COMMAND Dataset 😃

🤗 Dataset Availability

The evaluation data from this dataset is now available on Hugging Face:

• Repository: maum-ai/COMMAND
• Content: 60 evaluation scenarios (240 files, ~38GB)
• File Types: ROS bagfiles (.bag), metadata (.json), instructions (.npy), visualizations (.png)
• Purpose: Standardized evaluation set for benchmarking navigation models
• Upload Date: September 25, 2025
• Status: ✅ Successfully uploaded and verified

Dataset Structure 🗂️

The COMMAND_dataset is organized as follows:

COMMAND_dataset/
├─ README.md                      # Documentation for the dataset
├─ map_asset/                     # Directory containing map assets 🗺️
│  ├─ map_list.yml                # Information of all maps
│  └─ maps/                       # Folder containing individual maps
│     └─ <map_name>/             # Folder for a specific map
│        ├─ .png          # Image representation of the map
│        └─ .yml          # Map info (for ROS map server)
├─ bagfiles/                      # Directory containing ROS bagfiles 🎥
│  └─ .bag                  # ROS bagfile with unique identifier
├─ scenarios/                     # Directory containing scenario information 📜
│  ├─.json                 # Scenario schema
│  └─.npy                  # Sketch instructions
└─ eval/                          # Directory containing evaluation scenarios 🧪
   ├─.bag                   # ROS bagfile for evaluation
   ├─.json                  # Evaluation scenario schema
   ├─.npy                   # Evaluation sketch instructions
   └─.png                   # Evaluation scenario visualization

How to extract the dataset:

# To recombine the parts, use:
cat COMMAND_part_* > COMMAND.tar.zst
# Then decompress the archive:
tar --use-compress-program=zstd -xf COMMAND.tar.zst

---

Dataset Overview 📈

Statistics

Environment	Office (sim_office)	Street (Road) (sim_street_road)	Street (Sidewalk) (sim_street_sidewalk)	Orchard (sim_orchard)
Count	2,253	400	402	262
Avg. Time	31s	57s	103s	172s
Avg. Distance	32.8m	80.4m	150.0m	191.6m
Avg. FD (P / M)	1.05 / 1.77	0.97 / 2.02	3.03 / 3.50	1.91 / 3.76
% of Misleading	31%	51%	50%	39%

Training Data: 3,317 bagfiles, includes 48 hours of driving data over 219 kilometers. Evaluation Data: 60 bagfiles for standardized model evaluation.

Total: 3,377 bagfiles in the complete dataset.

Note: Training data values differ slightly from the paper because the data version history is missing.

Directory Details 🏗️

map_asset/

• map_list.yml: Contains metadata about all maps in the dataset.
• maps/: Holds individual map folders. Each folder typically includes:
  • .png: An image representation of the map.
  • .yml: Detailed map information (compatible with the ROS map server).

bagfiles/

• Contains ROS bagfiles, each identified by a unique UUID (e.g., <uuid>.bag).

scenarios/

• <uuid>.json: Scenario schema file describing the scenario.
• <uuid>.npy: Scenario annotations (sketch instructions). It contains a list of (y, x) pixel coordinates on the corresponding sketch_map_name map image.

eval/

• <uuid>.bag: ROS bagfile containing expert demonstration data for the evaluation scenario with robot sensor readings and control commands.
• <uuid>.json: Evaluation scenario schema file with the same structure as training scenarios but specifically designed for model evaluation.
• <uuid>.npy: Evaluation sketch instructions containing (y, x) pixel coordinates on the corresponding map image.
• <uuid>.png: Visualization image of the evaluation scenario showing the trajectory and environment.

Evaluation Statistics:

• Total scenarios: 60 evaluation scenarios (10 scenarios per each of the five environments + unseen environment Art museum (sim_art_museum))
• File types: 4 files per scenario (.bag, .json, .npy, .png)
• Purpose: Standardized evaluation set for benchmarking navigation models trained on the COMMAND dataset

---

Topics in ROS Bagfiles 🎥

1. /clock

   • Message Type: rosgraph_msgs/msg/Clock
   • Description: Publishes the current simulated time in a ROS system. Useful for synchronizing sensor data and events in simulation.

2. /cmd_vel

   • Message Type: geometry_msgs/msg/Twist
   • Description: Represents commanded linear and angular velocities for a robot, typically used to drive mobile robots (e.g., differential drive, holonomic platforms).

3. /imu

   • Message Type: sensor_msgs/msg/Imu
   • Description: Provides data from an Inertial Measurement Unit, including orientation (as a quaternion), angular velocity, and linear acceleration.

4. /joy

   • Message Type: sensor_msgs/msg/Joy
   • Description: Reports the state of joystick (gamepad) axes and buttons, enabling teleoperation or manual input to the robot.

5. /odom

   • Message Type: nav_msgs/msg/Odometry
   • Description: Offers an estimate of the robot's position and velocity in free space. This serves as ground-truth data in simulation.

6. /rgb_front/camera_info, /rgb_left/camera_info, /rgb_right/camera_info

   • Message Type: sensor_msgs/msg/CameraInfo
   • Description: Contains intrinsic and extrinsic parameters of a camera.

7. /rgb_front/compressed, /rgb_left/compressed, /rgb_right/compressed

   • Message Type: sensor_msgs/msg/CompressedImage
   • Description: Encoded image data (JPEG/PNG) from a camera feed.

8. /tf

   • Message Type: tf2_msgs/msg/TFMessage
   • Description: Provides an odom-to-base_link transform.

Schema of Scenario JSON ��

Both training scenarios (in scenarios/) and evaluation scenarios (in eval/) follow the same JSON schema:

• semantic_uuid: A unique identifier for the scenario. Combines map name, scenario name, and either "precise" or "misleading," followed by the scenario number.
• map_name: The name of the map where the scenario is located.
• sketch_map_name: The map where the sketch instructions are provided.
• start: The starting point of the robot (can be a specific point or "random").
• end: The ending point of the robot, predefined in annotated maps.
• drive_map_name: The map that the robot uses for navigation, usually an occupancy map.
• model_guideline: Instructions for the robot to follow during the scenario, serving as input for model training.

Example:

{
    "semantic_uuid": "sim_office/coffee_delivery/precise/3/42",
    "map_name": "sim_office",
    "sketch_map_name": "annotated_occupancy_map",
    "start": "(random)",
    "end": "4",
    "drive_map_name": "occupancy_map",
    "model_guideline": "You are an indoor food-serving robot.\n\nYou must follow these driving instructions:\n1. You must avoid collisions.\n2. You should prioritize reaching the final destination.\n3. You should follow the Trajectory Instruction.\n    a. If the Trajectory Instruction cannot be followed due to any obstacles, you should deviate to bypass the obstacle.\n    b. You should try to evade any identifiable obstacles.\n4. You should maintain a constant driving speed."
}

Citation

If you find this work useful, please cite our paper:

@article{choi2024canvas,
  title={CANVAS: Commonsense-Aware Navigation System for Intuitive Human-Robot Interaction},
  author={Choi, Suhwan and Cho, Yongjun and Kim, Minchan and Jung, Jaeyoon and Joe, Myunchul and Park, Yubeen and Kim, Minseo and Kim, Sungwoong and Lee, Sungjae and Park, Hwiseong and others},
  journal={arXiv preprint arXiv:2410.01273},
  year={2024}
}