MINK-WARP

mink-warp is batched differential inverse kinematics on MuJoCo Warp, with a Mink-shaped API.

Given a robot’s current configuration and a stack of task-space objectives, mink-warp computes joint velocities (or configuration updates) that reduce weighted task error — for many worlds in parallel on the GPU.

Purpose

mink-warp exists for real-time, GPU-batched differential IK in MuJoCo Warp pipelines — not as a replacement for Newton IK [Newton]. It targets the same niche as Mink (composable tasks, velocity output, control loops), scaled to nworld on device. See Why mink-warp (and not Newton IK)? for a full comparison with Newton IK and when to use each.

Key features

• Mink-shaped API — Configuration, FrameTask, RelativeFrameTask, EqualityConstraintTask, solve_ik(), and SE3 / SO3 helpers.

• Hard limits — ConfigurationLimit, VelocityLimit, CollisionAvoidanceLimit, LinearInequalityLimit via ConstrainedSolver.

• Device-native hot path — FK, Jacobians, residual assembly, and linear solves run as Warp kernels on wp.array buffers; NumPy only at boundaries.

• Batched by design — nworld parallel instances share one model and one launch grid (multi-agent IK, parameter sweeps).

• Interchangeable solvers — damped least squares (default), Levenberg–Marquardt, L-BFGS, and constrained ADMM.

• CUDA graph capture — optional one-step solve_and_integrate graph for fixed task sets (disabled when a task/limit requires host reads).

• Runnable demos — numbered examples/01_… through 05_… (mjviser); see Examples.

Minimal example

import mujoco
import mink_warp as mw

model = mujoco.MjModel.from_xml_path("robot.xml")
cfg = mw.Configuration(model, nworld=512, device="cuda")

frame = mw.FrameTask("ee", "site", position_cost=1.0, orientation_cost=1.0)
frame.set_target_from_configuration(cfg)
posture = mw.PostureTask(model, cost=1e-2)
posture.set_target_from_configuration(cfg)

solver = mw.DLSSolver(cfg)
solver.solve_and_integrate([frame, posture], dt=0.01, use_graph=True)

Table of Contents

Getting Started

• Installation
• Quickstart: single-world IK
• Batched IK
• Examples

Concepts

• Concepts

User Guide

• Tutorial
• Solver backends
• Constrained IK (hard limits)
• CUDA graphs
• Benchmarks

API Reference

• API Reference

Further Reading

• References
• Roadmap

License

mink-warp is licensed under Apache-2.0. See the LICENSE file.