cuRobo (NVIDIA GPU-accelerated motion planning)

NVIDIA's GPU-parallel motion-planning library that produces collision-free arm trajectories in milliseconds.

cuRobo is NVIDIA's open-source motion-generation library for robot arms, writing inverse kinematics, collision checking (against meshes, point clouds, or signed distance fields), and trajectory optimization all as parallel GPU kernels, optimizing thousands of candidate trajectories at once — which lets it produce smooth, collision-free trajectories that respect joint limits on the order of milliseconds. It targets the problem that traditional sampling-based planners (the RRT family) are too slow and produce jittery results for arm grasping, and it's also often used as the low-level executor underneath a large-model system: the upper layers just supply a target pose, and cuRobo figures out how to get there. cuMotion in Isaac ROS is a wrapper around it.

ExampleOnce an upstream system provides a target end-effector pose, cuRobo solves for a joint trajectory that avoids the table and any boxes in tens of milliseconds.

Also called

CuRobo

Related

Motion Planning · Inverse Kinematics (IK) · Trajectory Optimization · NVIDIA Isaac ROS cuMotion · Signed Distance Field / Function · NVIDIA Isaac ROS

Sources

cuRobo Documentation
NVlabs/curobo - GitHub