BAT-Nav: 長期セマンティックナビゲーションのための予算認識型調停と終了
BAT-Nav: Budget-Aware Arbitration and Termination for Long-Horizon Semantic Navigation
限られた行動予算で複数の目標を探索する長期ナビゲーションにおいて、低収益な目標が予算を独占する問題を解決するため、凍結されたVLMベースのナビゲーションの上で動作する訓練不要のオンライン目標調停機構を提案した。
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: Xi Lin, Kangyi Wu, Jiayi Li, Jiaqiao Tang, Qingrong He, Lin Zhao
分類: cs.RO
原文アブストラクト
Long-horizon semantic navigation asks a robot to localize multiple open-vocabulary targets under a finite action budget. This setting exposes an execution failure that is largely hidden in single-goal ObjectNav: a low-yield or occluded target can monopolize the action budget of a reactive navigator and leave later goals unattempted. We present BAT-Nav, a training-free online goal arbitrator above a frozen VLM-guided navigation backbone. Rather than modifying the low-level policy, BAT-Nav monitors execution telemetry and updates the mission goal queue through four interventions: PERSIST, SWITCH, ABORT, and COMMIT. The controller separates allocation from verification: progress stagnation and budget retention drive ABORT/SWITCH, whereas temporal variance filtering governs COMMIT. In the deployable observable-proxy setting, BAT-Nav improves the CR-WSF frontier over fixed-patience, single-signal, dynamic-cap, cap-plus-verification, revisit-cap, and frontier-utility controllers. On HM3D, BAT-Nav-Observable reaches 0.345 CR / 0.682 WSF, improving CR by 0.063 and reducing WSF by 0.053 over DynamicCapOnly. On MP3D, it reaches 0.263 CR / 0.748 WSF, improving CR by 0.061 and reducing WSF by 0.047 over DynamicCapOnly. Behavior diagnostics show that BAT-Nav attempts more goals per episode and reduces unresolved-goal monopoly rate from 0.39 to 0.18 on HM3D and from 0.44 to 0.27 on MP3D. Oracle telemetry is reported as an upper-bound ceiling rather than a deployable setting.
