What is TartanDrive?
Dataset Overview
TartanDrive is a large-scale dataset for learning off-road dynamics models. The dataset consists of 630 trajectories in total. These trajectories equate to approximately 200,000 off-road driving interactions. In total, the collected data represents five hours of driving time. The testing site for the data collection is located in Pittsburgh, PA. Research using the TartanDrive dataset was supported by ARL awards #W911NF1820218 and #W911NF20S0005.
Data Collection and Sensor Configuration
The dataset was collected using a modified Yamaha Viking ATV. Human tele-operated controls were used for data collection instead of random controls. Trajectories were ended whenever the driver intervened by applying the brakes. Data collection involved seven unique sensing modalities. The dataset supports the use of AckermannDriveStamped, Float64, Image, IMU, Odometry, PoseStamped, and TwistStamped rostopics. The sensor data was collected at a frequency of 10Hz. The dataset includes an IMU sensor with a data frequency of 200Hz. RGB image resolution in the dataset is 1024 x 512 pixels. State data is recorded at 50Hz, while action data is recorded at 100Hz. Shock position and wheel RPM are recorded at 50Hz. Pedal controls are recorded in the dataset at 50Hz. The authors used ResNet features on RGB images to cluster driving scenarios.
Technical Implementation and Access
The TartanDrive dataset is available on GitHub and the repository is licensed under the MIT license. The dataset file format uses compressed folders of several rosbags. Each dataset file is approximately 100GB in size. The dataset requires custom message types for shock travel, wheel RPM, and pedal position data. The dataset uses torch format for training trajectories. The multi-convert-bag.py script can be used to generate torch training trajectories from rosbags. The default configuration file for the dataset is provided at specs/2021-atv-3.yaml. Users can utilize the visualize-traj.py script to generate videos of the trajectory. The partition-dataset.py script allows creating a train/test split for the dataset. The script eval-prob-world-model is used to calculate the RMSE of a model over a directory of trajectories. TartanDrive provides a benchmark of recent neural network architectures for dynamics prediction from high-dimensional inputs.
Sources
- https://arxiv.org/html/2205.01791
- https://github.com/castacks/tartan_drive