> This page is for version Release 3.0 (default).
> For other versions, use one of these documentation indexes:
> - Release 3.0 (default): https://docs.nvidia.com/isaaclab/3.0/llms.txt
> - Release 2.3: https://docs.nvidia.com/isaaclab/2.3/llms.txt
> - Release 2.2: https://docs.nvidia.com/isaaclab/2.2/llms.txt
> - Release 2.1: https://docs.nvidia.com/isaaclab/2.1/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/isaaclab/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/isaaclab/_mcp/server.

# Training and Deploying COMPASS Policies with NuRec Assets

[COMPASS](https://arxiv.org/abs/2502.16372) (Cross-embodiment Mobility Policy via Residual RL and
Skill Synthesis) trains navigation policies that can transfer across robot embodiments. Isaac Lab supports
COMPASS training with NuRec scene assets, so policies can be developed in reconstructed real-world
environments instead of relying only on hand-authored simulation scenes.

This page gives a short overview of the supported NuRec assets and example COMPASS training clips. For
installation, asset registration, training, evaluation, export, and ROS 2 deployment instructions, see the
[COMPASS handbook](https://github.com/NVlabs/COMPASS/blob/real2sim/isaaclab_3.0/docs/handbook/workflows/nurec_real2sim.md).

## NuRec Assets for COMPASS

NuRec assets matter for COMPASS because they allow navigation policies to train in simulation on realistic
geometry, visual appearance, and obstacle layouts. This, in turn, should reduce the sim-to-real gap between
training scenes and real operating environments and improve zero-shot transfer. They also make it possible
to evaluate policies in reconstructed simulation scenes that better reflect target deployment spaces.

Some sample assets compatible with COMPASS training are available in the
[PhysicalAI-Robotics-NuRec dataset on Hugging Face](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-NuRec).

### Real2Sim Assets

[NVIDIA Neural Reconstruction (NuRec)](https://docs.nvidia.com/nurec/index.html) converts real-world
sensor data into digital simulation-ready assets for training and testing Physical AI agents. The Real2Sim
workflow converts stereo RGB captures into Isaac Sim-ready assets with aligned geometry, collision meshes,
and simulation-ready scene augmentation. The PhysicalAI-Robotics-NuRec dataset includes several robotics
scenes produced with this workflow; two examples are shown below.

![Galileo Real2Sim NuRec asset](https://download.isaacsim.omniverse.nvidia.com/isaaclab/images/nurec_real2sim_galileo.gif)

**Galileo**

* Embodiment: [Nova Carter \[wheeled robot\]](https://robotics.segway.com/nova-carter/)
* Sensor Rig: 4 x [Stereo Camera](https://leopardimaging.com/wp-content/uploads/2024/07/LI-AR0234CS-STEREO-GMSL2-30_Datasheet_V1.8.pdf)

![Living Room Real2Sim NuRec asset](https://download.isaacsim.omniverse.nvidia.com/isaaclab/images/nurec_real2sim_living_room.gif)

**Living Room**

* Embodiment: Hand-Held
* Sensor Rig: 1 x [Stereo Camera](https://leopardimaging.com/wp-content/uploads/2024/07/LI-AR0234CS-STEREO-GMSL2-30_Datasheet_V1.8.pdf)

### XGRIDS Assets

COMPASS also supports compatible assets reconstructed with XGRIDS. These assets are post-processed into the
NuRec format used by the COMPASS workflow, including mesh alignment and augmentation with simulation-ready
assets available in Isaac Sim.

![Wormhole XGRIDS reconstruction](https://download.isaacsim.omniverse.nvidia.com/isaaclab/images/xgrid_wormhole_with_sim_objects.gif)

**Wormhole**

* Embodiment: Hand-Held
* Sensor Rig: [XGRIDS PortalCam](https://www.xgrids.com/intl/portalcamp) \[2 x fish-eye cameras, 2 x front cameras, 1 x LiDAR]

## COMPASS Workflow

The complete COMPASS workflow is maintained in the COMPASS repository:

* **Repository:** [NVlabs/COMPASS](https://github.com/NVlabs/COMPASS)
* **Branch:** `real2sim/isaaclab_3.0`
* **Guide:** [NuRec asset workflow](https://github.com/NVlabs/COMPASS/blob/real2sim/isaaclab_3.0/docs/handbook/workflows/nurec_real2sim.md)

The guide covers the COMPASS code path for training with NuRec assets generated by the Real2Sim workflow,
including Isaac Sim and Isaac Lab setup, COMPASS installation, asset download and registration, residual RL
specialist training, evaluation, ONNX and TensorRT export, and ROS 2 sim-to-real deployment.

## Training Examples

The following clips show COMPASS training with different robot embodiments in NuRec scenes.

![COMPASS training with a G1 robot in the Galileo scene](https://download.isaacsim.omniverse.nvidia.com/isaaclab/images/compass_galileo_g1_training.gif)

**Galileo** with G1

![COMPASS training with a Carter robot in the XGRIDS Wormhole scene](https://download.isaacsim.omniverse.nvidia.com/isaaclab/images/compass_xgrid_wormhole_carter_training.gif)

**Wormhole** with Carter