[Question] Lack of temporal coherency results in teleportation of a custom robot

[PsorTheDoctor]

I'm working on a custom humanoid robot defined in a .urdf file converted to .usd and loaded in Isaac Lab using a custom ArticulationCfg object (defined in joule_cfg.py). The setup is adapted from the H1 locomotion example (LocomotionVelocityRoughEnvCfg class) and shares the same reward logic and simulation parameters.

teleportation.mp4

However, after switching from the H1 model to my custom humanoid, the simulation becomes unstable with a noticeable lack of temporal coherency. After a few simulation steps (or immediately after reset), the robot behaves unpredictably:

• teleports or catapults itself with very high velocity,
• individual links vibrate or oscillated rapidly,
• occasionally the robot detaches or stretches apart

These issues appear inconsistently - sometimes the simulation runs stably for a few seconds before diverging.

Questions

• Are there recommended PhysX parameters (solver iterations, depenetration velocity, damping) for articulated humanoids?
• What’s the correct workflow to validate a .urdf before import (checking mass, inertia, and collisions)?
• Should self-collisions be disabled for humanoid training unless explicit collision filters are defined?
• Is there a built-in PhysX diagnostic tool to detect overlapping collisions or penetration at spawn?

[RandomOakForest]

Thank you for posting this. It's a great post for our Discussions section, I'll move it there for follow up. Here is a summary that to consider, may be of help. Please share the version of your Isaac Lab and Isaac Sim, as we are moving quickly across releases and you may be hitting a migration issue.

Recommended PhysX Parameters for Articulated Humanoids

• Solver Iterations: Increase both num_position_iterations (e.g., 8–16) and num_velocity_iterations (e.g., 2–4) for humanoids to improve stability under high-contact and stacking scenarios. The default for simpler robots (6/1) often needs to be doubled for complex humanoids.¹
• Depenetration/Contact Offset: Tune contact_offset and rest_offset. Recommended contact_offset is 0.01–0.02, and rest_offset is usually 0.0. These control when contacts are detected and help avoid instability during stacking or close interactions.²¹
• Damping and Friction: Always set joint drive damping (damping parameter) nonzero (e.g., > 5.0 for large joints; start with 10–20 for humanoid legs), regardless of URDF defaults. Add joint friction and optionally velocity clamps to constrain extreme velocities.³⁴
• Timestep/Substeps: Use a smaller simulation timestep (e.g., $dt = \frac{1}{120}$) and increase substeps if simulation is diverging quickly. This will slow simulation but greatly help with temporal coherency.¹

Workflow to Validate URDF Files Before Import

1. Check Mass and Inertia: Ensure every link has nonzero mass and realistic inertia. URDF must specify these; otherwise, import tools will estimate or use defaults, which can cause erratic movement.⁴⁵
2. Validate Collisions: Make sure each link has appropriate collision geometry. Prefer simple primitives (capsule, box, sphere) for stable contact—complex meshes cause instability and expensive computations. Avoid overlapping or intersecting meshes.⁶⁵
3. Self-collision Setup: Before import, review whether adjacent link collision meshes overlap. Intersecting meshes at joints cause instability: adjust mesh or disable self-collision if not explicitly needed.⁵⁴
4. Asset Validation: Isaac Sim has an "Asset Validator" tool (Window > Asset Validator) that checks mass, inertias, collision mesh intersections, correct drive/mimic APIs, and joint configuration. Use this immediately after conversion to USD, before simulation.⁵

Self-collision in Humanoid Training

• For most humanoid locomotion training, disable self-collision unless explicit non-overlapping collision filters are defined. Self-collision between adjacent links can cause severe instability if collision meshes are touching or intersect. If you require self-collision, carefully filter adjacent links and make sure collision meshes are separated.⁷⁴

PhysX Diagnostic Tools for Overlapping Collisions/Penetrations

• Isaac Sim provides overlap diagnostic functionality. You can use built-in overlap tests and demo scenes (see Physics/Demo Scenes > Overlap) to visually detect intersecting or overlapping bodies on spawn.⁸⁵
• The Asset Validator extension automatically checks for overlap between non-adjacent collision meshes after import, helping catch common mistakes before simulation.⁵
• For programmatic detection, use the PhysX API's overlap tests or visual debugging in Isaac Sim. There is no dedicated inspector for articulation penetrations at spawn, but overlap tests and asset validation fulfill the requirement.

Footnotes

1. https://isaac-sim.github.io/IsaacLab/release/2.1.0/source/migration/migrating_from_isaacgymenvs.html ↩ ↩² ↩³

2. https://nvidia-omniverse.github.io/PhysX/physx/5.1.2/docs/AdvancedCollisionDetection.html ↩

3. https://nvidia-omniverse.github.io/PhysX/physx/5.1.3/docs/Articulations.html ↩

4. https://docs.isaacsim.omniverse.nvidia.com/4.2.0/features/environment_setup/ext_omni_isaac_urdf.html ↩ ↩² ↩³ ↩⁴

5. https://docs.isaacsim.omniverse.nvidia.com/5.1.0/robot_setup/asset_validation.html ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

6. http://wiki.ros.org/urdf/Tutorials/Adding Physical and Collision Properties to a URDF Model ↩

7. https://github.com/isaac-sim/IsaacLab/discussions/1428 ↩

8. https://forums.developer.nvidia.com/t/simple-way-to-check-for-collision/177947 ↩