Guidance for running SIMPLE Isaac rendering on GH200/aarch64

[DanielLSM]

Summary

We are trying to run SIMPLE on a GH200/H100 aarch64 cluster and would appreciate guidance on the intended or recommended porting path for Isaac Sim rendering/data generation.

The upstream Dockerfile currently defaults to Isaac Sim 4.5.0. That seems to be the tested path, but the public nvcr.io/nvidia/isaac-sim:4.5.0 container is amd64-only, while GH200 systems are aarch64.

Environment

• Hardware: NVIDIA GH200 / H100 class nodes
• Architecture: aarch64
• Execution substrate: Slurm + container runtime
• SIMPLE mode tested: mujoco_isaac
• Target task tested: G1WholebodyCloseDoorTeleop-v0
• Policy/eval route tested: psi0_decoupled_wbc / eval_decoupled_wbc.py

What we tried

Because Isaac Sim 4.5.0 is not available for aarch64, we tried a GH200-compatible Isaac Sim 6.x ARM64 route and translated the runtime/container setup while keeping SIMPLE task semantics as close as possible.

The simulation/policy path can run: the environment initializes, actions are served, episodes complete, and MP4 files are written. However, the Isaac RGB camera output in headless mujoco_isaac runs is not visually usable on this stack. The saved camera frames appear heavily clipped/stippled/edge-like before video encoding.

We tried several render-side changes without resolving the underlying RGB buffer issue:

• changing camera pose and adding an overview camera
• higher-resolution camera products
• disabling/tuning exposure and motion blur
• lowering/clamping task lighting
• neutral scene materials
• RTX realtime vs path tracing settings
• capture-time cleanup before MP4 writing

These affected the final files but did not restore normal RGB render output, which suggests the issue is likely below SIMPLE's MP4 writer and closer to the Isaac/renderer/GPU/container compatibility path.

Questions

1. Is there a recommended way to run SIMPLE's Isaac rendering path on GH200/aarch64?
2. Is Isaac Sim 6.x expected to work with SIMPLE, or is SIMPLE currently tied to Isaac Sim 4.5.x APIs/behavior?
3. Are GH200/H100 nodes intended to be used only for policy/training/physics while rendering/data generation should happen on x86_64 RTX GPUs?
4. For training new policies, is visually correct Isaac RGB output required at data-generation time, or can training proceed using existing recorded datasets / non-photorealistic renders while Isaac rendering is treated as an evaluation/visualization step?
5. Do you have a minimal smoke command or compatibility checklist for validating a correct SIMPLE Isaac port on non-standard clusters?

Any guidance on a supported GH200/aarch64 strategy, or confirmation that an x86_64 RTX render node is the right route, would be very helpful.

[songlin]

Hi @DanielLSM, thanks for reporting the issue.

We haven't tested SIMPLE on GH200. In general, I wouldn't recommend running Isaac Sim on platforms like GH200 that are primarily designed for large-scale training workloads, since Isaac Sim is only officially supported on a limited set of hardware/software configurations. We also haven't validated SIMPLE with newer versions of Isaac Sim, so you may encounter rendering issues, although they are likely straightforward to resolve.

Our recommended workflow is:

• Use a rendering GPU such as an RTX 5090 for SIMPLE data collection.
• Train your policy on GH200 clusters using the RGB data rendered by the 5090.
• Switch back to the 5090 for evaluation, since high-quality rendering is required again.

The policy server can be deployed either remotely on the GH200 cluster or locally on the 5090, depending on whether your model fits within the 32 GB GPU memory.

Overall, I'd recommend sticking with the RTX 5090 (or a similar rendering-focused GPU) for running Isaac Sim during both data collection and evaluation, while reserving GH200 for policy training.