TartanDrive 2.0
CMU AirLab · 2024 · datasets.bot · datasets.bot page
One-liner. Seven-hour off-road driving dataset collected on a Yamaha Viking ATV at speeds up to 15 m/s, adding three LiDAR sensors to the original camera, IMU, GPS and proprioceptive data, with tooling to reconfigure the data for self-supervised off-road learning.
Setup
- Datasets / benchmarks: TartanDrive 2.0 is an expanded large-scale off-road driving dataset from CMU's AirLab, collected on a Yamaha Viking All-Terrain Vehicle (ATV named 'Vicky') over 255+ acres of diverse terrain in Western Pennsylvania (narrow paths, dense foliage, rocky areas, dirt paths, steep hills). It comprises roughly 7 hours of data at speeds up to 15 m/s. Building on TartanDrive 1.0, version 2.0 adds three LiDAR sensors: two Velodyne VLP-32 units (front-roof mounted, one tilted downward) and one Livox Mid-70 (under the camera), all at 10 Hz. The camera system is a Carnegie Robotics MultiSense S21 providing stereo greyscale images plus a left RGB image at 10 Hz. Inertial/positioning comes from a NovAtel PROPAK-V3-RT2i GNSS (IMU at 100 Hz, GPS-fused pose at 50 Hz) and the MultiSense IMU at 400 Hz. Proprioceptive data (Racepak G2X Pro logger plus joystick/pedals) includes per-wheel RPM, rear suspension shock travel, steering commands and angle, and accelerator/brake pedal positions. Post-processed products include TartanVO and Super Odometry estimates, registered point clouds, predicted point clouds, heightmaps, RGB bird's-eye-view maps, IMU-derived roughness cost, and 200x200m BEV LiDAR feature maps at 0.5m resolution (min/max/mean height, roughness, SVD features, ground height/slope). Rich per-sequence metadata (driver ID, vehicle, passenger count, date/time, weather, lighting, course ID, top/average speed, duration) plus timestamped event annotations and 1-5 driving-quality scores are included. Data is distributed both as original rosbags and in a KITTI-like time-synced sequence format, with scripts to regenerate data at custom sample frequency, map size and resolution. A GUI tool (scripts/tartandrive_gui.py) lets users browse runs and select which modalities to download to save storage; a point-cloud scan of the ATV is downloadable separately. Released alongside the ICRA 2024 paper (arXiv:2402.01913). License: MIT. Download: https://github.com/castacks/tartan_drive_2.0.
- Hardware / simulator: Embodiment: not listed. Environment: outdoor. Realness: physical.
Schema
KITTI-like time-synced sequence format (and original rosbags). Per sample/sequence: FPV left RGB + stereo greyscale images (10 Hz), point clouds from 2x Velodyne VLP-32 + Livox Mid-70 (10 Hz), TartanVO/Super Odometry pose, registered/predicted point clouds, heightmaps, RGB BEV maps, 200x200m BEV LiDAR feature maps @0.5m, IMU (100/400 Hz), GPS-fused pose (50 Hz), per-wheel RPM, suspension shock travel, steering angle/commands, pedal positions, plus per-sequence metadata, event annotations and driving-quality scores.
Links