URDF export issue: missing base & collapsed gripper

[d-kleine]

I’m working with Isaac Sim 6.0 GA and have created a robot arm (SO-101) equipped with a parallel gripper. I have exported the robot to URDF, but I’m running into issues when visualizing it in VS Code using the URDF Visualizer extension.

Observed issues

After export:

• The robot base is not displayed at all, not even at the origin.
• The parallel gripper components all appear collapsed at the origin.
• The rest of the kinematic chain appears partially correct, but the base and gripper are clearly wrong.

Context / setup

• All components are modeled as rigid bodies with colliders.

• The base consists of multiple parts.

• The gripper includes:

  • A wheel driven by a joint
  • Subsequent links driven via mimic joints

• The robot origin in the URDF appears to be set at the first revolute joint rather than a dedicated base frame.

• There is a fixed joint connecting the base, but this fixed joint does not appear in the exported URDF.

Suspicious export behavior

While inspecting the generated URDF, I noticed that:

• All base links have:

  scale="0.001 0.001 0.001"
  

  However, in Isaac Sim, all these parts have a scale of (1.0, 1.0, 1.0) in the properties panel.

• This mismatch seems likely related to the missing base in visualization.

Question

Could someone help identify what might be causing these issues in the URDF export? Any pointers to likely misconfigurations or known pitfalls in the UI would be really appreciated.

Thanks in advance!

[rgasoto]

are you able to share a USD that we can use to reproduce the issues?

[d-kleine]

Hi Michael,

Thanks for taking your time to further investigate this issue. With reference to your points:

So the recommended setup is:

1. Keep the world-to-base fixed joint.

2. Move ArticulationRootAPI from base onto that fixed joint.

3. Reopen the Physics Inspector, then export again.

Thank you, I have done so. I have deleted the previous fixed joint and articulation root, and create a new fixed joint at base_group and add the articulation root at that joint. I was not possible to do that at base because it has no rigid body. Now, the physics look correct.

Unfortunately, this didn't resolve this URDF export issue that there base is missing. One thing I have forgotten to add is that the robot model has been imported from a STEP file, and that I have other assets in the scene too. I have create a new Xform and move all robots parts there, and the parts were scaled down too (using scale instead of scale:unitsResolve param). I think I have exported the robot model in the wrong size format (in meters instead of millimeters) before importing it to Isaac Sim. When exporting the whole robot to URDF, this also only scaled the first link. Therefore, I am now 100% sure that my first issue (missing base, in my case base_group) is due to issue of the scaling only for the first link. The part is there, but too small and not visible in the viewport at this size. Btw, the bug has been confirmed in the meantime, see #670 (comment). Once there is a patch that the scaling will be applied to all parts of the robots when exporting to URDF, I think the base will have the correct size.

With reference to my second issue (that some parts of the parallel gripper are collapsed at the origin), this is due to my joints setup. I forgot to mention that the collapsed parts are joints in the USD that are excluded from articulation. The issue here is that I have a parallel gripper, where each arm has both an internal rod as well as an external rod. The internal rods are driven by a servo motor, but the external rods are passive and rotate against pins that are connected to the case of gripper. On each side, both internal as well as external rods are connected to a finger, resulting in an articulation loop. I have not yet found a working solution for these excluded joints that makes it possible to include them into the articulation without ending in a loop. Therefore, when exporting the robot to URDF, the parts of the external rods will be located at the origin in the URDF file.

[michaellin6]

@d-kleine this would be easier to help you debug if we had access to the asset you are having issues with. Would you be able to share the assets with us privately so we can debug it on our side? Github doesn't support that but The NVIDIA Forums does https://forums.developer.nvidia.com/u/michalin/summary Or if you can put together a minimal/stripped down asset you can share and has the same issues, we can also do that.

[d-kleine]

@michaellin6 Thank you being so considerate - I have sent you a PM via the Nvidia Forums. 🙂

[michaellin6]

Hi @d-kleine thank you for sharing the USD asset. After reviewing it here are a couple of notes on the two separate issues here:

1. Gripper collapses at the origin — a closed loop can't be a URDF tree. A parallel gripper is a closed kinematic loop, but URDF is strictly a tree (one root link, one parent per link). To emit a tree the exporter has to break the loops, dropping the loop-closure joints, which leaves some gripper links with no valid parent — so they fall back to the origin. Mimic joints don't save it: a URDF only couples a joint that's already in the tree. Fix: keep one spanning-tree path (wheel → driven finger) as a normal chain with a on the second finger, and drop the loop-closure joints. If you need the true parallel kinematics, URDF can't express it — use MJCF (supports loop constraints) or keep it in USD.

2. Base missing / scale="0.001" — the export mixes units. The properties panel shows (1,1,1) because the links carry a hidden units-resolve scale (xformOp:scale:unitsResolve = 0.001, mm→m) that the panel doesn't surface but the exporter writes as scale="0.001". The real problem is it isn't applied uniformly: you noticed the base links get scale="0.001", but the other links don't — their meshes stay in millimeters (no scale) and their joint origins come out in mm too, so the URDF mixes meters and millimeters. The mm links render ~1000× too big and too far, leaving the correctly-scaled base as a speck at the origin ("not displayed"). Fix: bake the 0.001 into the mesh geometry before export so every link is a true scale=1 in meters, then re-export.

[d-kleine]

Hi @michaellin6,

Thank you for taking a look into the asset! With reference to your points:

Mimic joints don't save it: a URDF only couples a joint that's already in the tree. Fix: keep one spanning-tree path (wheel → driven finger) as a normal chain with a on the second finger, and drop the loop-closure joints.

For the parallel gripper, I have set up a normal joint chain on one side (via the driven wheel and subsequent mimic joints that use the wheel as reference joint). On the other side, I have used mimic joints that mimic their counterparts from the normal side. With this setup, the gripper in the URDF file is now complete (no parts relocated at the origin anymore) - but the physics in the URDF (via visualizer extension) they don't seem to correct. I might need to debug the URDF again or just stick with the USD file...

The properties panel shows (1,1,1) because the links carry a hidden units-resolve scale (xformOp:scale:unitsResolve = 0.001, mm→m) that the panel doesn't surface but the exporter writes as scale="0.001".

Yes, I have noticed that I used the wrong units (mm instead of m) for exporting my robot to a STEP file, that was later imported into Isaac Sim. I was just wondering why only the base was exported in mm, but not the other parts (this seems to be a bug, see #670).

So, thank you again for your time - the issues are (more or less) resolved. ✅👍🏻

[michaellin6]

Sounds good, one example of a closed kinematic chain implemented with mimic joints you can use as reference is the Robotiq 85 2F.

[d-kleine]

I followed exactly this parallel gripper using with the same asset (Robotiq_2F_85_edit.usd) from the Content Browser, and ended up in the collapsed gripper parts. That's why I created this discussion. The Robotiq 2F 85 is also used in the Rig Closed-Loop Structures tutorial. The issue is that the Robotiq 2F 85 has a simple, linear kinematic, where the parallel gripper can be set up by mimic joints that are all driven by the same actuator (finger_joint). In my case, using the same articulation principle for the parallel gripper ended up in the collapsed gripper parts at the origin.

I have also just checked the Robotiq 2F 85 URDF you have provided. It's behavior in the URDF seems to be incorrect there: when driving finger_joint, both fingers move into the same directions instead of snapping together. In Isaac Sim, using the Robotiq_2F_85_edit.usd asset from the Content Browser, exports it as URDF also with an incorrect behavior.

I wasn't sure if I should add this feedback to my last reply, but from my point of view, the Robotiq 2F 85 is a good example for understanding what and how to deal with closed articulation loops. The critical aspect of the tutorial is that it doesn't explain how to correctly setup joints and their articulation of a gripper for exporting to URDF, ensuring that the exported URDF contains the intended structure and kinematics of the gripper.

[d-kleine]

@michaellin6 Would it be possible to pass this feedback along to the team? At the moment, the Isaac Sim / Lab docs explain how to rig closed loops for USD and export a robot to URDF, but there doesn't seem to be a resource that clearly explains how to correctly setup joints and their articulation so that the exported URDF contains the intended structure and kinematics. From my point of view, it would be really helpful to have some guidance on this.