Researchers at Carnegie Mellon University in the US are collecting data to train self-driving ATVs.
To handle the vehicle in the terrain to learn, it must also be trained. This succeeds with data sets that represent real situations that the vehicle has to deal with. This should work with the new record of the Carnegie Mellon University go faster. The dataset called Tartan Drive helps train the vehicles.
The team drove an ATV at 30 miles per hour through an off-road environment near Pittsburgh. During the test, they had the vehicle perform various manoeuvres. So does curves, climbs and slopes, and through mud. The data of the vehicle and the sensors were collected.
The demanding terrain with the driving style was a challenge for the technology and provided edge cases that were interesting for the researchers. In this way, Tartan Drive includes 200,000 real interactions for five hours as a training basis. It is said to be more effective than the previous approach using maps to orient the vehicle. A supposedly passable terrain could be muddy and therefore not usable.
The data allowed the team to create predictive models that outperformed non-dynamic models. According to the researchers, software that can understand the dynamics of off-road driving is better able to master the terrain.