LAVN Dataset

Accepted to HRI2025 Short Contributions

Preprint: arxiv.org/pdf/2308.16682

Dataset Organization

After downloading and unzipping the zip files, please reorganize the files in the following tructure:

LAVN
   |--src
      |--makeData_virtual.py
      |--makeData_real.py
      ...
   |--Virtual
      |--Gibson
         |--traj_<SCENE_ID>
            |--worker_graph.json
            |--rgb_<FRAME_ID>.jpg
            |--depth_<FRAME_ID>.jpg
         |--traj_Ackermanville
            |--worker_graph.json
            |--rgb_00001.jpg
            |--rgb_00002.jpg
            ...
            |--depth_00001.jpg
            |--depth_00002.jpg
            ...
         ...
      |--Matterport
         |--traj_<SCENE_ID>
            |--worker_graph.json
            |--rgb_<FRAME_ID>.jpg
            |--depth_<FRAME_ID>.jpg
         |--traj_00000-kfPV7w3FaU5
            |--worker_graph.json
            |--rgb_00001.jpg
            |--rgb_00002.jpg
            ...
            |--depth_00001.jpg
            |--depth_00002.jpg
            ...
         ...
   |--Real
      |--Campus
         |--worker_graph.json
         |--traj_480p_<SCENE_ID>
            |--rgb_<FRAME_ID>.jpg
         |--traj_480p_scene00
            |--rgb_00001.jpg

where the main landmark annotation scripts makeData_virtual.py and makeData_real.py are in folder (1) src. (2) Virtual and (3) Real store trajectories collected in the simulation and real world, respectively. Each trajectory's data is collected in the following format:

         |--traj_<SCENE_ID>
            |--worker_graph.json
            |--rgb_<FRAME_ID>.jpg
            |--depth_<FRAME_ID>.jpg

where <SCENE_ID> matches exactly the original one in Gibson and Matterport run by the photo-realistic simulator Habitat. Images are saved in either .jpg or .png format. Note that rgb images are the main visual representation while depth is the auxiliary visual information captured only in the virtual environment. Real-world RGB images are downsampled to a 640 × 480 resolution noted by 480p in a trajectory folder name.

worker_graph.json stores the meta data in dictionary in Python saved in json file with the following format:

{"node<NODE_ID>":
  {"img_path": "./human_click_dataset/traj_<SCENE_ID>/rgb_<FRAME_ID>.jpg",
   "depth_path": "./human_click_dataset/traj_<SCENE_ID>/depth_<FRAME_ID>.png",
   "location": [<LOC_X>, <LOC_Y>, <LOC_Z>],
   "orientation": <ORIENT>,
   "click_point": [<COOR_X>, <COOR_Y>],
   "reason": ""},
  ...
 "node0":
  {"img_path": "./human_click_dataset/traj_00101-n8AnEznQQpv/rgb_00002.jpg",
   "depth_path": "./human_click_dataset/traj_00101-n8AnEznQQpv/depth_00002.jpg",
   "location": [0.7419548034667969, -2.079209327697754, -0.5635206699371338],
   "orientation": 0.2617993967423121,
   "click_point": [270, 214],
   "reason": ""}
 ...
 "edges":...
 "goal_location": null,
 "start_location": [<LOC_X>, <LOC_Y>, <LOC_Z>],
 "landmarks": [[[<COOR_X>, <COOR_Y>], <FRAME_ID>], ...],
 "actions": ["ACTION_NAME", "turn_right", "move_forward", "turn_right", ...]
 "env_name": <SCENE_ID>
}

where [<LOC_X>, <LOC_Y>, <LOC_Z>] is the 3-axis location vector, <ORIENT> is the orientation only in simulation. [<COOR_X>, <COOR_Y>] are the image coordinates of landmarks. ACTION_NAME stores the action of the robot take from the current frame to the next frame.

Dataset Usage

The visual navigation task can be formulated as various types of problems, including but not limited to:

1. Supervised Learning by mapping visual observations (RGBD) to waypoints (image coordinates). A developer can design a vision network whose input (X) is RGBD and output (Y) is image coordinate, specified by img_path, depth_path and click point [<COOR_X>, <COOR_Y>] in the worker graph.json file in the dataset. The loss function can be designed to minimize the discrepancy between the predicted image coordinate (Y_pred) and the ground truth (Y), e.g. loss = ||Y_pred − Y||. Then Y_pred can be simply translated to a robot’s moving action, such as Y_pred in the center or top region of an image means moving forward while left/right regions represent turning left or right.

2. Map Representation Learning in the latent space of a neural network. One can train this latent space to represent two observations’ proximity by contrastive learning. The objective is to learn a function h() that predicts the distance given two observations (X1) and (X2): dist = h(X1, X2). Note that dist() can be a cosine or distance-based function, depending on the design choice. The positive samples can be nodes (a node includes information at a timestep such as RGBD data and image coordinates) nearby while further nodes can be treated as negative samples. A landmark is a sparse and distinct object or scene in the dataset that facilitates a more structured and global connection between nodes, which further assists in navigation in more complex or longer trajectories.

Long-Term Maintenance Plan

We will conduct a long-term maintenance plan to ensure the accessability and quality for future research:

Data Standards: Data formats will be checked regularly with scripts to validate data consistency.

Data Cleaning: Data in incorrect formats, missing data or contains invalid values will be removed.

Scheduled Updates: We set up montly schedule for data updates.

Storage Solutions: HuggingFace, with DOI (doi:10.57967/hf/2386), is provided as a public repository for online storage. A second copy will be stored in a private cloud server while a third copy will be stored in a local drive.

Data Backup: Once one of the copies in the aforementioned storage approach is detected inaccessible, it will be restored by one of the other two copies immediately.

Documentation: Our documentation will be updated regularly reflecting feedback from users.

Citation

@article{johnson2024landmark,
  title={A Landmark-Aware Visual Navigation Dataset},
  author={Johnson, Faith and Cao, Bryan Bo and Dana, Kristin and Jain, Shubham and Ashok, Ashwin},
  journal={arXiv preprint arXiv:2402.14281},
  year={2024}
}

@misc{visnavdataset_lavn,
  author       = {visnavdataset},
  title        = {LAVN Dataset},
  year         = 2025,
  doi          = {10.57967/hf/2386},
  url          = {https://huggingface.co/datasets/visnavdataset/lavn},
  note         = {Accessed: 2025-02-07}
}

Note: change the accessed date.