EAGLE-360 Qwen3-VL GRPO

Project page: EAGLE-360

Paper: arXiv:2607.02479

This repository contains the merged Hugging Face checkpoint used by EAGLE-360 for 360-degree panoramic object localization. Given a panoramic image and a target-object query, the model predicts the object's azimuth and elevation in degrees through multi-turn global-to-local reasoning.

The checkpoint is based on Qwen3-VL and fine-tuned with panoramic multi-turn reasoning and GRPO. It is intended to be used with the EAGLE-360 evaluation code and the panoramic Rolling RoPE patches provided in the project repository.

Data

Public test split: Sansjudge/eagle360_test

Usage

Please install the dependencies and apply the vLLM/Transformers panoramic patches from the EAGLE-360 project repository before running evaluation.

python eval.py \
  --model ./checkpoints/hf_merged \
  --test_file ./eagle360_test/test.json \
  --pano_dir ./eagle360_test/images \
  --n_samples 50

Citation

@misc{xu2026eagle360embodiedactiveglobaltolocal,
      title={EAGLE-360: Embodied Active Global-to-Local Exploration in 360$^\circ$},
      author={Jingtao Xu and Zizhuo Lin and Jianwen Sun and Yi Yang and Yawei Luo},
      year={2026},
      eprint={2607.02479},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.02479},
}