cuRobo: CUDA Accelerated Robot Library

[API] [GTC Talk] [Code] [Paper]

A unified, dynamics-aware GPU stack for safe, feasible, and reactive robot motion generation, scaling from single-arm manipulators to high-DoF humanoids. cuRobo provides trajectory optimization, GPU-native perception, and whole-body computation built on PyTorch, CUDA, and Warp.

GPU-native perception from dual RGBD streams, integrating C-RADIO features into TSDF blocks in 2ms.
Whole-body policy trained in MimicKit (Newton) using retargeted motions from cuRoboV2.

Updates

  • April 27, 2026: Add feature channel integration in TSDF mapper, improve TSDF mapper performance (1.5ms to 0.5ms).

  • April 2026 (v0.8.0): cuRoboV2, complete rewrite of cuRobo with new features and improvements.

  • 2024 (v0.7.x): Sub-millimetre IK precision, mimic joints, ESDF voxel collision checker, Grasp API planning, trajectory re-timing.

  • Oct 2023 (v0.6.0): Initial release with Technical Report.

See CHANGELOG for details.

Documentation

  • Getting Started: Step-by-step tutorials covering installation, robot setup, kinematics, IK, MPC, motion planning, and volumetric mapping.

  • Guides: Writing custom optimization problems, cost terms, and optimizers.

  • Concepts: How cuRobo’s rollouts, optimization solvers, and geometric planner work.

  • Reference: API, style guide, runtime configuration, benchmarks, sphere fitting, and self-collision.

Research

Read the cuRobo technical reports for algorithm details and benchmarks.

Technical Reports