Licences, platforms and the comparison grids

The cross-cutting document: what you are allowed to ship, what will run on your machine, how gripping fits into ROS 2, and every approach in this area side by side.

It exists separately because these questions do not belong to the hardware, to the geometry or to the models. They are the same questions whichever you are working on, and answering them three times is how the three quietly disagree.

If you read one section of this area, read section 1. Gripping has the worst licence hygiene of any subject in these documents, and the specific failure — a popular repository with no licence file at all — is the one people are least equipped to notice, because it produces no badge, no warning and no error.

Contents#

1. Licences, and the four traps in this area
2. What runs on an Apple Silicon Mac
3. Putting it in ROS 2
4. The comparison grids

1. Licences, and the four traps in this area#

Every licence in this document was read from the project's own licence file in September 2026, using the GitHub licence API or by listing the repository root. Not from a badge, not from a README, and not from a blog post. That distinction produced a different answer in a dozen cases.

1.1 No licence at all#

This is the commonest problem in learned grasping and the most serious, because default copyright grants nothing. Not commercial use, not research use, not the right to modify. A repository with no licence file is not permissive; it is closed.

The projects below have real traction and no licence file:

Seven projects, thousands of stars between them, and no grant of rights on any of them. Anyone who has built on these has done so without permission, usually without realising there was a question.

1.2 A licence you cannot read#

NVlabs/contact_graspnet, one of the most cited grasp models in the field, ships its licence as a file called License.pdf. GitHub's licence API reports nothing. Every dependency scanner you run will classify the project as unlicensed. To know the terms you have to download and open a PDF, which nobody does before a prototype turns into a product.

The maintained PyTorch reimplementation, elchun/contact_graspnet_pytorch, copied the PDF rather than replacing it.

1.3 A licence that contradicts itself#

The GraspNet-1Billion datasets page states that all data, labels, code and models are "licensed under a Creative Commons Attribution 4.0 Non Commercial License (BY-NC-SA)". That names three mutually inconsistent things in one sentence. CC BY 4.0 permits commercial use. "Non Commercial" does not. BY-NC-SA is a fourth licence again, adding a share-alike obligation that neither of the others carries. There is no identifier you can put in a dependency manifest that is faithful to that sentence.

The only safe reading is the most restrictive one, and the page's own commercial contact email tells you what the authors meant.

A related case: graspness_unofficial carries, verbatim, the same Shanghai Jiao Tong University non-commercial agreement as the official baseline — including a clause stating that any derivative you create becomes owned by the licensor, and a trademark clause forbidding use of the name "AlphaPose", which is a completely unrelated piece of software. A licence that has been copied without being read is a licence nobody thought about.

1.4 Code and weights under different licences#

Checking the repository is not checking the model. NVIDIA's GraspGen is the clearest case and it splits four ways:

The successor repository is more permissive than the original, which is the opposite of the usual direction. The weights you would actually run are still under NVIDIA's own model licence, so the Apache badge on the code does not make a deployment Apache-2.0.

GraspVLA is the same shape with the pieces in different places: no licence on the code, CC BY-NC-4.0 on the weights, and no licence tag at all on its billion-sample training dataset.

And there is a case of a third party granting rights they do not hold. A Hugging Face repository mirroring Contact-GraspNet's weights declares them MIT. The upstream licence is an NVIDIA agreement in a PDF and is certainly not MIT. A licence field on a mirror is not a licence.

1.5 Three more that are specific to this area#

Two vendors ship their own drivers under the GPL. Schunk's mechatronic gripper driver, their SVH hand driver and their force-torque sensor driver are all GPL-3.0, as is GelSight's gsrobotics SDK. Robotiq, OnRobot and Franka all use permissive licences, so this is a vendor decision rather than an industry norm — but if you link a GPL driver into a product, the obligation to publish your source follows.

A non-commercial Creative Commons licence on a hardware driver. Meta's digit-interface is CC BY-NC 4.0 and archived. The DIGIT sensor is sold; its driver may not be used in a product. A Creative Commons licence is also a poor fit for software, addressing neither patents nor source distribution.

A permissive licence covering a paid dependency. Two well-starred dexterous grasping projects, GraspXL and RobustDexGrasp, vendor the RaiSim physics simulator. Their licence file makes the MIT grant apply only to the wrapper directories and states that you must obtain a licence from raisim.com for the simulator, which is then activated with a key. The physics engine these projects need is paid software, and nothing on their GitHub pages says so. Isaac Lab has the same shape: a BSD-3 wrapper pinned to proprietary Isaac Sim.

One more that catches people who only want a simulation model. mujoco_menagerie's root licence file is a machine-generated concatenation of per-model licences, and GitHub reports the repository as unclassified. The gripper and hand models in it differ: the Robotiq 2F-85 model is BSD-2-Clause from ROS-Industrial, the Wonik Allegro model is BSD-2-Clause from SimLab, the Shadow Hand model is Apache-2.0, and the LEAP Hand and UMI gripper models are MIT. Read the per-model file.

The LEAP Hand is worth stating on its own, because the answer changes with the artefact. Its MuJoCo model is MIT and its driver API is CC BY-NC 4.0. You may simulate the hand commercially and you may not drive the real one.

1.6 What you can actually ship#

The short list, for a commercial product, verified from the files:

If you are doing research, most of the non-commercial licences permit what you are doing. The seven projects in section 1.1 do not, and that is worth knowing before a paper's artefact review asks.

2. What runs on an Apple Silicon Mac#

The honest headline for this area, unlike the perception area's, is that almost no learned grasp model does. The reason is more specific than people expect and it is worth knowing exactly, because it tells you what to check first on anything new.

2.1 One abandoned library is the gate#

MinkowskiEngine is NVIDIA's sparse convolution library. It is MIT-licensed, it has around 2,960 stars, and it was last pushed in March 2024 with 236 open issues. Its stated requirement is CUDA 10.1 or later, matching the CUDA version PyTorch was built against, and it installs by compiling with nvcc.

graspness_unofficial, anygrasp_sdk and EconomicGrasp all import it — three of the four significant models in the GraspNet-1Billion lineage. So the obstacle for a Mac is not the vague "grasping needs a GPU". It is one abandoned sparse-convolution library that three leading models depend on.

The other blockers, each fatal on its own, are custom pointnet2 CUDA operators, spconv-cu120, flash-attn and xformers. None has a CPU or Metal build.

2.2 The table#

Two of those rows carry a caveat rather than a yes. GPD is CPU-friendly and was last pushed in January 2022, so its OpenCV 3.4 and PCL dependencies are the practical obstacle rather than the GPU. GG-CNN is plain PyTorch and was last pushed in July 2020, so its Python 3.6-era pinning is the obstacle.

2.3 The pattern worth carrying#

A clean requirements.txt is not evidence that a project will install. graspnet-baseline's requirements file lists only torch, tensorboard, numpy, scipy, open3d, Pillow and tqdm, which looks entirely portable. Its README then tells you to compile pointnet2 operators and a CUDA knn operator by hand. Dependency scanners do not read prose. Reading the install instructions is the only reliable check.

2.4 What this means in practice#

Develop the geometry on the Mac and the model elsewhere, or do not use a model. The methods in choosing a grip run natively and are tested in a second with no simulator and no weights; MuJoCo and its gripper models run natively; the ROS 2 control stack runs through RoboStack, which is what this repository uses. A rule-based gripping pipeline can be complete, property-tested and working before any model is involved.

3. Putting it in ROS 2#

Gripping in ROS 2 is smaller than most people expect, and knowing its actual extent saves a lot of searching.

3.1 The message types#

A gripper is commanded through control_msgs (BSD-3-Clause). Two actions matter.

GripperCommand.action is the classic. You send a position and a maximum effort, and the result carries four fields: the current position, the current effort, and two booleans, stalled and reached_goal. The definition's own comment on stalled is "True iff the gripper is exerting max effort and not moving".

Those two booleans are the whole of ROS 2's grasp feedback, and they say exactly what holding on, section 2.3 warns about. A successful grasp appears as stalled true and reached_goal false: the fingers stopped early and are pushing. So does a grip on the wrong object, a grip on two objects, a grip on the tote wall, and a jammed finger. The message has no way to express what stopped the fingers, because the gripper does not know.

ParallelGripperCommand.action is the newer one, carrying a sensor_msgs/JointState so that position, velocity and effort limits can be set per joint. It returns the same two booleans.

3.2 The packages#

Three things about that list are worth stating outright.

There is no impedance controller in upstream ros2_controllers. Admittance only. If you want impedance control, crisp_controllers is the maintained permissive option and franka_ros2's examples are the vendor one. The difference between the two kinds, and why it is decided by your hardware, is holding on, section 3.

MoveIt 2 has no grasping and no force control. It plans collision-free arm motion to a pose you supply. There is no grasp package and no compliance package anywhere in the repository. This surprises people because MoveIt is what the pick-and-place tutorials use; the pick in those tutorials is a taught grip and a fixed squeeze. moveit_grasps filled the gap and has not been pushed since November 2022. moveit_task_constructor is the live route, and it sequences stages rather than choosing grips.

Gazebo Harmonic has no suction or vacuum gripper. Its full system list includes contact, detachable_joint, force_torque, optical_tactile_plugin and touch_plugin, and no gripper system of any kind. A simulated suction gripper in Harmonic is something you build from detachable_joint plus contact. Every "Gazebo vacuum gripper" repository on GitHub is either Gazebo Classic or an unlicensed personal fork.

3.3 Where the vendor drivers fit#

Every gripper driver in grippers and hardware, section 9 presents itself as a ros2_control hardware interface, so the controller above is the same whichever gripper you have. That is the right layering and it works. What it does not give you is any of the semantics: the driver reports finger position and the controller reports stalled, and everything about whether the grasp was correct is yours to build.

4. The comparison grids#

Everything in one place. These are for choosing, not for benchmarking, so speeds are orders of magnitude rather than measured figures.

4.1 The six mechanisms#

Read this as: what it holds, how much, how fast, and what disqualifies it.

4.2 Choosing where to grip#

4.3 The grasp models, with their licences#

4.4 Holding on#