Agent Skills: NVIDIA Isaac Sim

NVIDIA Isaac Sim

Use this skill when upgrading Isaac Sim applications, extensions, USD stages, OmniGraph graphs, sensors, Replicator workflows, or ROS 2 integrations. Start with the breaking-change checks below, then open the topic reference that matches the work.

Reference index

| Reference | Topics | | --- | --- | | references/extensions-and-core-apis.md | Extension renames and removals, Kit compatibility, Core APIs, templates | | references/rtx-sensors-and-rendering.md | Camera, lidar, radar, acoustic, annotators, render timing, sensor corrections | | references/ros2-integration.md | Bridge libraries, Simulation Control, publisher graphs, workspaces, DDS | | references/physics-and-robotics.md | PhysX and Newton, articulations, joints, deformables, robot assets and tools | | references/replicator-and-deployment.md | Replicator Agent, synthetic data, containers, artifacts, offline and Nucleus |

Upgrade triage

Before changing code:

1. Identify the installed Isaac Sim and Kit versions.
2. Inventory extension dependencies, Python imports, settings paths, USD prim types, asset paths, OmniGraph node types, and ROS 2 graph connections.
3. Treat an extension rename as a deliberate migration. Some old extensions split into several replacements, while others have no direct replacement.
4. Remove deprecated dependencies before crossing a major-version boundary.
5. Rebaseline tests around corrected physics, sensor transforms, timestamps, camera intrinsics, and ROS 2 message behavior.
6. Validate packaged artifacts and platform defaults separately from API migrations.

Highest-priority API migrations

Core and project scaffolding

For 6.0 work, move from isaacsim.core.api, isaacsim.core.prims, and isaacsim.core.utils to isaacsim.core.experimental.* and isaacsim.core.simulation_manager.

Replace isaacsim.examples.extension scaffolding with:

./repo.sh template new

The experimental Core interfaces are also the engine-neutral route for physics data when selecting PhysX or Newton.

RTX camera construction and acquisition

The 6.0 experimental camera API separates USD authoring from acquisition:

from isaacsim.sensors.experimental.rtx import CameraSensor, RtxCamera

sensor = CameraSensor(
    RtxCamera(
        "/World/Camera",
        tick_rate=30.0,
        translations=[[0.0, 0.0, 1.0]],
    ),
    resolution=(640, 480),
    annotators=["rgb"],
)
data, info = sensor.get_data("rgb")

Use tick_rate, plural N×3/N×4 transform arrays, explicit annotators, and get_data(). Do not preserve command-based camera construction, frequency/dt, or the unused name argument.

Replace:

• CameraView(prim_paths_expr=..., output_annotators=...) with TiledCameraSensor(paths=[...], annotators=[...]).
• Regex prim selection with an enumerated paths list.
• Tiled implicit output with get_data("rgb", tiled=True).
• SingleViewDepthSensor with SingleViewDepthCameraSensor(RtxCamera.create(...), annotators=[...]).
• Polynomial distortion setters with native OpenCV fisheye or pinhole USD schemas.

RTX lidar, radar, and acoustic construction

Author prims with Lidar.create(), Radar.create(), or Acoustic.create(), then wrap them in LidarSensor, RadarSensor, or AcousticSensor:

from isaacsim.sensors.experimental.rtx import Lidar, LidarSensor

sensor = LidarSensor(
    Lidar.create(
        "/World/Lidar",
        config="Example_Rotary",
        orientations=[[1.0, 0.0, 0.0, 0.0]],
    ),
    annotators=["generic-model-output"],
)
data, info = sensor.get_data("generic-model-output")

Use explicit annotator acquisition and parse generic model output with parse_generic_model_output_data. Use parse_stable_id_map_data and parse_object_ids for stable IDs. Timeline control replaces per-sensor initialize, pause, and resume methods.

Keep these exceptions in view:

• IDS has no experimental create-command replacement; use the deprecated command or author USD directly.
• The active point-cloud annotator remains in isaacsim.sensors.rtx.nodes.
• Radar Doppler requires Motion BVH.
• Auxiliary levels use _replicator:rendervar:GenericModelOutput:channels.
• Mixed auxiliary levels on attached render products are last-attach-wins.

Version-sensitive RTX point-cloud nodes

Do not choose a lidar point-cloud node from memory:

• In 5.0, legacy point-cloud and scan-buffer nodes moved to OgnIsaacExtractRTXSensorPointCloud.
• In 5.1, per-frame output from OgnIsaacCreateRTXLidarScanBuffer replaced that 5.0 path.
• In 6.0, use the experimental sensor API where applicable, while honoring the point-cloud annotator exception above.

Also attach every required lidar output explicitly; implicit annotator or writer attachment is no longer a safe assumption.

ROS 2 graph migration

Publisher nodes no longer perform direct USD lookup for forward-compatible graphs.

For transforms, add Isaac Compute Transform Tree from isaacsim.core.nodes, set targetPrims and optionally parentPrim, then connect:

• execOut
• parentFrames
• childFrames
• translations
• orientations

to the corresponding publisher inputs. An articulation-root target expands to its complete link tree.

For joint states, add Isaac Read Joint State from isaacsim.sensors.physics.nodes, set prim to the articulation root, and wire execOut, joint names, positions, velocities, efforts, DOF types, stage meters per unit, and sensor time to the publisher.

For 6.0 Publish TF and Odometry graphs, feed precomputed transforms from IsaacComputeTransformTree; direct targetPrims is deprecated.

Replicator Agent migration

Treat Replicator Agent 1.x as an architectural replacement:

| 0.x surface | 1.x surface | | --- | --- | | scene.asset_path | Required environment.base_stage_asset_path | | simulation_length in 30-FPS frames | simulation_duration in seconds | | sdg_scheduler.py | actor_sdg.py | | omni.anim.people | omni.anim.behavior.core | | Command text and JSON transition maps | Inline YAML routines and triggers | | Top-level response, event, incident | Actor-group triggers | | Character filters | Removed |

The environment path may be asset-relative, a URL, or absolute. Select robot onboard cameras with camera_prim_paths, and validate the Pydantic v2 configuration before long generation jobs.

Physics and robotics quick choices

• Use floating-base inverse dynamics for gravity/Coriolis compensation, mass matrices, center of mass, and centroidal momentum calculations.
• CPU signed-distance-field triangle-mesh collision is supported for small scenes.
• Use compliant mimic-joint constraints when manipulation stability benefits.
• Consider solve-articulation-contacts-last to reduce gripper penetration when tuning is imperfect.
• Use PhysxJointAxisAPI for Stribeck-like friction and PhysxDrivePerformanceEnvelopeAPI for speed-dependent motor torque.
• Enable the beta deformable schema in Physics preferences and restart before using it.
• When Newton is active, expect importers, policies, ROS 2, and tensor-backed graph nodes to select registered Newton-compatible paths.

Runtime and deployment defaults

• Multitick rendering can schedule camera and RTX lidar rates and offsets from physics simulation time.
• On 5.1, the ROS 2 bridge system_default selects internal Humble on Ubuntu 22.04 and internal Jazzy on Ubuntu 24.04.
• On 6.0 Jazzy workspaces, use_internal_libs defaults to false; Humble retains true.
• An empty dds_type preserves RMW_IMPLEMENTATION; an explicit value selects Fast DDS, Cyclone DDS, or Zenoh.
• Containers run rootless by default in 5.1, and live streaming is unsupported on aarch64 there.
• For the complete offline asset pack, acquire, verify, combine, and extract all five parts.
• The WebRTC desktop client is optional for headless Docker Compose web-viewer deployments.

Verification checklist

• Run the Compatibility Checker for the installation modality.
• Confirm every extension and import resolves without a compatibility alias.
• Inspect authored prim types and moved asset paths.
• Verify annotators and writers are attached explicitly.
• Recheck lidar timestamps, transforms, motion settings, and auxiliary levels.
• Compare camera intrinsics and distortion schemas with downstream consumers.
• Exercise ROS 2 world, spawn, transform, joint-state, and time semantics.
• Rebaseline physics assertions affected by corrected simulation results.
• Validate Replicator configurations before expensive runs.
• Test container user, architecture, ROS library, DDS, streaming, and offline asset assumptions on the actual deployment platform.