Licences, platforms and the comparison grids

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

It exists separately because these questions do not belong to reaching, or to planning, or to controlling. They are the same questions whichever part you are working on, and answering them three times is how the three quietly drift apart.

Every licence below was read from the project's own LICENSE file or package manifest, fetched in September 2026, and never from a badge or a blog. Every link was checked. Where code and weights differ, both are named.

If you read one section, read section 1. The default inverse kinematics solver in the main open-source motion planning framework is LGPL, not BSD, and almost nothing says so.

Contents#

1. Licences, and the four that will catch you out
2. The libraries, and what state each is in
3. What runs on an Apple Silicon Mac
4. ROS 2: distributions, packages and versions
5. The comparison grids

1. Licences, and the four that will catch you out#

The good news first. The core of this area is permissive and clean. MoveIt 2, OMPL, the Pilz industrial motion planner, MoveIt Task Constructor, moveit_servo, TRAC-IK, pick_ik and bio_ik are all BSD-3. ros2_control, ros2_controllers, the admittance controller and the motion primitives controller are all Apache-2.0. Pinocchio is BSD-2, Drake is BSD-3, and — the one that surprises people — cuRobo is Apache-2.0, despite being NVIDIA research code, which is usually not.

Four things are not clean, and each catches people in a different way.

1.1 MoveIt's default IK solver is LGPL-2.1#

KDL, the Kinematics and Dynamics Library, is LGPL-2.1. GitHub's licence detector does not identify it, because the licence text lives in orocos_kdl/COPYING rather than at the repository root, so the repository shows no licence at all in the usual places. The file itself is the GNU Lesser General Public License, version 2.1.

KDL is MoveIt's default kinematics plugin, so a very large number of ROS arm projects have LGPL code in the path between "here is a pose" and "here are the joint angles" without anyone having decided that. The LGPL is much weaker than the GPL and dynamic linking is the case it was written to permit, so for most people this is a thing to know rather than a thing to fix. It is worth knowing because the assumption in the room is invariably "MoveIt is BSD, so this is BSD", and that assumption is wrong about one specific component.

If it matters, the replacements are permissive and better: TRAC-IK and pick_ik are both BSD-3.

1.2 ViSP, the visual servoing library, is GPL-2.0 or later#

ViSP is the reference implementation of visual servoing, it is actively maintained, and its LICENSE.txt is the GNU General Public License version 2 or later. The ROS wrapper vision_visp is the same code.

Visual servoing is a small enough corner that there is no permissive equivalent of comparable depth, so the practical positions are to use it and accept the obligation, to write the specific controller you need yourself — image-based servoing on a known feature is not a large piece of code — or to buy it.

1.3 Ruckig is MIT and calls a cloud API#

Ruckig is MIT, and this is not a licence problem. It is a deployment problem hiding behind a licence that looks settled.

Ruckig's own README states that the Community Version supports intermediate waypoints using a cloud API for remote calculation, that it switches to that cloud API as soon as intermediate positions are given, and that this path is not real-time capable. Local calculation of intermediate waypoints is a Pro feature, as are position limits and interrupting a calculation.

A trajectory generator that makes a network call is worth knowing about before it is in a control loop. MoveIt is not in that case — its Ruckig use smooths an existing waypoint sequence rather than supplying intermediate positions — and anyone calling Ruckig directly with intermediate_positions is.

1.4 Isaac ROS cuMotion is proprietary, and cuRobo is not#

This is the same trap the perception area found in Isaac ROS, in a different package, and it is worth restating because the two halves point opposite ways.

cuRobo itself is Apache-2.0, read from its LICENSE file. The ROS 2 packaging of it, isaac_ros_cumotion, declares NVIDIA Isaac ROS Software License in its own package.xml. That is NVIDIA's proprietary licence, not an open one, and GitHub's detector reports the repository as having no recognised licence.

So the library is permissive and the convenient way to use it from ROS is not. Both also require CUDA, so neither runs here at all.

1.5 The patterns worth recognising#

• Code permissive, weights not. GR00T N1.5's repository is Apache-2.0 and its weights are under the NVIDIA One Way Noncommercial License, with the model card stating the model is ready for non-commercial use.
• No licence is worse than a restrictive one. openpi's repository is Apache-2.0 and its checkpoints ship from a Google Cloud Storage bucket with no stated terms. Absent a statement, nothing is granted.
• A licence file at the root need not describe the whole repository. Tesseract ships a LICENSE that says the package contains Apache-2.0, BSD-2 and BSD-3 code, each file marked at the top. It has to be read per file.
• GitHub's detector is not the licence. It misses KDL entirely and reports Isaac ROS cuMotion as unlicensed. Fetch the file.

2. The libraries, and what state each is in#

Read these as: what it is, what its licence is, and what state it is in, verified by its last push in September 2026. A stale library is a slow problem rather than an obvious one.

2.1 Planning#

These are the libraries that decide a route. MoveIt is the framework and the others are either plugins inside it or alternatives to it.

2.2 Kinematics and trajectories#

These turn poses into joint angles, and paths into timed trajectories. The first row is the licence trap from section 1.1.

2.3 Control#

These execute the motion and decide what happens on contact. The ViSP row is the licence trap from section 1.2.

2.4 Reachability#

There are only two tools worth naming and both are quiet, which is itself the finding: reachability analysis has no healthy open-source ecosystem.

2.5 Learned motion#

Every one of these is covered with its weights licence in section 2 of learned motion. The short version: LeRobot is Apache-2.0 and is where the ecosystem now lives; ACT, Diffusion Policy, Octo, OpenVLA, RDT-1B and Motion Policy Networks are MIT; openpi's code is Apache-2.0 and its weights are unstated; GR00T N1.5's code is Apache-2.0 and its weights are non-commercial.

3. What runs on an Apple Silicon Mac#

The headline is better than its reputation, and it is not documented anywhere obvious, so this section is the result of checking rather than of reading.

ROS 2 has no official Apple Silicon support at all. In REP-2000 — whose last distribution section is Kilted Kaiju, so Lyrical's own dates are not in it — defines the target platforms for every distribution, macOS appears only in the amd64 row at tier 3, for both Jazzy and Kilted. The arm64 row has no macOS cell. This is the same finding the perception area reports, and the practical route is the same: RoboStack, a community conda redistribution, which is what this repository uses.

And through RoboStack, nearly the whole motion stack is there. Checked against the robostack-jazzy channel: ros-jazzy-moveit 2.12.4, ros-jazzy-moveit-py, ros-jazzy-moveit-servo, ros-jazzy-ompl, ros-jazzy-pilz-industrial-motion-planner, ros-jazzy-moveit-task-constructor-core, ros-jazzy-ros2-control 4.47.0, ros-jazzy-ros2-controllers 4.42.1, ros-jazzy-ur-robot-driver, ros-jazzy-octomap, ros-jazzy-behaviortree-cpp and ros-jazzy-py-trees all publish an osx-arm64 build. So planning, Cartesian servoing, industrial motion primitives, task construction and control all run on the Mac, which is more than most people assume.

The table below is what works against what does not. Read the right column as the list of things to plan around rather than to fight.

Two of those deserve a sentence rather than a cell.

LeRobot's Apple Silicon support is in the code, not merely in the README. Its device selection tries CUDA, then torch.backends.mps, and returns torch.device("mps") when Metal is available. Training a small policy and running inference both work on a Mac.

The ViSP split is the awkward one. The library has an osx-arm64 build on conda-forge, and the ROS wrapper has no RoboStack build. So visual servoing on a Mac means calling ViSP directly rather than through ROS, and the GPL obligation from section 1.2 comes with it.

4. ROS 2: distributions, packages and versions#

ROS 2's current release is Lyrical, from May 2026; ros2_controllers carries a ros2_controllers.lyrical.repos alongside its Jazzy and Kilted files. Jazzy Jalisco runs from May 2024 to May 2029 and is the safer choice today, which is what this repository uses. Kilted Kaiju ends in November 2026, so starting on it now is starting on something with two months left. Those dates are from REP-2000, and the distribution list is in rosdistro.

The packages that matter for motion, with the licence of each:

One thing worth stating because it is a category error people make. MoveIt's perception pipeline builds an octomap so the planner can avoid unknown obstacles. It is not a measuring tool and was never meant to be. The perception area says the same thing, and the mistake in this direction is the mirror image: using the octomap as the model of the part you are about to grasp, rather than as the model of the things you must not hit.

5. The comparison grids#

Everything in one place.

5.1 The five kinds of move#

Read this as a decision table: the four middle columns are what each kind of move demands of you, and the last says where it is covered properly.

5.2 Choosing a planner#

Speed is an order of magnitude on ordinary hardware rather than a benchmark figure, because the benchmark figure depends on a card you probably do not have.

5.3 Choosing a controller#

Read this as: what you hand the controller, what hardware that needs, and what it is for. The licence column is the one that decides whether you may ship it.

5.4 The silent failures, in one place#

Every failure in this area that produces no error message, with the setting or the assumption it hides behind and the section that works it out.

Back to the overview.