Known Issues

Each entry below names the backends it affects. An issue listed under one backend does not apply to the others unless it says so.

PhysX backends

Surface grippers require CPU simulation

Affects: physics=isaacsim_physx surface-gripper tasks.

Surface grippers require CPU simulation. This includes the UR10 Long/Short Suction stacking tasks, the Galbot Right Arm Suction stacking task, and its relative and absolute Mimic variants. Pass --device cpu when running teleoperation. Zero and random agents preserve these tasks’ CPU defaults when --device is omitted; an explicit GPU override is unsupported.

GPU articulation solver partition aliasing

Affects: physics=isaacsim_physx on CUDA when PhysxCfg.gpu_max_num_partitions > 1.

The PhysX GPU articulation solver can corrupt articulation state when a scene contains more than 32 articulation instances and at least one articulation has more than 64 links. The corruption can appear as runaway joint velocities or NaNs in otherwise separate environments. See Isaac Lab issue #8121 for the reported repro.

As a workaround, set PhysxCfg(gpu_max_num_partitions=1) until the underlying PhysX issue is fixed. Isaac Lab logs a warning for managed PhysX ~isaaclab.assets.Articulation assets when the known conditions are detected.

Sensor readings are stale immediately after a reset

Affects: physics=isaacsim_physx, physics=ovphysx, and any RTX-based renderer.

Many physics engines do a simulation step as a two-level call: forward() and simulate(), where the kinematic and dynamic states are updated respectively. PhysX has only a single step() call where the two operations are combined. Because of computations through GPU kernels, it is not straightforward to split them. As a result, writing a root or joint state does not by itself run a full forward pass.

For articulation link poses this is handled: reading isaaclab.assets.ArticulationData.body_link_pose_w (or the deprecated isaaclab.assets.ArticulationData.body_state_w) triggers a PhysX kinematic update, so link poses reflect a preceding root-state or joint-state write without an intervening step().

For RTX rendering-based sensors — cameras in particular — the data is still not refreshed by a state write. The rendering engine update is bundled with the simulator’s step() call, so the sensor data is only refreshed when the simulation is stepped forward, and a read taken between a reset and the next step returns the previous frame.

There is currently no direct workaround for the sensor case. From our experience, the reset values affect agent learning in proportion to how frequently the environment terminates; as an agent learns successfully, the termination rate drops and the effect becomes less significant.

Exiting the process

Affects: physics=isaacsim_physx.

When exiting a process with Ctrl+C, occasionally the below error may appear:

[Error] [omni.physx.plugin] Subscription cannot be changed during the event call.

This is due to the termination occurring in the middle of a physics event call and should not affect the functionality of Isaac Lab. It is safe to ignore the error message and continue with terminating the process. On Windows systems, please use Ctrl+Break or Ctrl+fn+B to terminate the process.

Newton backends

OpenUSD can crash while parsing collider-rich rigid bodies

Affects: Kitless Newton workflows using OpenUSD releases earlier than 26.05, and Isaac Sim releases earlier than 6.1. Isaac Sim 6.1 and later contain the upstream fix in their bundled OpenUSD runtime.

Older OpenUSD releases have a thread-safety issue in UsdPhysics.LoadUsdPhysicsFromRange when multiple colliders belong to the same rigid body. Newton uses this API while importing a USD stage, so affected processes can terminate during scene creation with a segmentation fault, access violation, heap-corruption message, or hang. The issue and upstream fix are described in OpenUSD PR #4002.

As a workaround for an affected kitless or pre-6.1 Isaac Sim runtime, limit OpenUSD to one worker thread. Set PXR_WORK_THREAD_LIMIT before launching Python so the limit is present before any pxr module initializes:

PXR_WORK_THREAD_LIMIT=1 uv run isaaclab train --rl_library rsl_rl --task Isaac-Reach-Franka physics=newton_mjwarp

This environment variable limits OpenUSD’s process-wide worker pool to one thread and can reduce USD import performance. Use it only with affected runtimes, and remove it after upgrading to an OpenUSD provider that contains the fix or to Isaac Sim 6.1 or later.

Closed-loop articulations are not validated on Kamino (e.g. Agility Digit)

Affects: physics=newton_kamino.

Robots whose USD encodes a closed kinematic loop — such as the achilles rod and toe push-rods on the Agility Digit — are not validated on newton_kamino.

Renderers

Blank initial frames from the camera

Affects: RTX-based renderers (renderer=isaacsim_rtx, renderer=ovrtx).

When using the ~isaaclab.sensors.Camera sensor in standalone scripts, the first few frames may be blank. This is a known issue with the simulator where it needs a few steps to load the material textures properly and fill up the render targets. It is most likely on a cold asset cache and in scenes with many or large textures; simple scenes with locally cached assets often render content on the very first frame.

If you do see blank frames, add the following after initializing the camera sensor and setting its pose:

from isaaclab.sim import SimulationContext
sim = SimulationContext.instance()
# note: the number of steps might vary depending on how complicated the scene is.
for _ in range(12):
    sim.render()

Scene partitioning is capped at 15625 partitions

Affects: renderer=isaacsim_rtx with scene partitioning enabled, and renderer=ovrtx.

The underlying rtx.scenedb.plugin allocates a fixed-size pool of scene partitions and caps it at 15625, regardless of which renderer requests them. Isaac Lab assigns one scene partition per environment when isaaclab_physx.renderers.IsaacRtxRendererCfg.enable_scene_partitioning is enabled for the Isaac RTX renderer, and OVRTX always assigns one scene partition per environment, so runs with more than 15625 environments exceed the pool on either backend. Once the cap is hit, rtx.scenedb.plugin logs a warning and discards the remaining partitions:

[Warning] [rtx.scenedb.plugin] SceneDbContext : Maximum number of scene partitions
(15625) reached. Additional scene partitions will be discarded.

Environments beyond the cap are left without their own partition and end up sharing one with another environment, so their tiled camera views can render another environment’s geometry instead of their own.

Large RTX camera workloads can return black or corrupted images

Affects: RTX-based renderers (renderer=isaacsim_rtx, renderer=ovrtx), independent of the physics backend.

Increasing the camera count, resolution, or path-tracing samples per pixel (SPP) can cause camera views to become black or corrupted without raising an error. Simulation and training can continue with invalid observations.

Camera views are combined into a tiled image (an atlas). On affected renderer versions, the rendering workload for a large atlas can exceed the supported Vulkan shader-access range for a buffer, approximately 4 GiB in the reported configuration. This can happen even with free GPU memory. The threshold depends on the padded atlas dimensions and rendering settings, so there is no universal camera-count limit. See Isaac Lab issue #8412 for affected configurations, reproduction steps, and fix status.

As a workaround, reduce the camera resolution, camera count, or path-tracing SPP, and inspect images from all cameras before training. Reducing the camera count may leave the atlas dimensions unchanged because of tile padding. Path-tracing SPP is distinct from the direct-lighting samples_per_pixel setting.

If RTX image quality is not required, the Newton Warp renderer (renderer=newton_renderer) does not render through Vulkan and is not subject to this limit.

Using instanceable assets for markers

Affects: all Kit-based renderers.

When using instanceable assets for markers, the markers do not work properly, since Omniverse does not support instanceable assets when using the UsdGeom.PointInstancer schema. This is a known issue and will hopefully be fixed in a future release.

If you use an instanceable assets for markers, the marker class removes all the physics properties of the asset. This is then replicated across other references of the same asset since physics properties of instanceable assets are stored in the instanceable asset’s USD file and not in its stage reference’s USD file.

Asset import

URDF importer: unresolved references for fixed joints

Affects: the URDF importer, independent of the physics backend.

Links connected through fixed_joint elements are not merged when their URDF link entries specify mass and inertia, even if merge-joint is set to True. This is expected behaviour — those links are treated as full bodies rather than zero-mass reference frames. However, the USD importer currently raises ReportError warnings showing unresolved references for such links when they lack visuals or colliders. This is a known bug in the importer; it creates references to visuals that do not exist. The warnings can be safely ignored until the importer is updated.

Environment and setup

GLIBCXX errors in conda environments

Affects: conda-based installations, independent of the physics backend.

Some workflows exit with an OSError indicating version 'GLIBCXX_3.4.30' not found when running from a conda environment. The issue appears to stem from importing torch or torch-related packages, such as tensorboard, before Isaac Sim starts. As a workaround, ensure that all torch imports happen inside ~isaaclab.app.launch_simulation, after the runtime has started, which should resolve the error.