VLA2026/9/14arXiv:2609.15169

GRAVA: 自動運転のための接地推論から行動への表現と学習

GRAVA: Grounded Reasoning-to-Action Representation and Learning for Autonomous Driving

自動運転VLAモデルにおいて、推論を物理的な場面証拠に接地させ、実行可能な行動と結びつけるGRAフレームワークを提案し、NAVSIMベンチマークで最先端性能を達成した。

詳しい要約

1. どんなもの？

- 自動運転向けのVLAモデルGRAVAを提案 - Grounded Reasoning-to-Action (GRA)を中核に据える - grounding・reasoning・action生成を単一のautoregressive streamで統合 - 言語参照を2D視覚領域とego-centric物理状態に結びつける - 軌跡にanchorしたtyped graphで物体相互作用と決定を構造化 - 単一VLMがreasoningとExecutable Planner actionを生成 - actionは決定論的に連続軌跡へdecodeされる

2. 先行研究と比べてどこがすごい？

- 従来のdriving VLAは中間reasoningが物理scene evidenceに弱くgroundされ、実行可能なbehaviorとの結びつきも緩い - GRAVAはgrounding・reasoning・actionを単一autoregressive streamで統合 - action-relevantな言語参照を2D視覚領域とego-centric物理状態に明示的にリンク - 純粋にautoregressiveなdrivingモデルの中でNAVSIM full benchmarkのSOTAを達成 - 人間のdriving demonstrationの約60%のみをaction supervisionに使用 - 内部long-tail benchmarkでkey-object complianceとClosed-loop Driving Scoreをaction-only比で19.3%・20.5%改善

3. 技術・手法の肝は？

- Grounded Reasoning-to-Action (GRA)を提案 - action-relevantな言語参照を2D視覚領域とego-centric物理状態にリンク - 物体相互作用と決定をtrajectory-anchored typed graphで整理 - この構造をgrounded reasoningとしてserialize - 単一VLMがreasoningに続きcompactなExecutable Planner actionを生成 - actionは決定論的に連続trajectoryへdecode - agentic GRA data construction pipelineを導入 - forward scene groundingとbackward trajectory anchoringを組み合わせ - GR-NavSimを構築(2.2M grounded QA pairs、70K GRA reasoning traces) - progressive training: pre-trainingでgrounded cognition、imitatio…

4. どうやって有効だと検証した？

- NAVSIM full benchmarkで評価 - 純粋にautoregressiveなdrivingモデルの中でSOTAを達成 - 人間のdriving demonstrationの約60%をaction supervisionに使用 - 内部long-tail benchmarkで評価 - full GRAはaction-only prediction比でkey-object complianceを19.3%改善 - Closed-loop Driving Scoreを20.5%改善

5. 議論はある？

- 要旨からは不明 - 限界や失敗事例、計算コスト、一般化性に関する議論は要旨に記載なし

6. 次に読むべき論文は？

- 要旨で参照/比較されている研究は明示されていない - 関連手法としてdriving VLA、autoregressive driving models、NAVSIM benchmark、Executable Planner、GR-NavSimが挙げられる - 同分野の定番としてvision-language-action models、imitation learning、reinforcement learning、closed-loop driving evaluationが考えられる

※ 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: Xiao Liu, Haoyu Li, Jianghao Leng, Lin Wang, Chao Sun

分類: cs.CV, cs.RO

原文アブストラクト

Driving vision-language-action (VLA) models increasingly reason before acting, but their intermediate reasoning is often weakly grounded in physical scene evidence and loosely connected to executable behavior. We present GRAVA, a framework built around Grounded Reasoning-to-Action (GRA), which unifies grounding, reasoning, and action generation in a single autoregressive stream. GRA links action-relevant language references to 2D visual regions and ego-centric physical states, organizes object interactions and decisions in a trajectory-anchored typed graph, and serializes this structure into grounded reasoning. A single VLM generates this reasoning followed by a compact Executable Planner action that is deterministically decoded into a continuous trajectory. We further introduce an agentic GRA data construction pipeline that combines forward scene grounding with backward trajectory anchoring, and use it to build GR-NavSim with 2.2M grounded question-answer pairs and 70K GRA reasoning traces. A progressive training strategy develops grounded cognition through pre-training, establishes the reasoning-to-action interface through imitation, and improves driving behavior through reinforcement learning and exploration. Using about 60% of the available human driving demonstrations for action supervision, GRAVA-8B achieves state-of-the-art performance among purely autoregressive driving models on the full NAVSIM benchmark. On an internal long-tail benchmark, full GRA improves key-object compliance and Closed-loop Driving Score by 19.3% and 20.5% over action-only prediction, respectively. These results show the benefit of preserving action-relevant physical evidence from grounded reasoning through executable action generation.

arXivで原文を読む →arXiv概要ページ