GraF-VTA: グラフ融合による視覚と時相論理からのニューロシンボリック軌道生成
GraF-VTA: Neuro-Symbolic Trajectory Generation from Vision and Temporal Logic via Graph-Based Fusion
視覚観測と信号時相論理(STL)から、グラフベースの双方向クロスアテンションと時間ピラミッド記憶、フローマッチング生成、再計画を組み合わせてSTLを満たす軌道を生成するフレームワーク。
詳しい要約
1. どんなもの?
2. 先行研究と比べてどこがすごい?
3. 技術・手法の肝は?
4. どうやって有効だと検証した?
5. 議論はある?
6. 次に読むべき論文は?
※ 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: Zezhi Liu, Zhiwei Zheng, Hanqian Luo, Deyun Qin, Shizhen Wu, Yongchun Fang
分類: cs.RO
原文アブストラクト
Robotic planning from visual observations and signal temporal logic (STL) requires grounding task requirements in scene evidence while adapting plans to execution feedback. We present GraF-VTA, a graph-based framework that extends Vision-TL-Action with structured task representations, causal observation history, and closed-loop trajectory generation. Bidirectional cross-attention fuses STL syntax tokens with visual features, while an optional frozen automaton embedding supplies complementary untimed task structure. A temporal pyramid preserves detailed current observations and compresses older frames to provide recent execution context. Conditioned on these features and robot state, a flow-matching model generates plans whose action prefixes are executed before replanning from refreshed observations. Across five simulation environments, the single-graph configuration achieves higher open-loop Success@$K$ than TeLoGraF and Vision-TL-Action on four environments at the reported budgets $K\geq16$. On Panda, matched-update ablations increase Success@256 from 53.91\% without observation history to 62.50\% with pyramid memory. In closed-loop Dubins evaluation, it reaches 64.06\% Success@1024, compared with 12.50\% for adapted S-MSP and 27.34\% for adapted Logic-VLA. These results support combining structured visual conditioning with observation history and replanning to improve coverage of STL-satisfying trajectories.
