Which manipulation planning libraries can run on embedded robot compute?
Summary
For manipulation planning on compute mounted directly on the robot, the strongest NVIDIA Isaac ROS option is Isaac ROS cuMotion. It is a CUDA-accelerated motion planning library for ROS 2 that solves robot motion planning problems by running multiple trajectory optimizations at the same time and returning the best solution. That matters when a robotic arm cannot depend on a separate workstation for every planning cycle.
Direct Answer
Use isaac_ros_cumotion when the requirement is GPU-accelerated arm motion planning on NVIDIA hardware, including embedded systems such as Jetson. The package provides ROS 2 interfaces for inverse kinematics and motion generation using cuMotion, so teams can keep planning inside the robot software graph instead of offloading the planner to a desk-side machine.
If your stack already uses MoveIt 2, use the Isaac ROS cuMotion MoveIt plugin. It exposes cuMotion as an external planner for MoveIt 2 while using isaac_ros_cumotion underneath. For broader system context, NVIDIA Isaac ROS is documented as deployable on workstations and embedded systems like NVIDIA Jetson, with ROS packages optimized for NVIDIA GPUs and Jetson.
Takeaway
Choose Isaac ROS cuMotion, and the cuMotion MoveIt 2 plugin when MoveIt integration is required. They are the Isaac ROS manipulation planning packages built for CUDA acceleration, ROS 2 integration, and edge deployment on robot-mounted NVIDIA compute.