Hugging Face's logo

Datasets:
maum-ai
/
CostNav-Teleop-Dataset

Tasks:
Robotics
Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
1K - 10K
ArXiv:
Tags:
navigation
teleoperation
ros2
mcap
delivery-robot
imitation-learning
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
2
CostNav-Teleop-Dataset / README.md
PurpleSand's picture
PurpleSand
Improve dataset card with full documentation, usage examples, and data processing pipeline
de5dd4d
preview code
|
raw
history blame contribute delete
10 kB
metadata 
license: apache-2.0
task_categories:
  - robotics
language:
  - en
tags:
  - navigation
  - teleoperation
  - ros2
  - mcap
  - delivery-robot
  - imitation-learning
  - physical-ai
pretty_name: >-
  CostNav Teleop Dataset: Human Teleoperation Data for Navigation Cost
  Benchmarking
size_categories:
  - 1K<n<10K
arxiv: '2511.20216'

CostNav Teleop Dataset

arXiv Blog Post GitHub Project Page Documentation

Dataset Summary

The CostNav Teleop Dataset is a large-scale collection of human teleoperation recordings for robot navigation in an urban sidewalk simulation environment. It was collected as part of the CostNav benchmark, which evaluates navigation systems using real-world economic cost and revenue metrics rather than purely technical metrics.

The dataset contains 2,203 teleoperation episodes totaling 50.2 hours of driving data, collected by 4 human operators using joystick control of a Segway E1 delivery robot in NVIDIA Isaac Sim. Each episode is stored as a ROS 2 bag file in MCAP format with synchronized multi-modal sensor data.

CostNav Simulation Environment

Navigation Comparison

CostNav Demo

Physics Simulation

Physics Simulation - Popcorn Dynamics

CostNav's high-fidelity physics simulation enables the modeling of real-world economic scenarios, including critical failures like food spoilage and robot rollovers.

Economic Model

CostNav Economic Model

Key Statistics

Statistic Value
Total episodes 2,203
Total duration 50.2 hours
Total ROS messages ~71.4M
Collectors 4 (A, B, C, D)
Episodes per collector A: 376, B: 744, C: 73, D: 446
Episode outcomes Success (O): 1,639 / Partial (△): 59 / Fail (X): 505
Food camera available Yes: 1,136 / No: 1,067
Storage format ROS 2 bag (MCAP)
Simulation platform NVIDIA Isaac Sim
Robot Segway E1 delivery robot
ROS distribution Jazzy

Dataset Structure 

CostNav-Teleop-Dataset/
├── collector_A/                          # 376 episodes
│   ├── recording_YYYYMMDD_HHMMSS/
│   │   ├── metadata.yaml                 # ROS bag metadata (topics, QoS, message counts)
│   │   └── recording_*_0.mcap            # MCAP bag file with all sensor data
│   └── ...
├── collector_B/                          # 744 episodes
├── collector_C/                          # 73 episodes
├── collector_D/                          # 446 episodes
└── costnav data collection_v3 recording - rosbags_time_summary.csv

Summary CSV Columns

Column Description
recording_dir Path to recording directory
metadata_path Path to metadata YAML
mcap_files MCAP bag filename
start_time_utc / end_time_utc Episode timestamps (UTC)
start_time_kst / end_time_kst Episode timestamps (KST)
duration_seconds Episode duration in seconds
duration_hms Episode duration (H:MM:SS)
message_count Total ROS messages in episode
episode_success Outcome: O (success), △ (partial), X (fail)
food_cam Whether food camera data is included (O/X)

ROS 2 Topics

Each MCAP recording contains the following topics:

Topic Message Type Description
/front_stereo_camera/left/image_raw/compressed sensor_msgs/CompressedImage Front stereo camera (left, compressed)
/front_stereo_camera/left/camera_info sensor_msgs/CameraInfo Camera intrinsics
/front_3d_lidar/lidar_points sensor_msgs/PointCloud2 3D LiDAR point cloud
/scan sensor_msgs/LaserScan 2D laser scan
/chassis/odom nav_msgs/Odometry Wheel odometry
/chassis/imu sensor_msgs/Imu Chassis IMU
/front_stereo_imu/imu sensor_msgs/Imu Stereo camera IMU
/cmd_vel geometry_msgs/Twist Teleop velocity commands
/joy sensor_msgs/Joy Raw joystick input
/is_model std_msgs/Bool Model control flag
/tf / /tf_static tf2_msgs/TFMessage Transform tree
/map nav_msgs/OccupancyGrid Occupancy grid map
/goal_pose geometry_msgs/PoseStamped Goal position
/initialpose geometry_msgs/PoseWithCovarianceStamped Initial pose estimate
/start_marker / /goal_marker / /robot_marker visualization_msgs/Marker RViz markers
/clock rosgraph_msgs/Clock Simulation clock
/diagnostics diagnostic_msgs/DiagnosticArray System diagnostics

Usage

Loading with MCAP 

from mcap.reader import make_reader

with open("collector_A/recording_20260202_052239/recording_20260202_052239_0.mcap", "rb") as f:
    reader = make_reader(f)
    for schema, channel, message in reader.iter_messages():
        print(f"{channel.topic}: {schema.name} @ {message.log_time}")

Loading with ROS 2 

# Play back a recording
ros2 bag play collector_A/recording_20260202_052239/

# Inspect bag info
ros2 bag info collector_A/recording_20260202_052239/

Data Processing: MCAP to Training Format

The CostNav repository includes a full data processing pipeline to convert raw MCAP bags into training-ready formats for imitation learning baselines (ViNT, NavDP, etc.).

Step 1: Convert MCAP bags to MediaRef format (extracts images to H.264 video for efficient lazy-loading):

python costnav_isaacsim/il_training/data_processing/converters/rosbag_to_mediaref.py \
    --input data/sample_rosbags/recording_20260109_061808 \
    --output data/processed/recording_20260109_061808

Step 2: Batch convert multiple bags in parallel using Ray:

python costnav_isaacsim/il_training/data_processing/converters/ray_batch_convert.py \
    --input-dir data/sample_rosbags/ \
    --output-dir data/mediaref/ \
    --num-workers 4

Step 3: Convert MediaRef to ViNT training format (synchronized images + odometry at configurable Hz):

python costnav_isaacsim/il_training/data_processing/process_data/process_mediaref_bags.py \
    --input-dir data/mediaref/ \
    --output-dir data/vint_format/ \
    --config costnav_isaacsim/il_training/data_processing/configs/vint_processing_config.yaml

Output structure (ViNT format):

data/vint_format/
└── trajectory_name_0/
    ├── 0.jpg, 1.jpg, ...    # Extracted images (160x120, 4:3)
    └── traj_data.pkl         # Position and yaw arrays

Alternative: Convert to NavDP format (with depth estimation via DepthAnythingV2):

python costnav_isaacsim/il_training/data_processing/process_data/process_mediaref_to_navdp.py \
    --input-dir data/mediaref/ \
    --output-dir data/navdp_format/ \
    --config costnav_isaacsim/il_training/data_processing/configs/navdp_processing_config.yaml

See the full Data Processing Documentation and Teleoperation Guide for details on configuration, the recording pipeline, and robot control setup.

Related Resources

Citation

If you use this dataset in your research, please cite the CostNav paper:

@misc{seong2026costnavnavigationbenchmarkrealworld,
      title={CostNav: A Navigation Benchmark for Real-World Economic-Cost Evaluation of Physical AI Agents},
      author={Haebin Seong and Sungmin Kim and Yongjun Cho and Myunchul Joe and Geunwoo Kim and Yubeen Park and Sunhoo Kim and Yoonshik Kim and Suhwan Choi and Jaeyoon Jung and Jiyong Youn and Jinmyung Kwak and Sunghee Ahn and Jaemin Lee and Younggil Do and Seungyeop Yi and Woojin Cheong and Minhyeok Oh and Minchan Kim and Seongjae Kang and Samwoo Seong and Youngjae Yu and Yunsung Lee},
      year={2026},
      eprint={2511.20216},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2511.20216},
}

License

This dataset is released under the Apache 2.0 License.