redtail

Why we included this project

For anyone building a drone or rover that has to navigate cluttered outdoor terrain without GPS or a map, this NVIDIA project is one of the more complete references around. It pairs the TrailNet neural network for visual navigation with a Stereo DNN that estimates depth from a stereo camera, and includes the control code and hardware assembly instructions needed to get a working vehicle running on Jetson hardware. The training pipeline is documented and pre-trained weights ship with the repo, so you can either retrain for your own environment or start from existing checkpoints. Development has been quiet since 2020, so treat it as a well-documented reference architecture and research baseline rather than an actively maintained product. The linked papers and demo videos, including a one-kilometer autonomous flight, make it easier to judge whether this vision-based approach.