EAGLE-360 is a framework for active visual search in 360° panoramic environments. It uses global priors to build an initial holistic view and narrows the search space through iterative reasoning, with an adapted RoPE Rolling position encoding to handle panoramic topology.

EAGLE-360 achieves nearly 8× higher accuracy than baseline models on 360° visual search, addressing polar distortion and limited field-of-view challenges. It significantly enhances exploration efficiency and error recovery for embodied AI in 3D environments.

The EAGLE-360 dataset (14,000 4K panoramas + 70,000 VQA dialogues) is open-sourced, filling a critical data gap. Future work can explore multimodal fusion, real-time interaction, and cross-domain generalization for embodied AI applications.

EAGLE-360：基於全局先驗的360度全景主動探索與視覺搜尋框架

針對多模態大語言模型在360度全景環境中主動視覺搜尋時面臨的極區畸變建模困難及局部視野局限問題，本文提出EAGLE-360框架。該方法利用全局先驗建立初始整體視角，通過迭代推理逐步縮小搜尋空間，克服傳統方法對碎片化局部視角的依賴。技術上，透過適配RoPE Rolling座標移位位置編碼機制，無縫建模全景圖的連續圓柱拓撲結構。同時，構建包含1.4萬張4K全景圖及7萬輪高品質VQA對話的EAGLE-360數據集，並結合監督微調與組相對策略最佳化訓練策略。實驗表明，該方法在360度視覺搜尋任務上達到最新水準，準確率較基線模型提升近8倍，顯著增強了探索效率與錯誤恢復能力，為具身智能在全景環境中的空間推理提供了新範式。

Sources

arXiv