3D foundation models enable generalizable geometric reasoning from RGB images but remain limited in persistent memory, scalability, and renderable scene modeling. We present a memory-centric 3D foundation model for scalable robotic localization, reconstruction, and Gaussian rendering. Its core is an adaptive world memory mechanism that combines transformer-based gated updates with test-time temporal-spatial regulation. Learned gates control recurrent memory propagation, while temporal state evolution and spatial observation-state consistency regulate token-wise updates and forgetting over long image sequences. To support large-scale mapping, we organize memory into local submaps and integrate progressive mapping and tracking, loop closure, and SL(4)-based global refinement to maintain local accuracy and global consistency. A Gaussian reconstruction head decodes memory-enhanced features into renderable primitives, unifying camera pose estimation, dense point-cloud reconstruction, and photorealistic rendering within a single model. Experiments on public benchmarks and self-collected datasets from diverse robotic platforms demonstrate improved trajectory accuracy, reconstruction completeness, and rendering quality over existing 3D foundation reconstruction and SLAM baselines. These results support adaptive memory as a foundation for persistent robotic world modeling. The dataset and code will be made publicly available at \href{https://github.com/dtc111111/AWM-3DFM}{https://github.com/dtc111111/AWM-3DFM}.

arXivで原文を読む →arXiv概要ページ

← 論文一覧