// the find
Zhefan-Xu/NavRL
[IEEE RA-L'25] NavRL: Learning Safe Flight in Dynamic Environments (NVIDIA Isaac/Python/ROS1/ROS2)
NavRL is the official implementation behind an IEEE RA-L 2025 paper on RL-based collision avoidance for drones (and other velocity-controlled robots) in dynamic environments, trained in NVIDIA Isaac Sim via OmniDrones. It's aimed at robotics/RL researchers who want a working UAV navigation stack to build on or benchmark against, not at app developers.
Ships a pretrained model and three ready-to-run demo scripts, so you can see the policy working in a few minutes without touching the training pipeline. Covers both ROS1/Gazebo and ROS2/Isaac Sim deployment paths with a real robot example (Unitree Go2), which is more than most RL navigation papers bother to release. Training config exposes practical knobs (robot count, static/dynamic obstacle count) rather than hiding them in code.
Hard-pinned to Isaac Sim 2023.1.0-hotfix.1, a version no longer available through normal install channels — getting it requires pulling a Docker image and manually copying files out, which is a fragile, easy-to-break setup step. Full-scale training (1024 envs, 350+ obstacles) explicitly requires an RTX 4090, so reproducing the paper's results isn't accessible on typical hardware. The repo vendors OmniDrones wholesale as a third-party dependency inside isaac-training/third_party rather than pinning it as a proper submodule/package, which makes upstream fixes and version tracking harder. No visible tests or CI, which is typical for a paper release but means the ROS1/ROS2 deployment code is unverified beyond manual demos.