Triton AI's home-grown, open-source autonomous driving platform in pseudo-ROS for robotic vehicles
Triton-Racer is a computer-vision and deep-learning based autonomous driving platform for robotic cars from 1/16 to 1/5 scales in speed racing.
With current support for 2D camera, odometery, and 2D LiDAR (SICK), and planned support for deapth-perception camera (Intel RealSense) and proximity sensors, we are designing the system towards more general applications of autonomous vehicles, and potentially transform the platform to be used on 1/1 scale autonomous vehicles.
The backend research in deep learning never ceases. Starting from a simple open-loop CNN regression structure for car control, we experiment with a variety of architectures including Categorical CNN, LSTM, and Transformer, some of which yields exciting results in competitions. We’ve also developed computer vision pipelines that assist the feature extraction of neural networks.
We are part of the global DIY robocar community which hosts regular competitions of autonomous robotic vehicles. We fly to different parts of the state with our vehicles, and achieve top results in the races. Even durning the pandemic, with everything happening in the simulator, we are still actively competing against racers joining us online from around the world.