EAGLE-360: Embodied Active Global-to-Local Exploration in 360$^\circ$

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Expert, quick

Summary

EAGLE-360 is a novel Embodied Active Global-to-Local Exploration framework designed to overcome limitations of Multimodal Large Language Models (MLLMs) in 360-degree panoramic visual search. Standard MLLMs struggle with inherent panoramic properties like severe polar distortion and continuous cylindrical topologies, which significantly reduces target detection accuracy. Current panoramic search methods often rely on fragmented local viewpoints, resulting in myopic, inefficient exploration and poor error recovery. EAGLE-360 addresses this by leveraging global priors to establish an initial holistic perspective, then iteratively narrowing the search space. Its architecture adapts RoPE Rolling, a coordinate-shifting positional encoding, to model continuous panoramic topologies. The framework is supported by the large-scale EAGLE-360 dataset, comprising over 14,000 4K panoramas and 70,000+ VQA dialogues. Training integrates Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO) to develop complex spatial reasoning and tool-calling. Experiments demonstrate EAGLE-360 achieves an 8-fold increase in accuracy over the base model and significantly enhances exploration efficiency, setting a new benchmark.

Key takeaway

For Computer Vision Engineers developing 360-degree visual search systems, consider adopting a global-to-local exploration strategy. Your current MLLM approaches likely struggle with panoramic distortions and continuous topologies, leading to inefficient searches. Implementing techniques like RoPE Rolling and leveraging global priors, as demonstrated by EAGLE-360, can significantly boost your target detection accuracy. This also enhances exploration efficiency. Explore constructing large-scale VQA datasets and integrating SFT with GRPO for robust spatial reasoning in your models.

Key insights

EAGLE-360 improves 360-degree visual search by using global priors and adapting MLLMs for panoramic topologies, achieving an 8-fold accuracy increase.

Principles

• Global priors enhance panoramic search efficiency.
• Continuous topology modeling improves MLLM adaptation.
• Iterative search space narrowing is effective.

Method

EAGLE-360 employs global priors for initial perspective, then iteratively narrows the search space. It adapts RoPE Rolling for continuous panoramic topologies and uses SFT with GRPO for training.

In practice

• Integrate RoPE Rolling for panoramic MLLMs.
• Utilize global priors for initial search.
• Construct large-scale VQA datasets.

Topics

• EAGLE-360
• 360-degree Visual Search
• Multimodal LLMs
• Panoramic Vision
• RoPE Rolling
• Reinforcement Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• MMD-SLAM: Structure-Enhanced Multi-Meta Gaussian Distribution-Guided Visual SLAM — MMD-SLAM enhances 3DGS-based SLAM by integrating structural priors and point-line fusion for superior tracking and mapping.
• LOCUS: Local Visual Cue Search for Enhancing Fine-Grained Perception in Multimodal Large Language Models — LOCUS enhances MLLM fine-grained perception by internalizing local visual cue search via a training-time proxy task.
• MDE-VIO: Enhancing Visual-Inertial Odometry Using Learned Depth Priors — Integrating temporally consistent learned depth priors into VIO backend optimization significantly improves accuracy and robustness on edge devices.
• Global-Local Monte Carlo Tree Search in Vision-Language Models for Text-to-3D Indoor Scene Generation — PRM-guided MCTS with hierarchical scene representation enables LVLMs to generate realistic 3D indoor scenes by correcting errors and balancing search.
• LongSpace: Exploring Long-Horizon Spatial Memory from Perception to Recall in Video — LongSpace enhances MLLMs' long-horizon spatial reasoning by integrating explicit 3D structural memory and question-guided retrieval.
• Learning Geometric Representations from Videos for Spatial Intelligent Multimodal Large Language Models — GeoVR fundamentally restructures MLLM's internal representations for 3D spatial intelligence using 2D video and 3D foundation model distillation, without inference overhead.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.