AI learned to drive a car in rough terrain. In the future, this system can be implemented in unmanned vehicles.
Researchers at McGill University have developed a remote-control cross-country mini-SUV training system. At the same time, she is trained using aerial photography and first-person shots. The hybrid approach takes into account uneven terrain and obstacles using on-board sensors, which makes it possible to generalize the environment around with vegetation, rocks and sand.
In the future, this work can be used to train autonomous cars such as Wayve, Tesla, Mobileye and Comma.ai. Now they rely on camcorders to train their navigational AI.
Researchers used an SUV with an electric motor and a mechanical brake that plugs into an open-source Intel i7 NUC computer. The device is equipped with a LiDAR short-range sensor and a front-facing camera in combination with a microcontroller, which transfers all the information from the sensor to a computer.
Before deploying the device on the track, the team took pictures of the track from a height of 80 m. Then they extracted data from these frames in order to orient and center the route. Pictures were taken with a resolution of 0.01 m per pixel and aligned within 0.1 m using four visual landmarks.
Given the data from the images, the AI can interpret them as an “obstacle”, “smooth road”, “rough road”, which predicts the likelihood of collisions along the way.