PyRoki

Modular Python toolkit for kinematic optimization across CPU, GPU, and TPU via JAX

License

MIT

Overview

A modular, extensible, and cross-platform Python toolkit for kinematic optimization. Unifies inverse kinematics, trajectory optimization, and motion retargeting with composable costs and constraints. Runs on CPU, GPU, and TPU via JAX. Supports URDF loading with automatic differentiation, collision detection, and multiple solver methods including Levenberg-Marquardt and Augmented Lagrangian. Developed at UC Berkeley.

PyRoki is a modular Python toolkit that provides a unified framework for solving diverse kinematic optimization problems in robotics, including inverse kinematics, trajectory optimization, and motion retargeting between different kinematic structures. Developed at UC Berkeley, the toolkit leverages JAX as its computational backend, enabling seamless execution across CPU, GPU, and TPU hardware while providing automatic differentiation for gradient-based optimization. PyRoki supports standard URDF file loading for robot description, built-in collision detection capabilities, and implements the Levenberg-Marquardt algorithm for robust convergence. By unifying multiple kinematic problem types within a single framework, PyRoki eliminates the need for researchers to maintain separate specialized tools for different motion planning tasks, streamlining workflows for applications ranging from teleoperation to humanoid retargeting.

Why it matters

PyRoKi provides a clean, modern toolkit for robot kinematics that works across platforms. Its composable optimization approach unifies IK, trajectory planning, and motion retargeting — tasks that previously required separate, incompatible libraries.

Problem it solves

Robot kinematics tools are typically platform-specific and solve narrow problems. PyRoKi provides a unified optimization framework where IK, trajectory planning, and retargeting are all expressed as composable optimization problems.

Key takeaways

• Modular Python toolkit for kinematic optimization
• Composable costs and constraints for flexible problem formulation
• Unifies IK, trajectory optimization, and motion retargeting
• Cross-platform with no heavy dependencies
• Extensible architecture for custom kinematic problems

Who it is for

• Robotics engineers solving kinematics problems
• Researchers in motion retargeting and teleoperation
• Developers needing lightweight IK without ROS dependency
• Teams building cross-platform robot software

Limitations

• Optimization-based approach slower than analytical IK for simple robots
• Newer library with smaller community than established alternatives
• Limited to kinematic problems — no dynamics or force control
• Documentation still growing