cuRobo: a CUDA Accelerated Robot Motion Generation Toolkit
, Senior Research Scientist, NVIDIA
, Senior Research Scientist, NVIDIA
Global motion generation for manipulators remains a very slow process, taking anywhere from a few seconds to minutes, predominantly done on CPU. We will present cuRobo, a CUDA-accelerated library containing a suite of robotics algorithms that can generate minimum-jerk, collision-free trajectories within 30 milliseconds leveraging parallel compute on a NVIDIA RTX 4090. cuRobo leverages GPU compute to run optimization over many seeds in parallel, converging to good solutions for many robotics optimization problems, including collision-free inverse kinematics and trajectory optimization.
cuRobo provides fast implementations of kinematics, collision-free inverse kinematics, trajectory optimization, graph planning, batched numerical optimization solvers (L-BFGS, MPPI), and global motion generation. cuRobo also provides fast collision functions to query signed distance between a robot and the world represented by cuboids, meshes (warp), and depth images (nvblox), leveraging several NVIDIA technologies, all on the GPU. cuRobo is integrated with PyTorch, so you can integrate either the full stack for motion generation or only sub modules as needed. We'll explain the approach that cuRobo takes to solve the global motion generation problem, the API design of the toolkit, and provide example integrations with NVIDIA Isaac Sim.