How are camera images stored in HDF5 recorder, and can this dataset be used for π0.5 training?

[Manu752]

Question

Hello,

I am using the ActionStateRecorderManagerCfg with a ManagerBasedRLEnv that includes an RGB camera observation (e.g., robot_pov_cam) and concatenate_terms = False.

I would like clarification on the following:

HDF5 structure

What is the exact structure used by the recorder when exporting image observations?

Are RGB frames stored directly inside the HDF5 as (T, H, W, C) or (T, C, H, W)?

Is the dtype guaranteed to be uint8 when normalize=False?

Is there any recommended way to efficiently store high-frequency camera data (compression settings, chunking, etc.)?

Dataset compatibility with π0.5 (OpenPI)

Is the exported HDF5 format intended to be directly usable for training vision-action models such as π0.5?

Or is the expected workflow to convert Isaac Lab recordings into a LeRobot dataset format first?

Are there any official or recommended conversion scripts for Isaac Lab → LeRobot/OpenPI?

My goal is to record teleoperated demonstrations (RGB + proprio + actions) and fine-tune a π0.5 policy. I want to ensure the recording format is compatible and scalable before collecting large datasets.

• Isaac sim 5.1
• Isaac Lab 2.3.1