// the take
cuRobo (v2) is NVIDIA's CUDA-accelerated motion generation stack for robot arms and high-DoF humanoids, covering kinematics, collision checking, trajectory optimization, and GPU-native perception (TSDF/ESDF) in one pipeline. It's aimed at robotics researchers and engineers who need batched, GPU-parallel motion planning for sim-to-real pipelines or large-scale training, not at someone who just wants a quick IK solver.
The whole pipeline is GPU-parallel end to end — IK, collision checking, and trajectory optimization all batch on the GPU rather than bolting a GPU solver onto a CPU planner. The perception side (TSDF/ESDF via parallel banding and jump-flooding algorithms, plus marching cubes extraction) runs natively on GPU too, which most mapping libraries don't bother with. There's real low-level engineering here: custom CUDA kernels for RNEA dynamics, B-spline trajectories, and self-collision, plus a Warp backend alongside the pybind/CUDA one, not just a thin Python wrapper over an existing solver. It's backed by an actual paper (arXiv 2603.05493), so the algorithms have been through review, not just shipped as code.
v2 is a breaking rewrite of the public API with no migration path — the README just tells you to pin to the old v0.7.8 tag if you depend on v1, so anyone with an existing integration eats a real port. The README itself is thin for a library this complex: no install instructions, no quickstart code, nothing beyond a bullet list of capabilities and a pointer to external docs — you have to leave the repo to find out how to actually use it. There's a hard dependency on a specific NVIDIA stack (CUDA, PyTorch, Warp) with no CPU fallback, and the README doesn't state which CUDA/driver versions are actually required. The internal structure (_src split across cost, optim, perception, graph_planner, dozens of .cuh kernel files) means debugging anything past the public API means reading real CUDA code, not just Python.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →