ml-resources/Indoor_UAV

Mirror of valyentine/Indoor_UAV from ModelScope. All credit to the original authors. Not an official release from them. License unchanged (apache-2.0).

IndoorUAV Dataset

IndoorUAV: Benchmarking Vision-Language UAV Navigation in Continuous Indoor Environments. This project contains the dataset of our AAAI 2026 paper. The usage of the dataset can be found in our github repo.

[Paper] [Github]

VLN File Hierarchy

scene_group_id/       # e.g. replica, gibson_1, hm3d_15.
└──scene_id/          # e.g. Adrian, 226REUyJh2K.
    └──traj_id/       # Trajectory collected from the scene; the traj_-* is the inverse edtion of the corresponding traj_*.
        ├── screenshots/              # The 720P, 10FPS camera view in the trajectory.
        └── saved_transformations/    # The raw posture matrix obtained from habitat-sim. 
        ├── posture.txt               # The formatted posture information, equivalent to saved_transformations.
        ├── real_action.json          # The actual movement type and displacement magnitude per frame.
        ├── annotation.json           # Frame numbers annotated with standard action labels.
        ├── key_frames.json           # Key frame numbers and their corresponding descriptive captions.
        ├── instruction.json          # Simplified edition of the natural language instruction.
        └── instruction_pro.json      # Detailed edition of the natural language instruction.

The posture.txt file contains the agent's pose for each frame, stored as 4-dimensional coordinates corresponding to x,y,z (in meters), yaw (in degrees, range 0-360°). For a concrete example of how to load and use this data, please refer to the load_position function in utils.py and test_sim.py.

VLA File Hierarchy

scene_group_id/       
└──scene_id/          
    └──traj_id/       
        └── vla_ins_*.json           # VLA instruction and the start & end frame in the corresponding VLN trajctory.

Other Files

scene_datasets.zip: scenes data in gibson, mp3d, hm3d and replica for habitat-sim.

without_screenshots.zip: IndoorUAV-VLN dataset without images to save the space (use the posture.txt to get images instead~).

train.csv, test_seen.csv, test_unseen.csv: contain the trajectories for the training set, seen test set and unseen test set, respectively.

• [Feb 2026] depth_images: depth information of the VLN trajctories collected from the habitat-sim, similiar file hierarchy to VLN part.