ROS 2 packages for deploying LEAPP-exported neural-network policies on real and simulated robots.
Isaac ROS Deploy provides ROS 2 packages for deploying neural-network policies on real and simulated robots. It can support a variety of robot-control policies, including policies derived from reinforcement learning in Isaac Lab and vision-language-action (VLA) policies exported through nvidia-isaac/gr00t-leapp-export, as long as the policy can be exported as a LEAPP bundle.
Isaac ROS Deploy bridges the gap between a Python training stack and a robot
control system: it loads LEAPP bundles, runs ONNX inference through NVIDIA
Triton, maps policy terms onto ROS topics or ros2_control interfaces, and
optionally gates outputs through a safety controller.
Use Isaac ROS Deploy when you need to deploy a policy as either a ROS 2 node
graph or inside a ros2_control loop.
Please visit the Isaac ROS Documentation to learn how to use this repository.
Update 2026-09-21: Added optional InferenceController debug topics for flattened model inputs and the selected output tensor, and added LEAPP runtime, robot description, ROS communication, and safety controller troubleshooting guides. Isaac Sim deployment support is not included in this release.