// the find
NVIDIA-AI-IOT/redtail
Perception and AI components for autonomous mobile robotics.
Redtail is NVIDIA's reference implementation of TrailNet, a DNN that steers drones and ground robots along forest trails and sidewalks, plus a Stereo DNN depth model. It runs on Jetson TX1 and TX2 boards with ROS nodes and TensorRT. It is for people building visual-navigation rigs on Jetson hardware who want working code and pretrained weights, and it stopped development in 2020.
- The pretrained TrailNet SResNet-18 caffemodel and the dataset scripts in models/dataset make the training pipeline reproducible from public data, which most navigation papers do not provide.
- The stereoDNN directory includes custom TensorRT plugins for cost volume, softargmax, conv3d and transposed conv3d, with FP16 and INT8 paths and serialized engines. The Jetson inference work is implemented here, not sketched.
- The ROS launch files for gscam, ZED, and MAVROS, along with the PX4 controller node and simulation Docker image, cover a practical path from bench testing to an airframe.
- It targets JetPack 3.x, TensorRT 2 to 4, Caffe, and Kinetic-era ROS. Building it on a current Jetson stack means porting the plugins and Caffe layers, and the README gives no migration path.
- The last push was November 2020, and the README's news section says the project is no longer developed. Issues and PRs will likely go unanswered, and the README points to a community fork (mtbsteve/redtail) as the continuation.
- Binary artifacts such as .caffemodel files, TensorRT .bin weights, and a prebuilt TX2 .plan are checked in, and most setup instructions live in the external wiki. The repo alone does not give you a reproducible build.