What it does

Redtail is a full-stack toolkit for building drones and ground robots that navigate visually through unstructured environments—forest trails, sidewalks, office corridors—without GPS or pre-mapped routes. It bundles deep neural networks (the TrailNet controller and a Stereo DNN for depth estimation), computer vision pipelines, control code, hardware build instructions, and pre-trained models, all tuned for NVIDIA’s Jetson TX1/TX2 embedded boards.

The interesting bit

The project is essentially a research-to-robotics pipeline: TrailNet was published at IROS 2017, Stereo DNN at CVPR 2018, and NVIDIA open-sourced the whole flight-tested system with 1 km autonomous drone demos. Then in 2020, NVIDIA walked away. The twist: the README explicitly points to a community fork (@mtbsteve) that kept the project alive—an unusual admission for a corporate repo.

Key highlights

• TrailNet DNN runs visual path-following; Stereo DNN adds depth from stereo cameras (TensorRT-optimized, ~20 fps on TX2 in FP16)
• Pre-trained models included; training-from-scratch instructions and public datasets documented
• Full simulation environment packaged in Docker with MAVROS integration
• Hardware guides for TBS Discovery drone + ZED stereo camera assembly
• INT8 inference support for compatible hardware

Caveats

• Officially unmaintained since February 2020; JetPack/CUDA/TensorRT versions frozen at 2018–2019 era (JetPack 3.2, TensorRT 4.0)
• Community continuation exists but is unofficial; expect friction with modern Jetson boards
• “Experimental” APM Rover support noted in 2017; unclear if matured before abandonment

Verdict

Grab this if you’re a robotics researcher or stubborn hobbyist who wants a proven, citation-backed visual navigation baseline and doesn’t mind archaeology. Skip it if you need active upstream support or are targeting current Jetson hardware without porting work.