RoboKit
GPU toolkit that solves thousands of robot IK, planning and retargeting problems at once
License
MIT
Code: MIT, per LICENSE in the RoboKit, RetargetingKit and GraspKit repositories. Robot models and datasets are downloaded from the project's Hugging Face datasets and keep their own licences: several are CC BY-NC or CC BY-NC-SA (for example the Inspire hand model and DexYCB), and AMASS, ARCTIC, GRAB, MANO and SMPL-X data have restricted research terms. Some models list no licence (for example the XHand), as the project's own asset licence list says. Check the asset licence list before commercial use or redistribution. Checked 2026-11-07 against LICENSE and the asset licence list.
Overview
RoboKit is an open-source Python toolkit for robot optimization from Xiaolong Wang's group at UC San Diego; its first author is Jianglong Ye. It loads a robot from URDF or MJCF, computes batched and differentiable kinematics on the GPU, and solves tasks written as sums of cost terms with sampling and gradient optimizers, from a single problem up to thousands in parallel. Built-in tasks cover inverse kinematics, collision-free motion planning, hand and humanoid motion retargeting, hand-eye calibration, grasp synthesis and model predictive control through MuJoCo Warp. It is the base of two companion toolkits released with it, RetargetingKit for teleoperation and GraspKit for dexterous grasping. Version 1.0.0 is on PyPI, and according to the authors the paper describing it was accepted as a spotlight at CoRL 2026.
RoboKit treats most of the small optimization problems in robotics as one kind of problem. A robot is loaded from URDF or MJCF into batched, differentiable kinematics; a task is a set of cost terms such as pose targets, collisions, joint limits and contacts; and a population solver moves many candidate solutions at once with Levenberg-Marquardt, L-BFGS or gradient descent, or samples them with MPPI-style and evolution-strategy schedules. The optimization guide shows how these pieces compose. Kinematics and geometry run on NVIDIA Warp, PyTorch is optional, and the same code runs on one problem or thousands in parallel, on the GPU when one is present.
The built-in tasks are inverse kinematics with fixed or floating bases, offline and online motion planning around obstacles, hand, humanoid and full-body motion retargeting with optional object contacts, camera calibration from rendered robot masks, and force-closure grasp synthesis for any hand. With the MuJoCo Warp extra it also does sampling-based model predictive control through the simulator, and the examples include a Sharpa hand spinning a pen, a Unitree G1 climbing onto a box and a Go2 landing jumps on chosen footholds. A reinforcement learning path trains policies such as a Franka assembling NIST board parts and a Go1 walking stairs. Robot models and datasets are fetched from the project's Hugging Face dataset and pinned to one revision, covering arms such as the Franka Panda and xArm, a dozen humanoids and many robot hands and grippers. Interactive galleries in Viser run straight from the package, as the quickstart shows.
RoboKit was announced on 6 November 2026 as part of OpenRoboKit, together with two toolkits built on it. RetargetingKit maps live or recorded human motion to robot hands, arms and humanoids, from Apple Vision Pro, PICO, Manus gloves or motion capture, for real-time teleoperation or batch retargeting, with targets that include xArm arms and Inspire and Ability hands. GraspKit runs a full grasp pipeline: it optimizes grasps on the GPU, renders point clouds, trains a policy that maps a point cloud to a grasp trajectory and deploys it in ManiSkill or on hardware; its real-robot presets pair xArm6 and xArm7 arms with XHand, Sharpa and Ability hands. The team's announcement shows a 16-hour outdoor grasping run in San Diego with this setup. A fourth repository, RoboSkills, packages agent workflows for robotics. In its README the team reports large wall-clock speedups over the tools it compares against, among them PyRoki for inverse kinematics, cuRobo for motion planning, GMR for humanoid retargeting and mjlab for time to train an RL policy; the benchmark details are in the paper, which is not yet public.
Why it matters
Robot learning pipelines run the same small optimizations again and again: inverse kinematics for every frame of a retargeted motion, a grasp for every object in a dataset, a plan for every start and goal. Today each of those usually means a different library with its own robot loader, such as PyRoki or mink for kinematics, cuRobo for planning and a separate retargeting tool. RoboKit puts them on one batched, differentiable base, so a lab can generate retargeted motions or grasp datasets at scale with one dependency and then reuse the same robot model for planning and calibration. The companion toolkits show that the base reaches real hardware: GraspKit's policies run on xArm arms with dexterous hands, and RetargetingKit drives hands and humanoids live. For readers of our frax and PyRoki pages it is the newest entry in GPU-batched robot kinematics, with a wider set of tasks; for mjlab and MuJoCo users it adds sampling MPC on MuJoCo Warp. It is days old, so the speed claims and the breadth still need outside users to confirm them.
Problem it solves
Inverse kinematics, motion planning, retargeting, calibration and grasp synthesis are usually solved one problem at a time, each with its own library and robot description. RoboKit expresses them as cost terms over one batched, differentiable robot model and solves thousands of them in parallel on the GPU, so producing large amounts of robot data or solving many problems at once stops being the bottleneck.
Key takeaways
- One toolkit for inverse kinematics, motion planning, retargeting, hand-eye calibration, grasp synthesis and MPC
- Tasks are sums of cost terms, solved for one problem or thousands in parallel with sampling and gradient optimizers
- Batched, differentiable kinematics from URDF or MJCF on NVIDIA Warp; PyTorch is optional
- Sampling MPC through MuJoCo Warp and an RL training path, with examples on hands, humanoids and quadrupeds
- RetargetingKit and GraspKit build teleoperation and dexterous grasping on top of it, tested on xArm arms with robot hands
- Version 1.0.0 on PyPI under MIT; robot models and datasets come under their own licences
Who it is for
- Robot learning researchers who generate retargeted motions, grasps or trajectories in bulk
- Teams teleoperating dexterous hands or humanoids from VR, gloves or motion capture
- Users of PyRoki, cuRobo or mink who want one batched base for several tasks
- Labs working on dexterous grasping with xArm arms and robot hands
Limitations
- Announced on 6 November 2026 by a single lab, so it has had little outside use yet
- The speedups are the authors' own measurements; the paper with the benchmark details is not yet public
- Robot models and datasets carry their own licences: several are non-commercial (CC BY-NC) and some motion and hand data (AMASS, ARCTIC, GRAB, MANO, SMPL-X) have restricted research terms, as the project's asset licence list states
- GPU-heavy features need NVIDIA hardware: GraspKit's optimization, rendering and training need a CUDA GPU, and camera calibration needs nvdiffrast built for your exact PyTorch and CUDA
- GraspKit and RetargetingKit are at version 0.1.0, and their real-robot paths are set up for specific rigs: xArm arms with XHand, Sharpa or Ability hands