CVPR 2026: DAAAM - Describe Anything, Anywhere, at Any Moment

[arXiv] [Project Page]

Real-time foundation-model-first robot mapping: SAM segmentation + BotSort tracking + VLM grounding feed into Hydra to build 3D Dynamic Scene Graphs on the fly.

Key contributions:

• Novel optimization-based frontend for semantic descriptions from localized captioning models
• Hierarchical 4D scene graph construction with real-time performance
• State-of-the-art results on NaVQA and SG3D benchmarks

The ROS 2 interface lives in DAAAM-ROS.

Installation | Running | Evaluation | Codebase | DAAAM-ROS

Paper

If you use this code in your work, please cite the following paper:

Nicolas Gorlo, Lukas Schmid, and Luca Carlone, "Describe Anything, Anywhere, at Any Moment". arXiv preprint arXiv:2512.00565, 2025.

@inproceedings{Gorlo26cvpr-DAAAM,
      title={Describe anything anywhere at any moment},
      author={Gorlo, Nicolas and Schmid, Lukas and Carlone, Luca},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
      pages={35002--35013},
      year={2026}
}

This work was supported by the ARL DCIST program and the ONR RAPID program.