TartanDrive (Autopilot Training Dataset)

2022: Announcement of a datacet for training cars autonomous off-road driving

In late May 2022, researchers at Carnegie Mellon University took an all-terrain vehicle on a wild ride through tall grass, loose gravel and mud to gather data on how the ATV interacts with complex off-road environments. The resulting dataset, called TartanDrive, includes about 200,000 real interactions, and five hours of data can be useful for learning autonomous off-road driving. Dataset is publicly available.

The researchers believe the data is the largest multimodal dataset of actual SUV driving, both in terms of interactions and sensor types. An all-terrain vehicle with a large number of devices drove at speeds of up to 50 km per hour. The researchers covered a long distance, went up and down the hills and even got stuck in the mud and all that when collecting data, such as video, the speed of each wheel and the magnitude of the suspension shock absorbers, from seven types of sensors.

An open datacet has been created to teach cars autonomous off-road driving

Previous off-road driving work often used annotated maps that gave designations such as dirt, grass, vegetation or water to help the robot understand the area. But such information is infrequent, and even if it is, the data may be useless. For example, a portion of the card marked as dirt may or may not be suitable for movement.

The scientists found that the multimodal sensor data they collected for TartanDrive allowed them to build prediction models superior to those developed from simpler, non-dynamic data. Aggressive driving also pushed the ATV into the performance area, where understanding the dynamics became necessary, the researcher said. The scientist added that during the test, a person drove an ATV, although the Drive-by-Wire system was used to control steering and speed.^[1]

TartanDrive dataset

Notes