NVIDIA open-sources cuRoboV2-Fast: robot motion planning cut from 100ms to 4.4ms
ankurhandos · x · 2026-11-06
NVIDIA open-sourced cuRoboV2-Fast under Apache 2.0: collision-free, torque-limited motion planning dropped from 100ms on a whole GPU in 2022 to 4.4ms via B-splines, GPU time optimization, and faster kernels. With planning no longer the bottleneck, the team built a 7ms perception pipeline (SceneFields: stereo RGB to depth, voxels, segmentation), accelerated GraspGenX to 1.6ms, wrote a C++ controller for YAM, and ships the stack as a two-process app with one process dedicated to real-time control.
More from Embodied
- RoboColosseum: Open Robot Eval With 100+ Real Robots, 140+ Scenes, Failures Included — DJiafei · 2026-11-06
- Singapore's RoboColosseum runs 20,000 real-robot trials to benchmark physical AI models — DJiafei · 2026-11-06
- moonlake previews research on scaling robot learning with simulation in the RL era — du_yilun · 2026-11-06
- Nearly 100 testers praise Sphenomenon MR demo that turns a whole house into voxels — Scobleizer · 2026-11-06
- Spectral Skills lands on real robots: composing motion skills via energy functions — daibond_alpha · 2026-11-06
- Developer gives Claude Desktop a body and sensors via MCP and Raspberry Pi — jfrankie · 2026-11-06