![Illustration of a multi-story indoor house layout used in the CapNav benchmark. The diagram asks: “Can [AGENT] go from the basement foyer to the top floor rest area?” The basement foyer (START) is at the bottom left, and the top floor rest area (END) is at the top right. Colored paths show different possible routes through labeled areas including Garage, First Floor Hallway, and Elevator. Icons on the left represent four agent types: a wheelchair user (allowed, green check), a quadruped robot (allowed, green check), a human (allowed, green check), and a sweeping robot (not allowed, red X). Red X markers and warning labels indicate constraints such as “Cannot go up stairs” and “Cannot operate elevator,” showing how different mobility capabilities affect route feasibility.](/media/projects/images/Screenshot_2026-02-23_at_11.51.22AM.png.500x300_q85_box-277%2C0%2C1094%2C491_crop_upscale.png)
CapNav
2025 – Present
Jingwei Ma
PhD Student, Allen School of Computer Science and Engineering
University of Washington
Jingwei Ma is a third year PhD student in UW Graphics and Imaging Laboratory (GRAIL) working with Prof. Ira Kemelmacher-Shlizerman, Prof. Steve Seitz and Prof. Brian Curless.
Projects
Papers
CapNav: Benchmarking Vision Language Models on Capability-conditioned Indoor Navigation
Proceedings of CVPR 2026 To Appear | Acceptance Rate: 25.4% (4090 / 16092)
FlyMeThrough: Human-AI Collaborative 3D Indoor Mapping with Commodity Drones
Proceedings of UIST 2025 | Acceptance Rate: 22.2% (210 / 947)
A Demo of DIAM: Drone-based Indoor Accessibility Mapping
Extended Abstract Proceedings of UIST 2024
PDF doi Cite Poster DIAM: Drone-based Indoor Accessibility Mapping