Hello everyone,
Some of you may remember my earlier post about curobo_ros, a lightweight ROS 2 wrapper around NVIDIA’s cuRobo planner. I’m happy to share that the package has now been migrated to cuRobo v2 (v0.8.0), which was a major rewrite upstream, and the wrapper has been updated accordingly.
What’s new in v2:
- Native perception pipeline: cuRobo v2 ships a block-sparse TSDF + GPU ESDF mapper built in, so we could drop our homegrown
nvblox/nvblox_torchforks entirely. Depth cameras now feed a single unifiedSceneobject instead of juggling separate obstacle/voxel/mesh caches. - Unified
MotionPlannerAPI: single-pose, batch, goal-set, and multi-environment planning are now handled through one consistent interface instead of several separate solver classes. - Dynamics-aware trajectory optimization (B-spline based), and a cleaner composition-based architecture that should make the wrapper easier to extend going forward.
- ROS 2 Jazzy support, in addition to Humble.
- Jetson support: Docker images now also build and run on Jetson AGX Orin with JetPack 7.2, in addition to x86 desktop/server setups.
- Updated prerequisites: Python ≥ 3.10, CUDA ≥ 12, Torch ≥ 2.5 (cu12) / ≥ 2.9 (cu13).
As before, the goal is to keep this a simple, modular, lightweight alternative to the larger stacks, without requiring Isaac Sim.
Repo and docs:
GitHub - Lab-CORO/curobo_ros: wrapper of curobo for ros · GitHub
Migration notes (mostly relevant if you’re extending the wrapper’s internals):
curobo_ros/docs/MIGRATION_V2.md at main · Lab-CORO/curobo_ros · GitHub
Introduction / getting started:
curobo_ros/docs/getting-started/introduction.md at main · Lab-CORO/curobo_ros · GitHub
As always, feedback, issues, and PRs are very welcome!