Listen to this article
Researchers at Carnegie Mellon University have created a dataset, called the Tartan Drive, that could help train self-driving all-terrain vehicles (ATVs).
The research team at the university drove an ATV aggressively through an off-road environment near Pittsburgh at 30 miles an hour. During the test, the team drove the ATV around turns, up and down hills and through mud, all while gathering data about how the vehicle was performing. The data included video, the speed of each wheel and the amount of suspension shock that traveled from seven types of sensors.
“The dynamics of these systems tend to get more challenging as you add more speed,” Samuel Triest, lead author on the team’s paper and master’s student in robotics, said. “You drive faster, you bounce off more stuff. A lot of the data we were interested in gathering was this more aggressive driving, more challenging slopes and thicker vegetation because that’s where some of the simpler rules start breaking down.”
All the data gathered resulted in the Tartan Drive, which includes around 200,000 real-world interactions and five hours of data that could help train self-driving ATVs handle off-road driving.
Typically, off-road driving is done with an annotated map that provides information about what terrain to expect. Areas are labelled as mud, grass, vegetation or water so that the robot can understand what areas it will be able to navigate. While these labels can be helpful, they don’t provide enough information. For example, a muddy area could be navigable, or it could result in the robot getting stuck.
“Unlike autonomous street driving, off-road driving is more challenging because you have to understand the dynamics of the terrain in order to drive safely and to drive faster,” Wenshan Wang, a project scientist in the Robotics Institute (RI), said.
The data the research team gathered helped them to build prediction models that worked better than models developed with simpler, non-dynamic data. By driving the ATV aggressively during tests, the team put the vehicle into a performance realm where an understanding of dynamics was essential. Robots that can understand dynamics are more likely to be able to reason about the physical world.
The research team that worked on the paper included Sebastian Scherer, an associate research professor in the RI, Aaron Johnson, an assistant professor of mechanical engineering, Wang and Triest.