NVlabs/curobo

What it solves

cuRobo addresses the computational bottlenecks in robot motion generation, particularly for high-degree-of-freedom (DoF) robots like humanoids. It enables fast, collision-free motion planning and trajectory optimization that accounts for robot dynamics and environmental obstacles in real-time.

How it works

Built on PyTorch, CUDA, and Warp, the library uses GPU-parallel algorithms to handle complex robotic tasks. It implements GPU-native perception to create dense signed distance fields (ESDF) from depth images for fast collision checking, and employs B-spline representations for trajectory optimization to ensure smooth movements while respecting torque limits.

Who it’s for

This library is designed for robotics researchers and engineers working with complex robotic systems, ranging from single-arm manipulators to high-DoF humanoids, who require high-performance motion generation and perception.

Highlights

• GPU-Accelerated Kinematics: Parallel algorithms for forward and inverse kinematics.
• Dynamics-Aware Optimization: Trajectory optimization that enforces smoothness and torque limits.
• High-Speed Perception: GPU-native ESDF generation from depth images, performing up to 10x faster than existing state-of-the-art methods.
• Scalable Whole-Body Computation: Topology-aware kinematics and map-reduce self-collision checking for high-DoF robots.