Make reconstruction an active task
The next observation is part of the reconstruction process.
Unmanned Systems · 2026 / PERCEPTION & CONTROL
AIR-Embodied.
Look closer. Move around. Reach what is hidden.
Multimodal agents that choose viewpoints and interact with objects to build more complete 3D reconstructions.
- 3D Gaussian Splatting
- Multimodal agents
- Active perception
01 / OVERVIEW
The idea.
AIR-Embodied: Active Interactive Reconstruction for 3D Gaussian Splatting with Embodied Multimodal Agents
Unmanned Systems · 2026
Active reconstruction requires deciding what to observe next. AIR-Embodied brings multimodal reasoning, viewpoint planning, and physical interaction into a 3D Gaussian Splatting pipeline, allowing the embodied system to address parts of a scene that passive observation leaves occluded.
- Viewpoint selection guided by scene understanding
- Physical interaction to reveal occluded regions
- Reconstruction updated with new observations
RESEARCH VISUALS
Watch the experiment.
Method preview from the coauthor’s project materials. The original video shows the full demonstration.
02 / IMPLEMENTATION
How it works.
- Observe
Capture the current view of the scene.
- Reason
Identify incomplete or occluded regions.
- Act
Choose a new viewpoint or interact with an object.
- Reconstruct
Integrate the additional observations.
Connect reasoning and action
Multimodal agents connect scene interpretation with embodied actions.
Recover hidden structure
Interaction creates opportunities to observe surfaces that were previously hidden.
03 / RESOURCES
Read further.
The paper, original demonstrations, and available project materials.